Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

A new tool for assessing Pectus Excavatum by a semi-automatic image processing pipeline calculating the classical severity indexes and a new marker: the Volumetric Correction Index

A new tool for assessing Pectus Excavatum by a semi-automatic image processing pipeline... Background: In clinical assessment of Pectus Excavatum (PE), the indication to surgery is based not only on symp- toms but also on quantitative markers calculated from Computed Tomography (CT ) or Magnetic Resonance Imaging (MRI) scans. According to clinical routine, these indexes are measured manually by radiologists with limited computer support. This process is time consuming and potentially subjected to inaccuracy and individual variability in measure- ments. Moreover, the existing indexes have limitations, since they are based on linear measurements performed on single slices rather than on volumetric data derived from all the thoracic scans. Results: In this paper we present an image processing pipeline aimed at providing radiologists with a computer-aid tool in support of diagnosis of PE patients developed in MATLAB and conceived for MRI images. This framework has a dual purpose: (i) to automatize computation of clinical indexes with a view to ease and standardize pre-operative evaluation; (ii) to propose a new marker of pathological severity based on volumetric analysis and overcoming the limitations of existing axial slice-based indexes. Final designed framework is semi-automatic, requiring some user interventions at crucial steps: this is realized through a Graphical User Interface (GUI) that simplifies the interaction between the user and the tools. We tested our pipeline on 50 pediatric patients from Gaslini Children’s Hospital and performed manual computation of indexes, comparing the results between the proposed tool and gold-standard clinical practice. Automatic indexes provided by our algorithm have shown good agreement with manual measure- ments by two independent readers. Moreover, the new proposed Volumetric Correction Index ( VCI) has exhibited good correlation with standardized markers of pathological severity, proving to be a potential innovative tool for diagnosis, treatment, and follow-up. *Correspondence: trorosella@gmail.com Department of Informatics, Bioengineering Robotics and System Engineering (DIBRIS), University of Genoa, Viale Causa 13, 16143 Genova, Italy Full list of author information is available at the end of the article © The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons. org/ publi cdoma in/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Trò et al. BMC Medical Imaging (2022) 22:30 Page 2 of 16 Conclusions: Our pipeline represents an innovative image processing in PE evaluation, based on MRI images (radiation-free) and providing the clinician with a quick and accurate tool for automatically calculating the classical PE severity indexes and a new more comprehensive marker: the Volumetric Correction Index. Keywords: Pectus Excavatum, Magnetic resonance imaging, Image processing pipeline Background highlighted the ability of chest MRI to detail anatomi- Pectus Excavatum (PE) is the most common congenital cal information such as displacement and rotation of chest-wall deformity in children [1]. It is characterized by the heart or great vessels anomalies, promoting the a sunken deformity of the anterior chest wall, involving adoption of this modality in pre operative workup for both sternum and costal cartilages. The deformity wors - patients with PE. ens during adolescence and is primarily male-dominated, A particular MRI technique, the Cardiac Magnetic with a male/female ratio of 5:1 [2]. Although originally Resonance Imaging (CMRI) [24], represents an added considered an aesthetic condition without clinical impli- value in the evaluation of the influence of sternum cations, several studies conducted in the past decades impingement on cardiac function [25, 26]. Specifically, have demonstrated that PE has a substantial psychosocial CMRI allows for a careful surgical evaluation and pre- impact among developing children [3] and may also lead operative cardiac function assessment, overcoming to disabling cardiopulmonary manifestations in worst technical difficulty as well as subjectivity inherent to cases [4–6]. Indeed, when the deformity is moderate to cardiac ultrasound imaging [27]. Despite being the gold severe, it can reduce the volume of the chest, restrict the standard to evaluate the cardiac function for all cardi- pulmonary movement, and force the heart into a rotated opathies, the use of CMRI in patients with PE dates to position [7]. These important cardiopulmonary implica - recent times. tions can be substantially improved with surgical correc- The first study which propelled momentum for CMRI tion [5, 6, 8, 9]. to be used in preoperative assessment of PE dates to In order to assess the severity of the malformation and 2010. Saleh et al. [28] showed how CMRI could unravel determine treatment options, including surgical repair, findings associated with severe PE condition not patients with PE are evaluated through thoracic imaging, detectable with cardiac ultrasound, corresponding to a particularly Computed Tomography (CT) and Magnetic significant reduction of the Right Ventricular Ejection Resonance Imaging (MRI). These imaging modalities Fraction (RVEF) along with a distortion in the right allow to extract several indexes used as markers to quan- ventricle geometry. Similar findings were confirmed in tify the degree of severity [10]. For years, CT has been the a more recent study by Dore et al. [18]. gold standard for preoperative evaluation of PE, provid- In 2013, Humphries et  al. [29] employed CMRI for ing bone details, anatomic relations, and an option for 3D perioperative evaluation of sternal eversion technique reconstruction [11–16]. However, considering the young used for PE repair. They found improvement of anatom - age of PE patients, the efforts to avoid unnecessary radia - ical chest wall contour and cardiac function, suggesting tion exposure should be maximized [11, 17]. Additionally, once again CMRI as a promising tool for delineating CT provides static results, which do not allow to know the anatomical and physiological components of PE as the changes in chest compression during the breathing well as measuring the results of surgical repair. cycle. A dynamic measurement during the normal res- More recently, Deviggiano et  al. [25] combined CT piratory cycle is only possible with a high radiation dose, and CMRI modalities to evaluate the impact of the that should be avoided in such young patients [18, 19]. malformation severity on both morphological and For these reasons, in the last decade MRI has acquired functional cardiac parameters, respectively. Patients an important role in the assessment of this pathology. affected by PE showed significant alterations of car - Several studies have validated this modality as an alter- diac morphology and function that were related to the native radiation-free diagnostic tool for the assessment of severity of the deformation and that manifested as an malformation indexes [19–23]. exaggerated interventricular dependence. Despite the variety of MRI sequences adopted, all In 2019, Vina et  al. [26] demonstrated an excel- aforementioned works agreed to prove reliability, fea- lent agreement between chest CT and standard CMR sibility, and image quality  of fast chest MRI protocols for the evaluatiom of chest wall malformations, thus for preoperative evaluation of PE. Indeed, they showed potentially enabling preoperative assessment of PE that severity indexes of chest deformity collected from severity and cardiac involvement with a single non- CT scan and fast MRI were comparable. They also invasive diagnostic tool. Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 3 of 16 In the same year, Lai et  al. [30] showed that, in Implementation patients with mild PE deformity and minimal symp- Current framework is organized in four interconnected toms at rest, cardiac MRI might reveal additional modules summarized in Fig.  1: Pre-processing, Depres- functional information than echocardiography, able to sion quantification, Inner chest contour segmentation explain exertional symptoms. They also demonstrated and Thoracic indexes computation. The software code resolution of cardiac dysfunction with surgical repair of has been developed in MATLAB 2020a (https:// it. PE. mathw orks. com/), running in Windows 10. In 2021, Stagnaro et  al. [31] analyzed cardiovascular As a preliminary step for subsequent analyses, a range effects beside of thoracic indexes with multiparametric of slices of interest must be selected, including the slice CMR, using a simple noninvasive device mimicking the of maximal sternal depression, on which measurements immediate, temporary effect of surgical correction with for PE indices are usually performed in clinic. Indeed, the Vacuum Bell (VB). not all axial images acquired are useful for our analysis, If some attempts to automatize image processing of CT but only the ones in which the chest depression and the scans of PE patients have been made very recently [15, lungs are clearly visible and thus could be quantified. Of 32], in all existing MRI studies, to the best of or knowl- course, this excludes marginal slices at the beginning and edge, the chest-wall malformation indices are manually at the end of the scan, where the amount of depression is computed by radiologists. According to a commonly negligible. adopted standard procedure, the latter measure specific This step has been implemented through a Graphical thoracic distances with a ruler on axial images of the User Interface for selecting range of slices, slice for PE patient’s chest on a standard DICOM viewer for medical indexes computation, as well as patient’s sex (Additional images. These thoracic measures are then used to calcu - file 1: Fig. 1). late clinical indexes according to their specific formula [33]]. The critical points of this working method are long Pre‑processing processing time, low reliability and low reproducibility in In order to improve low contrast inherent to MR images, measurements [14]. firstly we perform a contrast adjustment by remapping The aim of our work is to develop an image process - the values of the input intensity to fill the entire inten - ing framework for evaluation of PE using Magnetic Reso- sity range. Then, we focus exclusively on chest district nance Imaging (MRI), which can support, standardize by excluding arms placed at the borders of images, due and accelerate the diagnostic assessment of patients. to the small dimension of chest in pediatric patients. This Firstly, we want to automatize the computation of exist- is obtained by defining a proper mask, based on subject’s ing indexes on MRI images, given the lack of automatic thorax morphology (Additional file 1: Fig. 2). procedures for MRI modality. The other purpose is to introduce an innovative marker of pathological sever- Depression quantification ity, based on a volumetric analysis to quantify chest This module has the goal to quantify the depression, depression. Indeed, most of the existing clinical indexes based on a volumetric study. Indeed, rather than evaluat- are calculated on a single slice, usually corresponding ing the depression on a single slice, as traditional radio- to maximum sternal depression [26]. Thus, accuracy of logical indices commonly adopted in clinical practice do, these indexes largely depends on which images are cho- we propose to analyze multiple slices in order to meas- sen and how measurements are performed from them. ure the depression volume. The idea is to identify the This fact could determine a high degree of variability of two maximum and the minimum points of the outer measured indexes. Specifically, we want to elaborate an chest contour for each slice considered and thus define image processing method that first corrects the depres - an elliptic curve between the two maximum points to sion, by simulating the normal morphology of the chest, correct the depression and simulate the normal chest, in and then obtains the amount of depression by comparing absence of PE malformation. The difference between the the images of thorax before and after the image correc- chest image before and after image correction gives the tion. The ratio between the depression volume and the amount of the depression. chest volume post-correction gives the portion of chest that must be repaired. This new measure, that we named Analysis of outer chest contour Volumetric Correction Index (VCI), could represent a First of all, the algorithm turns the grey-scale image into more comprehensive marker, complementary to existing a binary image, by applying a manual threshold (T = 0.1) clinical indices, of effective patient pre-treatment condi - to separate the foreground from the background pixels. tion, assisting physicians in diagnosis process and proper Indeed, this value proves to apply for all examined sub- treatment choice. jects. Then, we get exterior boundaries of chest in terms Trò et al. BMC Medical Imaging (2022) 22:30 Page 4 of 16 Fig. 1 Image analysis framework is composed of four interconnected elements. a First module consists in pre-processing of selected slices, that is contrast stretching and cropping on the area of interest. Outcome of this step is a binary mask, used as input for subsequent pipeline. b Second module is represented by quantification of the chest depression. Outer chest contour detection serves as a preliminary step for depression computation, quantified as the portion between an elliptic profile and external contour. c The latter is exploited for next phase, which is inner chest contour segmentation. This is performed through consecutive sub-steps, which include lung segmentation and similarity between inner and outer wall contour. d Final outcome of this pipeline allows to obtain thoracic indexes on the reference slice as well as new volumetric marker. All these measures are saved in a Microsoft Excel file per each subject of Cartesian  x  and  y  coordinates through morphological where: gradient operators in order to identify the two maximum and the minimum points of outer chest contour. This is • (x, y): coordinates of ellipse points just a rough detection of maximum points, since binariza- • (x , y ): coordinates of right vertex of major axis 1 1 tion may alter upper profile of images. We thus resort to • (x , y ): coordinates of left vertex of major axis 2 2 morphological operators of closing to correct upper image • a: semi-major axis boundaries. By only considering the upper half of outer • b: semi-minor axis, found as b = a ∗ 1 − e chest boundary, the algorithm exploits the spatial discon- • e: ellipse eccentricity tinuities of boundary pixels along y direction in order to • t: variation angle of ellipse points, defined between 0 identify the two maximum points (Fig. 2a,b). As regards the and π (half ellipse) minimum point, we use boundary pixel locations before • α: rotation angle, defined as the angle between the morphological operations and find it as lowest peak in the horizontal line and major axis range of y coordinates between the two already identified maximum points (Fig. 2c–e). Eccentricity (e) and positions of right and left vertices of major axis ((x , y ) and (x , y )) represent the param- 1 1 2 2 Depression volume eters we have modified in order to simulate the profile By analyzing MR images of chest in normal patients, we of outer chest contour in normal patients. After several have noticed that the best curve, representing the chest tests, e has been set to 0.99. morphology between the two maximum points of outer Regarding the position of major axis vertices, we have chest contour, could be a roto-translated ellipse, whose separated patients based on their sex. Indeed, anatomi- points are calculated as follow: cal differences between male and female forced us to deal with the depression issue in a distinct way. By ana- x + x 1 2 x = + a ∗ cos(t) ∗ cos(α) − b ∗ sin(t) ∗ sin(α) lyzing normal chest images of male subjects, we were able to find a unique method to define the position of y + y 1 2 y = + a ∗ cos(t) ∗ sin(α) − b ∗ sin(t) ∗ cos(α) major axis vertices. Specifically, the algorithm identi - fies y coordinates (y and y ) by lowering the position of 1 2 Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 5 of 16 Fig. 2 Outer chest contour detection. a Plot of upper half of outer chest boundary pixel coordinates after morphological operations, among which research of the two maximum points, shown with light blue arrows, is performed. b Binary image with the two maximum points in red, identified after morphological operations. The image shows as the latter modifies the minimum position. c Plot of upper half of outer chest boundary pixel coordinate before morphological operations. Between the two already identified maximum points (dashed gray vertical lines), minimum point research is performed. d Binary image with the minimum point in red, identified before morphological operations. e Grey-scale image with the two maximum and the minimum points in red two maximum points of outer chest contour by a con- including them in the automatic computation of stand- stant value, while the x coordinates (× 1 and × 2) are ard indexes. found by searching the most extreme points at the same After the ellipse has been obtained, the algorithm finds y coordinates (Fig. 3a). the indices corresponding to x and y ellipse coordinates This correction method does not work in female in the image matrix and adds pixels to the binary image patients, because the presence of breast rises the posi- of chest in these specific locations. Consequently, the tion of major axis vertices by causing a wrong depres- depression is filled by applying a morphological opera - sion correction. As among female subjects there is tion of closing using a disk as structural element (Fig. 3b). a high variability in chest shape due both to age and The depression area is thus calculated as the difference differentiated anatomical growth, it is impossible between the image before application of morphological to define a single correction method for the depres - operators and the one after correction with the ellipti- sion. For these reasons, we decided to exclude female cal curve. Finally, the depression volume is computed by patients from depression quantification analysis, while summing the volumes obtained for each slice, resulting Trò et al. BMC Medical Imaging (2022) 22:30 Page 6 of 16 Fig. 3 Automatic procedure for depression area filling. a Binary image with two maximum points of outer chest contour in blue and two vertices of ellipse major axis in red, after operation of lowering. b Grey-scale image with elliptical curve of correction. c Grey-scale image with missing chest area, caused by PE malformation, in white from the product of depression areas by ‘slice thickness’ Firstly, the algorithm isolates the inner chest portion by image DICOM attribute. After repeating this operation exploiting lung segmentation and similarity between the for each slice, the algorithm is able to represent on grey- inner and outer wall contour. Then, it excludes the verte - scale images how the normal morphology of chest should bral body by thresholding method. Finally, it corrects the be (Fig. 3c). errors in the detection of inner chest contours through a comparison among consecutive slices. For this analy- sis step, we opted for working on a limited number of New pathological marker computation slices, by excluding the ones preceding the slice selected The absolute value of depression volume cannot be used for index computation. Indeed, the remaining range of as a pathological marker since it is strongly dependent on slices ensures an easily implementable segmentation of chest dimension and on the number of slices considered inner chest region thanks to an optimal lung-background for its computation, which is different from patient to contrast. patient. Thus, we decided to normalize it on the thorax volume after the correction, as it simulates the ‘normal’ condition of the chest. Specifically, the algorithm quan - Lung segmentation tifies the correct chest volume in the same way as for In view of performing segmentation of the lungs, we depression volume by considering the binary image after used histogram analysis for identification of the cor - depression correction. The new pathological marker, that rect threshold. The grey-level histogram of a MR image we named Volumetric Correction Index, is defined as is characterized by a high variability, both across sub- follow: jects and across slices within the same patients, in peaks’ shapes corresponding to the lungs and to cardiac struc- depression volume tures and thorax tissue, respectively. For this reason, Volumetric Correction Index = ∗ 100 correct chest volume Otsu thresholding technique [34], the standard approach for histogram partitioning, does not perform well due Therefore, the new index proposed represents the to its inability to correctly separate bimodal histograms percentage of depression that must be corrected in PE when the two classes are very different in size. There patients. fore, we developed a method to automatically partition a grey-level histogram, by adapting an algorithm presented Inner chest contour segmentation by [35]. The idea proposed by this study, that we applied This module aims at detecting the inner contour of the to our problem, is to locate the concavity between the two principal peaks in the curve representing the image chest, fundamental for PE indexes calculation. If this task histogram by maximizing divergence between the histo is difficult in CT images, where the attenuation coeffi - - cients of the heart and the chest-wall are quite close, it gram and a Gaussian fit. is even more challenging in MR images, where different After computing the histogram of the grey-scale image chest regions often have a high similarity in terms of grey in continuous form, the algorithm defines an auxiliary levels. curve P(x) on the same grey-level range of the histogram Therefore, in order to simplify the segmentation pro H(x). We assumed P(x) as a normal distribution, with mean given by µ, the average gray-level of H(x), and the cess, we designed this module by subdividing it in con- corresponding variance given by σ secutive steps, as described in the following sections. . We also considered Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 7 of 16 P(x) and H(x) to have an identical area α under their absence of lungs. Specifically, the algorithm computes the curves. Given that x ≤ x ≤ x , P(x) is defined lung area and relates it to the entire chest area. Then, it min max as:P(x) = G(x)where: selects only the slices in which the ratio is greater than 20%. Therefore, if the slice selected for indices calculation 1 −(x−μ) √ shows high similarity in grey values between different • G(x) = exp 2σ 2πσ chest areas, the algorithm automatically picks the first max • z(x) = G(x) x=x min following slice, where inner chest contour detection can max • α = H(x) x=x min be performed properly. 1 x=x max • μ = xH(x) x=x α min 1 x=x 2 2 max • σ = (x − μ) H(x) x=x α min Inner chest contour detection In order to isolate the inner thoracic region, we adapted As a normal distribution, P(x) presents its largest value an algorithm proposed by [36], for the inner curvature at x = µ and has a convex part that goes from µ − σ to detection of CT images. Specifically, they proposed a µ + σ. The oddity is that the highest peaks of H(x) are recursive algorithm that exploits outer wall contour as a close to µ, the average grey level, such that the concavi- starter point for inner contour segmentation, due to sim- ties surrounded by the highest peaks in H(x) are often in ilarity in morphology between them. contrast with the convex part of P(x). Hence, the line (l) We identified as algorithm inputs, obtained from pre - that divides the main peaks in H(x) can be easily found by vious module, the matrices composed of pixel locations maximizing the difference between P(x) and H(x), or for - of each lung and the matrix containing pixel locations of mally: l = arg max (P(x) − H(x)) , with outer curvature. μ − σ ≤ x ≤ μ + σ. The algorithm goes through steps along the outer The l value corresponds to the threshold separating the curvature in clockwise direction until the start point two main peaks in image histogram. We implemented is found again. Every 12 steps the actual point and the this histogram partitioning method in MATLAB and point 12 steps before are connected and a perpendicular applied it twice in our analysis. First, it is used to sepa- line in the mid-point of their connection is generated. rate the chest area from the background. Thus, it analyzes Then the algorithm finds the intersection point between the lower part of histogram, by considering as input the the perpendicular line and the first point crossed by it range of x values in the low grey-level region. Once found on the two lungs. In the area of the binary image where the threshold (l ) that removes the background from the perpendicular lines do not cross any lungs, a correc- bg the image, a new grey-level histogram H’(x) is generated, tion of the invalid points generated is necessary. u Th s, it considering only the pixels related to the chest. Hence, calculates for each line the distances between the mid- the method is reapplied to the new data to estimate the point and the intersection point and computes their correct threshold for the lung segmentation (l ). Spe- mean value (µ ) and the standard deviation (σ ). We have lung d d cifically, in order to enhance the threshold search, the designed a length filter by defining as invalid the inter - algorithm focuses the analysis on the lower part of H’(x), section points whose distances from the mid-points are since it corresponds to grey values belonging to lungs longer than a specific threshold that we identified as 2* (Additional file 1: Fig. 3). µ − σ . All points, corresponding to longer distances d d After finding appropriate thresholds for each slice with than this value, are deleted and replaced by new ones this strategy, these are used to segment lungs from chest located at the same distances as the previous valid point region. Specifically, a mask is created where pixels with (Fig.  4a). Once all intersection points have been found, intensity above the l are set as white and the remain- the inner curvature is calculated by an interpolating pro- lung ing ones are set as black. Other segmented elements cess. Initially, the algorithm prepares intersection points besides the lungs are removed and morphological opera- by separating them in two subsets: the ones related to the tion of closing are applied to smooth edges and fill holes upper half of the inner contour and those belonging to inside the lungs. the lower part. Additionally, it performs an initial correc- tion, by deleting points whose y locations are in discon- tinuity with y positions of neighboring points, in order Selection of appropriate slice for indices computation to avoid the eventual errors made by previous opera- Finally, before proceeding to inner contour detection, we tions. Then it applies a shape-preserving piecewise cubic created an automatic technique to exclusively select the interpolation method (‘pchip’) with a high sampling rate. slices where lungs are clearly visible. Indeed, we wanted Finally, we obtain a group of closely spaced points both to exclude from further analysis those slices where inner for superior and inferior half of inner contour. After re- chest region segmentation could be complex, due to the combining them in a unique set of points, we can define Trò et al. BMC Medical Imaging (2022) 22:30 Page 8 of 16 holes. Finally, we obtain a binary image representing the inner chest region from which to extract boundary pixel locations for each slice. However, a further correction may be necessary both around the vertebral body, having grey values close both to lungs and to cardiac structure intensities, and around inferior lung area, being often difficult separating pixels belonging to lungs and to thoracic tissue ones (Fig. 5b). We thus designed a method to correct the inner chest Fig. 4 Algorithm for preliminary inner wall contour a Binary image representing lung region. Yellow line represents outer chest contour by comparing consecutive slices thanks to high curvature. In blue there are the perpendicular lines, generated similarity in boundary pixel positions belonging to the every 12th step. In red there are the intersection points resulting lower half of inner contour of our interest. At this step, from recursive algorithm. b Gray-scale image on which inner mask user intervention is required such that correction pro- boundary points are indicated in green, while intersection points cess starts from a slice where inner contour detection found by recursive algorithm in red. This is clearly a rough contour of inner chest, including vertebral body, and thus further corrections are does not present errors. Once first slice is selected, the required algorithm starts the pair-wise comparison among adja- cent slices in both directions, by taking as reference the points belonging to the contour of slice selected. u Th s, the boundary of a mask that isolates the inner thoracic the algorithm computes the distance among them and all region (Fig.  4b). However, this mask also includes the the points belonging to the contour of the adjacent slice, vertebral body and is not accurate in all the slices, mostly which is assumed as incorrect. Then it creates a vector due to bad lung segmentation. For these reasons, it is that, for each point belonging to the contour to correct, necessary to improve the inner chest segmentation with only keeps the minimum distance among all those just further processing. computed. It also calculates the maximum value (d ) max and the standard deviation (σ ) of all minimum distances. Inner chest contour correction After several tests, we established that the algorithm The first step of inner chest contour correction consists must continue only if σ is greater than 1.8. Additionally, in excluding the vertebral body. For doing so, the algo- we identified d -2* σ as threshold value that sepa- max d rithm applies a thresholding method by using the inner rates correct points from incorrect ones. Thus, for each mask found in previous section as a tool to improve seg- incorrect point, the algorithm finds the range that must mentation. After masking out the external chest area, a be deleted, by identifying its extremities in the near- slice-wise threshold is defined to exclude out heart and est points to the correct curve. Then, it replaces them other cardiac elements through histogram partitioning with the points belonging to the correct curve by using method presented in previous section. Indeed, correction a shape-preserving piecewise cubic interpolation method is not possible without masking out cardiac district, due (‘pchip’) (Fig. 5c). Once a new curve is obtained, the algo- to high similarity in grey values between heart and tho- rithm proceeds to the next slice, by taking as reference racic tissue. Thus, the algorithm returns to original chest the just corrected contour. Such an algorithm allows to image, before the application of inner mask, and assigns satisfactorily correct errors in segmentations (Fig. 5d). to background the just segmented pixels belonging to cardiac structures (Fig. 5a). Then, it is able to separate the Thoracic indexes computation inner chest region from the outer chest one, by applying This module aims at computing PE indices used by phy - as threshold the same value found for lung segmentation sicians to classify the severity of patients’ malformation. (l ). To have the inner region as foreground, the com- As mentioned above, among multiple thoracic markers, lung plementary image is computed, and some morphological we focused on the severity (Haller index and Correc- operations are applied to smooth the edges and fill the tion index) (Fig.  6a, b) and deformity (Asymmetry index (See figure on next page.) Fig. 5 Algorithm for inner wall contour correction. a Grey-scale image after masking out cardiac structures in order to create a mask of inner thoracic area. b Grey-scale image with inner chest boundary pixel locations in red. The error appears around inferior lung area, that has grey values close to thoracic tissue ones. c Plot of x and y coordinates corresponding to inner chest boundary pixels. In red there is the reference curve used for correction, while in green the curve that need to be corrected by algorithm. Between two blue arrows there are the points resulting from interpolation process that substitute incorrect ones. d Inner chest boundary pixel locations after correction are represented in red while those before correction in green; in blue there is inner contour of the reference curve, belonging to previous slice Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 9 of 16 Fig. 5 (See legend on previous page.) Trò et al. BMC Medical Imaging (2022) 22:30 Page 10 of 16 minimum x coordinate and the second one as the point at its same y coordinate. Conversely, the first extremity for measuring min APd corresponds to sternum position. We approximate it as the point with maximum y coordi- nate (y values decrease toward the bottom of image) by only considering the range of inner chest contour points between x coordinate positions of two maximum points of outer chest contour upper half. Second extremity cor- responds to the vertebral body position. It is taken as the point with minimum y coordinate by only considering the range of inner chest contour points between position of x coordinates of two maximum points of outer chest contour lower half. Correction index Correction index (iCorrection) is calculated by dividing the amount of defect, measured as the difference between the maximum anteroposterior distance, i.e. the maxi- mum distance between the anterior spine and the anterior Fig. 6 Inner thoracic distances overlaid on slice of maximal sternal portion of the chest (max APd) and the minimum anter- depression. a Distances useful for iHaller computation: transverse oposterior diameter (min APd), to the maximum anter- diameter in red, min APd in yellow. In cyan there is the horizontal line at same y position of vertebral body. b Distances useful for oposterior distance (max APd), multiplied by 100 [37]. iCorrection computation: min APd in yellow, max APd in green. In (max APd − min APd) cyan there is the horizontal line at same y position of vertebral body. iCorrection = ∗ 100 c Distances useful for iAsymmetry computation: right hemithorax min APd APd in blue and left hemithorax APd in magenta. d Distances useful for iFlatness computation: right hemithorax APd in blue and For max APd computation, firstly, the algorithm draws transverse diameter in red a horizontal line at the same y coordinate of vertebral body position, that is assumed as the anterior spine posi- tion. Then, it identifies two points on inner chest contour and Flatness index) ones (Fig.  6c, d). The algorithm only at the same x coordinate of two maximum points of outer works on the first slice of images processed in the previ - chest contour. The latter are assumed as the positions of ous module. Indeed, it corresponds to the slice selected right and left anterior portion of the chest. Thus, for each by the user or to the first following one where inner chest point it computes the distances between them and the contour can be detected. horizontal line and gets the maximum diameter between Once inner distances and thoracic indices are com- the two distances. puted, the framework saves their results along with the new pathological marker obtained in Depression quan- Asymmetry index tification module in an Excel file, located in the same Asymmetry index (iAsymmetry) is calculated by divid- folder as patient’s images. Each quantified distance in fol - ing the longest anteroposterior distance of the right chest lowing computations has been multiplied by ‘pixel spac- wall (right hemithorax APd) to the longest anteroposte- ing’ attribute to have measures in mm. rior distance of the left chest wall (left hemithorax APd), multiplied by 100 [38]. Haller index Haller index (iHaller) is calculated by dividing the trans- right hemithorax APd iAsymmetry = ∗ 100 verse diameter, i.e., the widest horizontal distance of the left hemithorax APd inside of the ribcage, to the minimum anteroposterior diameter (min APd), i.e., the shorter distance between Right hemithorax APd’s extremities are identified as the vertebral body and the sternum [12]. the points on inner chest contour at the same x coor- dinate of first maximum point, as it is located in right transverse diameter iHaller = hemithorax. Right hemithorax APd’s bounds are identi- min APd fied as the points on inner chest contour at the same x coordinate of second maximum point, as it is situated in As regards transverse diameter, the algorithm identifies left hemithorax. its first extremity as the point on inner chest contour with Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 11 of 16 Flatness index msec (maximum) and a 32-element cardiac phased- Flatness index (iFlatness) is computed by dividing array coil for signal reception and cardiac synchroniza- the transverse diameter of the thorax to the longer of tion (with “retrospective gating” technique). Our MRI the two maximum anteroposterior diameters of the protocol borrowed cardiac gating and breath-holding right (right hemithorax APd) and left hemithorax (left techniques as well as specific sequences from CMRI, hemithorax APd) [13]. As all the distances have been in order to overcome motion-related artifacts and to already found, the algorithm can proceed with Flatness inspect with further detail cardiovascular morphology. index computation, as follow: The MR acquisition setting thus included scout images and Steady State Free Precession (SSFP) images in transverse diameter iFlatness = axial, coronal and sagittal planes, acquired at the end of max righthemithoraxAPd; lefthemithoraxAPd expiratory phases. Specifically, the SSFP sequence was a Gradient-Echo sequence, named Balanced Turbo Field Echo-Breath Hold (BTFE-BH). Total scanning time was User’s correction 5–8 min approximately. As mentioned above, the algorithm does not always perform indices computation on the same slice selected Results by user, due to its inability to segment images where Out of 50, just three subjects (2 male and 1 female) have different chest areas have similar grey values. In these been excluded from our analysis since characterized by cases, it selects the first following slice, where inner extremely low contrast images that the algorithm could chest contour detection can be performed. We noticed not process. We can thus conclude that, provided suffi - that by going through consecutive slices some inner cient contrast in the input raw image, proposed method distances maintain their value constant, while others, mantains its reliability and accuracy for the whole cohort specifically min APd and max APd are more likely to ander analysis. vary. For this reason, in the algorithm we add the pos- Our image processing pipeline has then been quan- sibility of user’s intervention when the slice is different titatively evaluated through comparison with manual from the one selected. Specifically, the user is asked procedure. As mentioned in previous section, the slice to insert two points on the image, useful for min APd selected by user for indices computation is often dif- and max APd calculation: sternum position and verte- ficult to segment due to similar grey values of different bral body position. Thus, the algorithm recomputes the thoracic regions, so that algorithm automatically picks indices by considering the modifications on these two the first following slice, where inner chest contour detec - inner distances. Finally, two sets of results are obtained: tion can be performed. Specifically, out of the 47 patients the ones calculated on the slice picked by the algorithm processed by algorithm, the latter was able to use the and those obtained on the same slice selected after cor- same slice as the one selected in 24 patients, while in the rection of sternum and vertebral body points. remaining it selected another scan. We thus separated the patients in two groups, depending on whether the Modules validation indices computation was performed on the same slice Current methodological framework has been devel- selected by user (group 1) or it was executed on a differ - oped from a small subset counting 5 subjects. In order ent slice picked by the algorithm (group 2). to test the overall quality of our algorithm, we extended its application to other 45 pediatric patients affected by Automatic framework agrees with manual procedure Pectus Excavatum from Gaslini Children’s Hospital, in for indexes computation Genoa, for a total of 50-subjects dataset. In the absence of a ground truth to test performance Additionally, two expert radiologists manually per- against, the accuracy of the thoracic indices resulting formed double-blind thoracic indices computation, from algorithm was evaluated by comparing them to as they routinely do in the clinical setting. The group results obtained by manual measures performed by two of patients consisted in 41 males and 9 females aged expert radiologists (through a double-blind analysis). 13.5 ± 2.78 (mean ± SD), age range 5–18 years. Each of Table  1 shows the results of patients belonging to them underwent MRI examination, in order to estab- group 1. Results of inner thoracic distances show a good lish the severity of malformation and thus the best agreement between measures obtained by the 2 read- treatment strategies. MRI examinations were per- ers and the algorithm. Naturally, the difference is higher formed on a 1.5 Tesla MR scanner (Achieva, Philips by comparing manual results to automatic ones, as it is Healthcare, Cleveland, OH, USA), equipped with 66 shown by a greater mean standard deviation. We can mT/m gradients (maximum), a slew rate of 180 mT/m/ notice that transverse diameter, min APd and max APd Trò et al. BMC Medical Imaging (2022) 22:30 Page 12 of 16 Table 1 Average inner thoracic distances and thoracic indices along with relative mean standard deviation between 2 readers and mean standard deviation (std) among readers and algorithm in case of appropriate user selection of main slice for indices computation Reader 1 Reader 2 Algorithm Std between 2 Std among readers readers and algorithm Thoracic distances (cm) Transverse diameter 24.1 ± 2.6 24.2 ± 2.6 24.4 ± 2.9 0.17 0.30 Min APd 4.9 ± 1.3 5.0 ± 1.4 5.1 ± 1.4 0.18 0.30 Max APd 7.2 ± 0.94 7.6 ± 0.93 7.5 ± 0.95 0.31 0.34 Right hemithorax APd 12.4 ± 1.2 12.6 ± 1.3 12.2 ± 1.2 0.24 0.37 Left hemithorax APd 12.3 ± 1.2 12.6 ± 1.1 11.9 ± 1.2 0.30 0.47 Thoracic indices Haller index 5.3 ± 1.9 5.3 ± 1.9 5.1 ± 1.6 0.31 0.34 Correction index (%) 32.5 ± 13.9 35.2 ± 14.1 32.6 ± 13.6 2.7 3.6 Asymmetry index (%) 101.7 ± 6.7 99.8 ± 5.2 102.8 ± 8.2 2.2 3.2 Flatness index 1.9 ± 0.19 1.9 ± 0.17 2.0 ± 0.20 0.038 0.057 are computed by the algorithm in a comparable way as readers. As we mentioned above, we noticed that some results obtained manually. Contrariwise, right hemitho- distances remained almost constant by measuring them rax APd and left hemithorax APd are characterized by a on consecutive slices. Contrariwise, two distances, spe- higher variability that, however, is stronger also between cifically min APd and max APd, showed more variability the 2 readers. Obviously, inner distances affect the results among consecutive slices. Consequently, we decided to of thoracic indices. Specifically, we notice that Haller apply a correction factor to these measurements, to be index and Flatness index results are comparable, whereas able to compare the algorithm results to the ones com- the differences between manual and automatic computa - puted by the readers. Specifically, as we observed that tion increase by considering Correction index and Asym- algorithm tends to overestimate both min APd and max metry index. APd, we subtracted to them a corrective factor that we identified as the mean standard deviation between read - Inner thoracic distances and thoracic indices in case ers and algorithm (0.50 in both cases). Table  2 shows of matching slice selection the results of patients belonging to group 2, after the Regarding the results belonging to group 2, the measure- just mentioned correction of the two inner distances. ments obtained from the algorithm were performed on The same considerations made for inner thoracic dis - a different slice compared to the one analyzed by the 2 tances results belonging to group 1 apply also in this case. Table 2 Average inner thoracic distances and thoracic indices along with relative mean standard deviation (std) between 2 readers and mean standard deviation among readers and algorithm in case of failed user selection of main slice for indices computation Reader 1 Reader 2 Algorithm Std between 2 Std among readers readers and algorithm Thoracic distances (cm) Transverse diameter 23.2 ± 1.6 23.3 ± 1.6 23.5 ± 1.7 0.06 0.26 Min APd 3.9 ± 1.4 4.0 ± 1.6 4.2 ± 1.6 0.16 0.36 Max APd 6.5 ± 0.85 6.9 ± 0.83 6.8 ± 0.94 0.25 0.35 Right hemithorax APd 11.3 ± 1.1 11.5 ± 1.2 11.2 ± 1.4 0.17 0.39 Left hemithorax APd 11.9 ± 0.96 12.0 ± 1.0 11.7 ± 1.1 0.14 0.33 Thoracic indices Haller index 7.2 ± 4.0 7.2 ± 4.3 7.0 ± 4.6 0.23 0.63 Correction index (%) 42.1 ± 17.9 42.5 ± 19.9 39.1 ± 18.8 2.5 4.7 Asymmetry index (%) 95.1 ± 5.5 96.0 ± 6.8 96.5 ± 8.1 1.4 4.0 Flatness index 2.0 ± 0.16 1.9 ± 0.17 2.0 ± 0.16 0.023 0.050 Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 13 of 16 However, by observing thoracic indices results, we notice the correlation increases with severity indices, i.e. Haller a higher variability among readers and algorithm than the index, and slightly more with Correction index, that are one found in Table 1. The reason is mainly due to the use the most used by physicians to assess the PE malforma- of a different slice for indices computation. Furthermore, tion in a quantitative way. As it is shown in Fig. 7c, small there are more severe cases of PE among patients belong- depressions (low VCI) correspond to low Haller indices. ing to this group. This aspect could be another cause for However, if Haller index begins to increase the linear the higher variability in indices, specifically Haller index. relation tends to disappear, as the same iHaller corre- Indeed, we noticed that differences among reader and sponds to different degrees of depression. Same behavior algorithm results increase when the min APd assumes is visible in Fig.  7d that shows the correlation between low values, as it is placed at the denominator in the index VCI and iCorrection. However, it should be noted that calculation formula. u Th s, variability is higher for high severe cases of PE with high iHaller and iCorrection are Haller indices rather than lower ones. few among all the patients analyzed. Thus, the lack of linear tendency could be caused by a limited number of Inner thoracic distances and thoracic indices in case cases with high degree of PE severity. of not matching slice selection Additional file  1: Fig. 4 displays scatter charts represent- Person correlation between Volumetric Correction Index ing comparison among results obtained by readers and and other traditional indices algorithm for each thoracic index, belonging to patients See Table 3. of group 1. Finally, the average time necessary to perform tra- Discussion ditional indices computation on a single patient was We introduce a set of tools to aid the pre- and post-sur- 50  s and 3  min 45  s for automatic (on a standard Win- gery assessment of PE patients. We opted for developing dows workstation with i7-core and 8  GB RAM) and this algorithm within an existing software rather than a manual processing, respectively. However, if radiologists new stand-alone tool in order to ensure later extensibility are beginners, the time could significatively increase, across different centers. even rising to twice the value indicated for manual The set of algorithms have been tested both qualita - computation. tively and quantitatively through a cohort of 50 pediat- ric patients with varying age, sex and disease severity. In New volumetric index as a promising marker of PE severity our study, we were able to show that automatic results Finally, the new index calculated by our algorithm exclu- obtained by our algorithm are comparable with the ones sively on male subjects (n = 39), named VCI, was com- manually computed by expert radiologist. The proposed pared to all the thoracic indices, in order to evaluate its algorithm offers different advantages First of all, it gives feasibility for quantitative evaluation of PE. Specifically, physicians an accurate tool not subjected to individual we calculated statistical Pearson correlation between VCI interpretation or errors and represents a useful sup- and other traditional indices, as shown in Table  3. Fur- port in establishing proper treatment decision, includ- thermore, scatter plot analysis between new pathological ing the need for surgical correction of malformation. marker and other indices are shown in Fig. 7. Moreover, it ensures a faster processing time compared The results of correlation show a very low correlation to manual measurements, which gets relevant in case of between VCI marker and Asymmetry index and Flatness large datasets and radiologists with limited experience. index (Fig.  7a, b). We expected this behavior since both Furthermore, for now limited to the male subset of our indices do not quantify the severity of depression but the cohort, we suggested a new pathological marker to better degree of chest asymmetry and flatness. Contrariwise, quantify the depression caused by PE: the VCI. Indeed, indexes used so far are based on linear measurements of chest diameters, but they do not evaluate the chest in the Table 3 Result of Pearson correlation between traditional tridimensional aspect of the deformity, which has clinical indices and new pathological marker computed by algorithm implications. A patient with a deep but very localized PE Indices comparison Pearson could have a worse Haller index or Correction index than correlation coefficient another patient with a less severe but more extended PE, even if the real impact of the deformity and the compres- iHaller – VCI (%) 0.79 sion on lung and heart could be globally worse in the sec- iCorrection (%) – VCI (%) 0.81 ond patient, due to the diffuse PE. Therefore, an index iAsymmetry (%) – VCI (%) 0.062 which considers all the missing volume of the thorax iFlatness – VCI (%) 0.22 and not only measures the severity of PE at a single level Trò et al. BMC Medical Imaging (2022) 22:30 Page 14 of 16 Fig. 7 Linear relationship of new volumetric index with existing clinical markers a Linear Regression between VCI and iHaller. b VCI and iCorrection. c VCI and iFlatness. d VCI and iAsymmetry. Estimates for the slope and intercept of the linear equation as well as R are reported for each measure could overcome the limits of traditional thoracic indices, By analyzing patient-wise results, we could notice that such as the dependence on the slice selected for measure- accuracy of algorithm outcomes is strongly dependent ments or chest shape. Theoretically, it could also have a on quality of MR images acquired. Thus, optimization of better clinical correlation than the current indexes. Nev- acquisition setting would lead to higher-quality images ertheless, further investigations are required to prove the and thus improve pipeline’s performance without the clinical relevance of VCI and incorporate it among the need of further corrections beside main modules. This clinical and radiological parameters considered in the improvement may also allow to properly handle with the decision regarding surgical indication. quantification of depression volume for female patients. Our work has made a relevant contribution to the lit- Indeed, one current limitation of this study is exclu- erature. Indeed, the other studies focused on automatic sion of female patients from computation of newly pro- or semi-automatic quantification of the markers of chest- posed VCI marker, given the higher variability in chest wall deformity are exclusively limited to CT scans [14, 15, shape caused by differential breast growth. A similar 36, 39, 40]. situation may apply for overweight patients with rel- The peculiarity of our algorithm is that it works on MR evant gynecomastia. However, both target patients images. The adoption of MRI in the evaluation of this substantially represent outliers for this kind of condi- condition is relatively recent since, despite their non- tion, mainly affecting under- or normal-weight male invasiveness, MRI scans are more complex to process adolescents [41–43]. As a result, current algorithm with automatic segmentation methods than standard CT already performs successfully for most PE candidates, ones [21, 23, 29, 30]. Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 15 of 16 Abbreviations being other cathegories statistically negligible in their CT: Computed tomography; CMR: Cardiac magnetic resonance; DICOM: amount. Digital imaging and communications in medicine; GUI: Graphical user inter- Another possible limitation to the accuracy in calcu- face; iAsymmetry: Asymmetry index; iCorrection: Correction index; iFlatness: Flatness index; iHaller: Haller index; max APd: Maximum anterior–posterior lation of VCI is the thickness of subcutaneous fat tis- diameter; min APd: Minimum anterior–posterior diameter; MRI: Magnetic sue in pectoral region. However, as anticipated, the vast resonance imaging; PE: Pectus excavatum; VCI: Volume Correction Index. majority of PE patients are very slim so the influence of the subcutaneous tissue is in our opinion neglectable in Supplementary Information most patients. Calculating VCI at the level of cartilage The online version contains supplementary material available at https:// doi. and bones instead of the inner skin level could potentally org/ 10. 1186/ s12880- 022- 00754-0. overcome this limitation. One last potential future development of present algo- Additional file 1: Figure 1. Graphical User Interface for slices selection. Figure 2. Image border correction. Figure 3. Histogram partitioning for rithm may include the computation of cardiac indices. lung segmentation. Figure 4. Comparison between double-blind manual Specifically, it may be useful to develop a new method for measurements and automatic algorithm for computation of thoracic quantification of cardiac compression caused by PE mal - indexes for patients of group 1 a. formation, in order to compare it with the new pathologi- cal marker proposed. Acknowledgements The authors thank all the Technicians and Nurses of our Cardiac MRI Unit for their contribution, and in general to their dedication for Cardio-Radiologic Group since the foundation in 2007 Conclusions In this work we present a piece of software specifically Authors’ contributions SM, RT and MF designed, coded, tested the software tools and performed the designed to support radiologists in diagnosis and best analyses. NS, MT, and VS provided the data and validated as expert the results. personalized treatment choice for patients affected by SM, RT, NS, MT, MF wrote the manuscript. All authors read and approved the PE condition. Indeed, our study proved its reliability and final manuscript. robustness in processing a discrete number of MR images Funding ranging across different degrees of PE severity. Our tool None. significantly eases assessment of pathology by improving Availability of data and materials accuracy of thoracic distances and subsequent clinical The dataset supporting the conclusions of this article is not publicly available indexes beyond subjectivity inherent to manual interven- due to privacy restrictions of clinical data imposed by the Gaslini Hospital’s tion, and by reducing the time required for computation administration but is available from the corresponding author on reasonable request. The Software has been released under MIT License. Code for our of these markers as well. pipeline is available at https:// github. com/ rosel la1234/ PE_ pipel ine. Current Given the relatively high incidence of this disease software has been designed and tested under MATLAB 2020a. A proprietary (1:400 live births), disposing of a novel semi-automatic license is required for using MATLAB . supportive tool enriched with an easily extensible, user- friendly interface may have a substantial clinical impact. Declarations Finally, formulation of a new relevant marker for PE scor- Ethics approval and consent to participate ing paves the way for exploring new strategies for PE This research study was conducted retrospectively from data obtained for clin- assessment. ical purposes. An IRB official waiver of ethical approval was granted from the IRB of IRCCS Istituto Giannina Gaslini. IRB number: 8/2020. Written informed consent was obtained from all patients or their guardians for each participant prior to examination in accordance with the Declaration of Helsinki. Availability and requirements Consent for publication The authors affirm that human research participants and their parents • Project name: e.g. PE_pipeline provided written informed consent for publication of all the images in the • Project home page: https:// github. com/ rosel la1234/ manuscript. PE_ pipel ine/ Competing interests • Operating system(s): Platform independent The other authors declare no competing interests. • Programming language: MATLAB Author details • Other requirements: none Department of Informatics, Bioengineering Robotics and System Engineering • License: e.g. MIT 2 (DIBRIS), University of Genoa, Viale Causa 13, 16143 Genova, Italy. Complex • Any restrictions to use by non-academics: e.g. licence Operative Radiology Unit, IRCCS Giannina Gaslini, Genova, Italy. Depar tment of Health Sciences (DISSAL), Radiodiagnostics, University of Genoa, Via A. Pas- needed for MATLAB tore 1, 16132 Genova, Italy. Pediatric Thoracic and Airway Surgery Unit, IRCCS Giannina Gaslini, Genova, Italy. Trò et al. BMC Medical Imaging (2022) 22:30 Page 16 of 16 Received: 10 September 2021 Accepted: 8 February 2022 23. Marcovici PA, LoSasso BE, Kruk P, Dwek JR. MRI for the evaluation of pectus excavatum. Pediatr Radiol. 2011. 24. Tseng WYI, Su MYM, Tseng YHE. Introduction to cardiovascular magnetic resonance: Technical principles and clinical applications. Acta Cardiol Sin. 2016. 25. Deviggiano A, Carrascosa P, Vallejos J, Bellia-Munzon G, Vina N, Rodríguez- Granillo GA, et al. Relationship between cardiac MR compression clas- References sification and CT chest wall indexes in patients with pectus excavatum. 1. Brochhausen C, Turial S, Müller FKP, Schmitt VH, Coerdt W, Wihlm JM, et al. Journal of Pediatric Surgery. 2018. Pectus excavatum: history, hypotheses and treatment options. Interact 26. Viña NA, Carrascosa P, Mogensen VC, Deviggiano A, Bellia-Munzon G, Cardiovasc Thoracic Surg. 2012. Martinez-Ferro M, et al. Evaluation of pectus excavatum indexes during 2. Fonkalsrud EW. Current management of pectus excavatum. World J Surg. standard cardiac magnetic resonance: potential for single preoperative tool. Clin Imaging. 2019. 3. Kelly RE, Cash TF, Shamberger RC, Mitchell KK, Mellins RB, Lawson ML, 27. Narayan RL, Vaishnava P, Castellano JM, Fuster V. Quantitative assessment et al. Surgical repair of pectus excavatum markedly improves body image of right ventricular function in pectus excavatum. J Thoracic Cardiovasc and perceived ability for physical activity: multicenter study. Pediatrics. Surg. 2012. 28. Saleh RS, Finn JP, Fenchel M, Moghadam A, Krishnam M, Abrazado M, 4. Rodriguez-Granillo GA, Raggio IM, Deviggiano A, Bellia-Munzon G, Capu- et al. Cardiovascular magnetic resonance in patients with pectus excava- nay C, Nazar M, et al. Impact of pectus excavatum on cardiac morphol- tum compared with normal controls. J Cardiovasc Magn Reson. 2010. ogy and function according to the site of maximum compression: effect 29. Humphries CM, Anderson JL, Flores JH, Doty JR. Cardiac magnetic reso- of physical exertion and respiratory cycle. Eur Heart J Cardiovasc Imaging. nance imaging for perioperative evaluation of sternal eversion for pectus excavatum. Eur J Cardio-thoracic Surg. 2013. 5. Jaroszewski DE. Physiologic implications of pectus excavatum. J Thoracic 30. Lai K, Colen TM, Dicken BJ. Integrating cardiac MRI into pre-operative Cardiovasc Surg. 2017. planning for patients with pectus excavatum and right ventricular dys- 6. Sigalet DL, Montgomery M, Harder J, Wong V, Kravarusic D, Alassiri A. function. J Pediatr Surg Case Rep. 2019. Long term cardiopulmonary effects of closed repair of pectus excavatum. 31. Stagnaro N, Trocchio G, Torre M, Rizzo F, Martucciello G, Marasini M. In: Pediatric Surgery International. 2007. Cardiovascular MRI assessment of pectus excavatum in pediatric patients 7. Kotagal M, Swanson JW, Avansino JR. Correlating Haller Index and cardio- and postoperative simulation using vacuum bell. J Pediatr Surg. 2021. pulmonary disease in pectus excavatum. In: Chest Wall Deformities. 2017. 32. Silva B, Pessanha I, Correia-Pinto J, Fonseca JC, Queiros S. Automatic 8. Nuss D, Kelly RE. Indications and technique of nuss procedure for pectus assessment of Pectus Excavatum severity from CT images using deep excavatum. Thoracic Surgery Clinics. 2010. learning. IEEE J Biomed Health Inf. 2021. 9. Jayaramakrishnan K, Wotton R, Bradley A, Naidu B. Does repair of pectus 33. Martinez-Ferro M. Indexes for pectus deformities. In: Chest wall deformi- excavatum improve cardiopulmonary function? Interactive Cardiovascu- ties and corrective procedures. 2016. p. 35–60. lar and Thoracic Surgery. 2013; 34. Otsu, N. A threshold selection method from gray-level histograms. IEEE 10. Abid I, Ewais MM, Marranca J, Jaroszewski DE. Pectus excavatum: a review Trans Syst Man Cybern 1996. of diagnosis and current treatment options. Journal of the American 35. Balan AGR, Traina AJM, Ribeiro MX, Marques PMA, Traina C. Smart histo- Osteopathic Association. 2017. gram analysis applied to the skull-stripping problem in T1-weighted MRI. 11. Sarwar ZU, DeFlorio R, O’Connor SC. Pectus excavatum: current imaging Comput Biol Med. 2012. techniques and opportunities for dose reduction. Seminars in Ultrasound, 36. Papp L, Juhasz R, Travar S, Kolli A, Sorantin E. Automatic detection and CT and MRI. 2014. characterization of funnel chest based on spiral CT. J X-ray Sci Technol. 12. Haller JA, Kramer SS, Lietman SA. Use of CT scans in selection of patients for pectusexcavatum surgery: a preliminary report. J Pediatr Surg. 1987. 37. st. Peter SD, Juang D, Garey CL, Laituri CA, Ostlie DJ, Sharp RJ, et al. A 13. Choi JH, Park IK, Kim Y T, Kim WS, Kang CH. Classification of Pectus Excava- novel measure for pectus excavatum: The correction index. In: Journal of tum according to objective parameters from chest computed tomogra- Pediatric Surgery. 2011. phy. Ann Thoracic Surgery. 2016; 38. Lawson ML, Barnes-Eley M, Burke BL, Mitchell K, Katz ME, Dory CL, et al. 14. Kim M, Lee KY, Park HJ, Kim HY, Kang EY, Oh YW, et al. Development of Reliability of a standardized protocol to calculate cross-sectional chest area new cardiac deformity indexes for pectus excavatum on computed and severity indices to evaluate pectus excavatum. J Pediatr Surg. 2006. tomography: feasibility for pre- and post-operative evaluation. Yonsei 39. Kim HC, Park HJ, Nam KW, Kim SM, Choi EJ, Jin S, et al. Fully automatic Med J. 2009. initialization method for quantitative assessment of chest-wall deformity 15. Lai L, Cai S, Huang L, Zhou H, Xie L. Computer-aided diagnosis of pectus in funnel chest patients. Med Biol Eng Comput. 2010. excavatum using CT images and deep learning methods. Sci Rep. 2020. 40. Barbosa D, Queirós S, Rodrigues N, Correia-Pinto J, Vilaça J. Semi- 16. Chu ZG, Yu JQ, Yang ZG, Peng LQ, Bai HL, Li XM. Correlation between automatic 3D segmentation of costal cartilage in CT data from Pectus sternal depression and cardiac rotation in pectus excavatum: evaluation Excavatum patients. In: Medical Imaging 2015: Image Processing. 2015. with helical CT. Am J Roentgenol. 2010. 41. Nguyen TM, Le VT, Nguyen HU, Pham HL, Phung HSD, Vu NT, et al. An 17. Rattan AS, Laor T, Ryckman FC, Brody AS. Pectus excavatum imaging: initial 5-year single-center experience of 365 patients undergoing the enough but not too much. Pediatr Radiol. 2010. video-assisted thoracoscopic surgery for nuss procedure for pectus 18. Dore M, Triana Junco P, Bret M, Gomez Cervantes M, Muñoz Romo M, excavatum in resource-scare setting. Front Surg. 2021. Jimenez Gomez J, et al. Advantages of cardiac magnetic resonance imag- 42. Kim JJ, Park HJ, Park JK, Cho DG, Moon SW. A study about the costoclavic- ing for severe Pectus Excavatum assessment in children. Eur J Pediatr ular space in patients with pectus excavatum. J Cardiothorac Surg. 2014. Surg. 2018. 43. Akkaş Y, Gülay Peri N, Koçer B, Gülbahar G, Baran Aksakal FN. The preva- 19. Lollert A, Funk J, Tietze N, Turial S, Laudemann K, Düber C, et al. Morpho- lence of chest wall deformity in Turkish children. Turk J Med Scie. 2018. logic assessment of thoracic deformities for the preoperative evaluation of pectus excavatum by magnetic resonance imaging. Eur Radiol. 2015. 20. lo Piccolo R, Bongini U, Basile M, Savelli S, Morelli C, Cerra C, et al. Chest Publisher’s Note fast MRI: an imaging alternative on pre-operative evaluation of Pectus Springer Nature remains neutral with regard to jurisdictional claims in pub- Excavatum. J Pediatr Surg. 2012. lished maps and institutional affiliations. 21. Sun J, Chen C, Peng Y, Zhang Y, Tian H, Yu J, et al. Comparison of magnetic resonance imaging and computed tomography to measure preoperative parameters of children with pectus excavatum. Pediatr Investig. 2019. 22. Birkemeier KL, Podberesky DJ, Salisbury S, Serai S. Limited, fast magnetic resonance imaging as an alternative for preoperative evaluation of pec- tus excavatum: a feasibility study. J Thoracic Imaging. 2012. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BMC Medical Imaging Springer Journals

A new tool for assessing Pectus Excavatum by a semi-automatic image processing pipeline calculating the classical severity indexes and a new marker: the Volumetric Correction Index

Download PDF
16 pages

Loading next page...
 
/lp/springer-journals/a-new-tool-for-assessing-pectus-excavatum-by-a-semi-automatic-image-jcIpgDZqcd

References (89)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2022
eISSN
1471-2342
DOI
10.1186/s12880-022-00754-0
Publisher site
See Article on Publisher Site

Abstract

Background: In clinical assessment of Pectus Excavatum (PE), the indication to surgery is based not only on symp- toms but also on quantitative markers calculated from Computed Tomography (CT ) or Magnetic Resonance Imaging (MRI) scans. According to clinical routine, these indexes are measured manually by radiologists with limited computer support. This process is time consuming and potentially subjected to inaccuracy and individual variability in measure- ments. Moreover, the existing indexes have limitations, since they are based on linear measurements performed on single slices rather than on volumetric data derived from all the thoracic scans. Results: In this paper we present an image processing pipeline aimed at providing radiologists with a computer-aid tool in support of diagnosis of PE patients developed in MATLAB and conceived for MRI images. This framework has a dual purpose: (i) to automatize computation of clinical indexes with a view to ease and standardize pre-operative evaluation; (ii) to propose a new marker of pathological severity based on volumetric analysis and overcoming the limitations of existing axial slice-based indexes. Final designed framework is semi-automatic, requiring some user interventions at crucial steps: this is realized through a Graphical User Interface (GUI) that simplifies the interaction between the user and the tools. We tested our pipeline on 50 pediatric patients from Gaslini Children’s Hospital and performed manual computation of indexes, comparing the results between the proposed tool and gold-standard clinical practice. Automatic indexes provided by our algorithm have shown good agreement with manual measure- ments by two independent readers. Moreover, the new proposed Volumetric Correction Index ( VCI) has exhibited good correlation with standardized markers of pathological severity, proving to be a potential innovative tool for diagnosis, treatment, and follow-up. *Correspondence: trorosella@gmail.com Department of Informatics, Bioengineering Robotics and System Engineering (DIBRIS), University of Genoa, Viale Causa 13, 16143 Genova, Italy Full list of author information is available at the end of the article © The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. The Creative Commons Public Domain Dedication waiver (http:// creat iveco mmons. org/ publi cdoma in/ zero/1. 0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Trò et al. BMC Medical Imaging (2022) 22:30 Page 2 of 16 Conclusions: Our pipeline represents an innovative image processing in PE evaluation, based on MRI images (radiation-free) and providing the clinician with a quick and accurate tool for automatically calculating the classical PE severity indexes and a new more comprehensive marker: the Volumetric Correction Index. Keywords: Pectus Excavatum, Magnetic resonance imaging, Image processing pipeline Background highlighted the ability of chest MRI to detail anatomi- Pectus Excavatum (PE) is the most common congenital cal information such as displacement and rotation of chest-wall deformity in children [1]. It is characterized by the heart or great vessels anomalies, promoting the a sunken deformity of the anterior chest wall, involving adoption of this modality in pre operative workup for both sternum and costal cartilages. The deformity wors - patients with PE. ens during adolescence and is primarily male-dominated, A particular MRI technique, the Cardiac Magnetic with a male/female ratio of 5:1 [2]. Although originally Resonance Imaging (CMRI) [24], represents an added considered an aesthetic condition without clinical impli- value in the evaluation of the influence of sternum cations, several studies conducted in the past decades impingement on cardiac function [25, 26]. Specifically, have demonstrated that PE has a substantial psychosocial CMRI allows for a careful surgical evaluation and pre- impact among developing children [3] and may also lead operative cardiac function assessment, overcoming to disabling cardiopulmonary manifestations in worst technical difficulty as well as subjectivity inherent to cases [4–6]. Indeed, when the deformity is moderate to cardiac ultrasound imaging [27]. Despite being the gold severe, it can reduce the volume of the chest, restrict the standard to evaluate the cardiac function for all cardi- pulmonary movement, and force the heart into a rotated opathies, the use of CMRI in patients with PE dates to position [7]. These important cardiopulmonary implica - recent times. tions can be substantially improved with surgical correc- The first study which propelled momentum for CMRI tion [5, 6, 8, 9]. to be used in preoperative assessment of PE dates to In order to assess the severity of the malformation and 2010. Saleh et al. [28] showed how CMRI could unravel determine treatment options, including surgical repair, findings associated with severe PE condition not patients with PE are evaluated through thoracic imaging, detectable with cardiac ultrasound, corresponding to a particularly Computed Tomography (CT) and Magnetic significant reduction of the Right Ventricular Ejection Resonance Imaging (MRI). These imaging modalities Fraction (RVEF) along with a distortion in the right allow to extract several indexes used as markers to quan- ventricle geometry. Similar findings were confirmed in tify the degree of severity [10]. For years, CT has been the a more recent study by Dore et al. [18]. gold standard for preoperative evaluation of PE, provid- In 2013, Humphries et  al. [29] employed CMRI for ing bone details, anatomic relations, and an option for 3D perioperative evaluation of sternal eversion technique reconstruction [11–16]. However, considering the young used for PE repair. They found improvement of anatom - age of PE patients, the efforts to avoid unnecessary radia - ical chest wall contour and cardiac function, suggesting tion exposure should be maximized [11, 17]. Additionally, once again CMRI as a promising tool for delineating CT provides static results, which do not allow to know the anatomical and physiological components of PE as the changes in chest compression during the breathing well as measuring the results of surgical repair. cycle. A dynamic measurement during the normal res- More recently, Deviggiano et  al. [25] combined CT piratory cycle is only possible with a high radiation dose, and CMRI modalities to evaluate the impact of the that should be avoided in such young patients [18, 19]. malformation severity on both morphological and For these reasons, in the last decade MRI has acquired functional cardiac parameters, respectively. Patients an important role in the assessment of this pathology. affected by PE showed significant alterations of car - Several studies have validated this modality as an alter- diac morphology and function that were related to the native radiation-free diagnostic tool for the assessment of severity of the deformation and that manifested as an malformation indexes [19–23]. exaggerated interventricular dependence. Despite the variety of MRI sequences adopted, all In 2019, Vina et  al. [26] demonstrated an excel- aforementioned works agreed to prove reliability, fea- lent agreement between chest CT and standard CMR sibility, and image quality  of fast chest MRI protocols for the evaluatiom of chest wall malformations, thus for preoperative evaluation of PE. Indeed, they showed potentially enabling preoperative assessment of PE that severity indexes of chest deformity collected from severity and cardiac involvement with a single non- CT scan and fast MRI were comparable. They also invasive diagnostic tool. Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 3 of 16 In the same year, Lai et  al. [30] showed that, in Implementation patients with mild PE deformity and minimal symp- Current framework is organized in four interconnected toms at rest, cardiac MRI might reveal additional modules summarized in Fig.  1: Pre-processing, Depres- functional information than echocardiography, able to sion quantification, Inner chest contour segmentation explain exertional symptoms. They also demonstrated and Thoracic indexes computation. The software code resolution of cardiac dysfunction with surgical repair of has been developed in MATLAB 2020a (https:// it. PE. mathw orks. com/), running in Windows 10. In 2021, Stagnaro et  al. [31] analyzed cardiovascular As a preliminary step for subsequent analyses, a range effects beside of thoracic indexes with multiparametric of slices of interest must be selected, including the slice CMR, using a simple noninvasive device mimicking the of maximal sternal depression, on which measurements immediate, temporary effect of surgical correction with for PE indices are usually performed in clinic. Indeed, the Vacuum Bell (VB). not all axial images acquired are useful for our analysis, If some attempts to automatize image processing of CT but only the ones in which the chest depression and the scans of PE patients have been made very recently [15, lungs are clearly visible and thus could be quantified. Of 32], in all existing MRI studies, to the best of or knowl- course, this excludes marginal slices at the beginning and edge, the chest-wall malformation indices are manually at the end of the scan, where the amount of depression is computed by radiologists. According to a commonly negligible. adopted standard procedure, the latter measure specific This step has been implemented through a Graphical thoracic distances with a ruler on axial images of the User Interface for selecting range of slices, slice for PE patient’s chest on a standard DICOM viewer for medical indexes computation, as well as patient’s sex (Additional images. These thoracic measures are then used to calcu - file 1: Fig. 1). late clinical indexes according to their specific formula [33]]. The critical points of this working method are long Pre‑processing processing time, low reliability and low reproducibility in In order to improve low contrast inherent to MR images, measurements [14]. firstly we perform a contrast adjustment by remapping The aim of our work is to develop an image process - the values of the input intensity to fill the entire inten - ing framework for evaluation of PE using Magnetic Reso- sity range. Then, we focus exclusively on chest district nance Imaging (MRI), which can support, standardize by excluding arms placed at the borders of images, due and accelerate the diagnostic assessment of patients. to the small dimension of chest in pediatric patients. This Firstly, we want to automatize the computation of exist- is obtained by defining a proper mask, based on subject’s ing indexes on MRI images, given the lack of automatic thorax morphology (Additional file 1: Fig. 2). procedures for MRI modality. The other purpose is to introduce an innovative marker of pathological sever- Depression quantification ity, based on a volumetric analysis to quantify chest This module has the goal to quantify the depression, depression. Indeed, most of the existing clinical indexes based on a volumetric study. Indeed, rather than evaluat- are calculated on a single slice, usually corresponding ing the depression on a single slice, as traditional radio- to maximum sternal depression [26]. Thus, accuracy of logical indices commonly adopted in clinical practice do, these indexes largely depends on which images are cho- we propose to analyze multiple slices in order to meas- sen and how measurements are performed from them. ure the depression volume. The idea is to identify the This fact could determine a high degree of variability of two maximum and the minimum points of the outer measured indexes. Specifically, we want to elaborate an chest contour for each slice considered and thus define image processing method that first corrects the depres - an elliptic curve between the two maximum points to sion, by simulating the normal morphology of the chest, correct the depression and simulate the normal chest, in and then obtains the amount of depression by comparing absence of PE malformation. The difference between the the images of thorax before and after the image correc- chest image before and after image correction gives the tion. The ratio between the depression volume and the amount of the depression. chest volume post-correction gives the portion of chest that must be repaired. This new measure, that we named Analysis of outer chest contour Volumetric Correction Index (VCI), could represent a First of all, the algorithm turns the grey-scale image into more comprehensive marker, complementary to existing a binary image, by applying a manual threshold (T = 0.1) clinical indices, of effective patient pre-treatment condi - to separate the foreground from the background pixels. tion, assisting physicians in diagnosis process and proper Indeed, this value proves to apply for all examined sub- treatment choice. jects. Then, we get exterior boundaries of chest in terms Trò et al. BMC Medical Imaging (2022) 22:30 Page 4 of 16 Fig. 1 Image analysis framework is composed of four interconnected elements. a First module consists in pre-processing of selected slices, that is contrast stretching and cropping on the area of interest. Outcome of this step is a binary mask, used as input for subsequent pipeline. b Second module is represented by quantification of the chest depression. Outer chest contour detection serves as a preliminary step for depression computation, quantified as the portion between an elliptic profile and external contour. c The latter is exploited for next phase, which is inner chest contour segmentation. This is performed through consecutive sub-steps, which include lung segmentation and similarity between inner and outer wall contour. d Final outcome of this pipeline allows to obtain thoracic indexes on the reference slice as well as new volumetric marker. All these measures are saved in a Microsoft Excel file per each subject of Cartesian  x  and  y  coordinates through morphological where: gradient operators in order to identify the two maximum and the minimum points of outer chest contour. This is • (x, y): coordinates of ellipse points just a rough detection of maximum points, since binariza- • (x , y ): coordinates of right vertex of major axis 1 1 tion may alter upper profile of images. We thus resort to • (x , y ): coordinates of left vertex of major axis 2 2 morphological operators of closing to correct upper image • a: semi-major axis boundaries. By only considering the upper half of outer • b: semi-minor axis, found as b = a ∗ 1 − e chest boundary, the algorithm exploits the spatial discon- • e: ellipse eccentricity tinuities of boundary pixels along y direction in order to • t: variation angle of ellipse points, defined between 0 identify the two maximum points (Fig. 2a,b). As regards the and π (half ellipse) minimum point, we use boundary pixel locations before • α: rotation angle, defined as the angle between the morphological operations and find it as lowest peak in the horizontal line and major axis range of y coordinates between the two already identified maximum points (Fig. 2c–e). Eccentricity (e) and positions of right and left vertices of major axis ((x , y ) and (x , y )) represent the param- 1 1 2 2 Depression volume eters we have modified in order to simulate the profile By analyzing MR images of chest in normal patients, we of outer chest contour in normal patients. After several have noticed that the best curve, representing the chest tests, e has been set to 0.99. morphology between the two maximum points of outer Regarding the position of major axis vertices, we have chest contour, could be a roto-translated ellipse, whose separated patients based on their sex. Indeed, anatomi- points are calculated as follow: cal differences between male and female forced us to deal with the depression issue in a distinct way. By ana- x + x 1 2 x = + a ∗ cos(t) ∗ cos(α) − b ∗ sin(t) ∗ sin(α) lyzing normal chest images of male subjects, we were able to find a unique method to define the position of y + y 1 2 y = + a ∗ cos(t) ∗ sin(α) − b ∗ sin(t) ∗ cos(α) major axis vertices. Specifically, the algorithm identi - fies y coordinates (y and y ) by lowering the position of 1 2 Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 5 of 16 Fig. 2 Outer chest contour detection. a Plot of upper half of outer chest boundary pixel coordinates after morphological operations, among which research of the two maximum points, shown with light blue arrows, is performed. b Binary image with the two maximum points in red, identified after morphological operations. The image shows as the latter modifies the minimum position. c Plot of upper half of outer chest boundary pixel coordinate before morphological operations. Between the two already identified maximum points (dashed gray vertical lines), minimum point research is performed. d Binary image with the minimum point in red, identified before morphological operations. e Grey-scale image with the two maximum and the minimum points in red two maximum points of outer chest contour by a con- including them in the automatic computation of stand- stant value, while the x coordinates (× 1 and × 2) are ard indexes. found by searching the most extreme points at the same After the ellipse has been obtained, the algorithm finds y coordinates (Fig. 3a). the indices corresponding to x and y ellipse coordinates This correction method does not work in female in the image matrix and adds pixels to the binary image patients, because the presence of breast rises the posi- of chest in these specific locations. Consequently, the tion of major axis vertices by causing a wrong depres- depression is filled by applying a morphological opera - sion correction. As among female subjects there is tion of closing using a disk as structural element (Fig. 3b). a high variability in chest shape due both to age and The depression area is thus calculated as the difference differentiated anatomical growth, it is impossible between the image before application of morphological to define a single correction method for the depres - operators and the one after correction with the ellipti- sion. For these reasons, we decided to exclude female cal curve. Finally, the depression volume is computed by patients from depression quantification analysis, while summing the volumes obtained for each slice, resulting Trò et al. BMC Medical Imaging (2022) 22:30 Page 6 of 16 Fig. 3 Automatic procedure for depression area filling. a Binary image with two maximum points of outer chest contour in blue and two vertices of ellipse major axis in red, after operation of lowering. b Grey-scale image with elliptical curve of correction. c Grey-scale image with missing chest area, caused by PE malformation, in white from the product of depression areas by ‘slice thickness’ Firstly, the algorithm isolates the inner chest portion by image DICOM attribute. After repeating this operation exploiting lung segmentation and similarity between the for each slice, the algorithm is able to represent on grey- inner and outer wall contour. Then, it excludes the verte - scale images how the normal morphology of chest should bral body by thresholding method. Finally, it corrects the be (Fig. 3c). errors in the detection of inner chest contours through a comparison among consecutive slices. For this analy- sis step, we opted for working on a limited number of New pathological marker computation slices, by excluding the ones preceding the slice selected The absolute value of depression volume cannot be used for index computation. Indeed, the remaining range of as a pathological marker since it is strongly dependent on slices ensures an easily implementable segmentation of chest dimension and on the number of slices considered inner chest region thanks to an optimal lung-background for its computation, which is different from patient to contrast. patient. Thus, we decided to normalize it on the thorax volume after the correction, as it simulates the ‘normal’ condition of the chest. Specifically, the algorithm quan - Lung segmentation tifies the correct chest volume in the same way as for In view of performing segmentation of the lungs, we depression volume by considering the binary image after used histogram analysis for identification of the cor - depression correction. The new pathological marker, that rect threshold. The grey-level histogram of a MR image we named Volumetric Correction Index, is defined as is characterized by a high variability, both across sub- follow: jects and across slices within the same patients, in peaks’ shapes corresponding to the lungs and to cardiac struc- depression volume tures and thorax tissue, respectively. For this reason, Volumetric Correction Index = ∗ 100 correct chest volume Otsu thresholding technique [34], the standard approach for histogram partitioning, does not perform well due Therefore, the new index proposed represents the to its inability to correctly separate bimodal histograms percentage of depression that must be corrected in PE when the two classes are very different in size. There patients. fore, we developed a method to automatically partition a grey-level histogram, by adapting an algorithm presented Inner chest contour segmentation by [35]. The idea proposed by this study, that we applied This module aims at detecting the inner contour of the to our problem, is to locate the concavity between the two principal peaks in the curve representing the image chest, fundamental for PE indexes calculation. If this task histogram by maximizing divergence between the histo is difficult in CT images, where the attenuation coeffi - - cients of the heart and the chest-wall are quite close, it gram and a Gaussian fit. is even more challenging in MR images, where different After computing the histogram of the grey-scale image chest regions often have a high similarity in terms of grey in continuous form, the algorithm defines an auxiliary levels. curve P(x) on the same grey-level range of the histogram Therefore, in order to simplify the segmentation pro H(x). We assumed P(x) as a normal distribution, with mean given by µ, the average gray-level of H(x), and the cess, we designed this module by subdividing it in con- corresponding variance given by σ secutive steps, as described in the following sections. . We also considered Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 7 of 16 P(x) and H(x) to have an identical area α under their absence of lungs. Specifically, the algorithm computes the curves. Given that x ≤ x ≤ x , P(x) is defined lung area and relates it to the entire chest area. Then, it min max as:P(x) = G(x)where: selects only the slices in which the ratio is greater than 20%. Therefore, if the slice selected for indices calculation 1 −(x−μ) √ shows high similarity in grey values between different • G(x) = exp 2σ 2πσ chest areas, the algorithm automatically picks the first max • z(x) = G(x) x=x min following slice, where inner chest contour detection can max • α = H(x) x=x min be performed properly. 1 x=x max • μ = xH(x) x=x α min 1 x=x 2 2 max • σ = (x − μ) H(x) x=x α min Inner chest contour detection In order to isolate the inner thoracic region, we adapted As a normal distribution, P(x) presents its largest value an algorithm proposed by [36], for the inner curvature at x = µ and has a convex part that goes from µ − σ to detection of CT images. Specifically, they proposed a µ + σ. The oddity is that the highest peaks of H(x) are recursive algorithm that exploits outer wall contour as a close to µ, the average grey level, such that the concavi- starter point for inner contour segmentation, due to sim- ties surrounded by the highest peaks in H(x) are often in ilarity in morphology between them. contrast with the convex part of P(x). Hence, the line (l) We identified as algorithm inputs, obtained from pre - that divides the main peaks in H(x) can be easily found by vious module, the matrices composed of pixel locations maximizing the difference between P(x) and H(x), or for - of each lung and the matrix containing pixel locations of mally: l = arg max (P(x) − H(x)) , with outer curvature. μ − σ ≤ x ≤ μ + σ. The algorithm goes through steps along the outer The l value corresponds to the threshold separating the curvature in clockwise direction until the start point two main peaks in image histogram. We implemented is found again. Every 12 steps the actual point and the this histogram partitioning method in MATLAB and point 12 steps before are connected and a perpendicular applied it twice in our analysis. First, it is used to sepa- line in the mid-point of their connection is generated. rate the chest area from the background. Thus, it analyzes Then the algorithm finds the intersection point between the lower part of histogram, by considering as input the the perpendicular line and the first point crossed by it range of x values in the low grey-level region. Once found on the two lungs. In the area of the binary image where the threshold (l ) that removes the background from the perpendicular lines do not cross any lungs, a correc- bg the image, a new grey-level histogram H’(x) is generated, tion of the invalid points generated is necessary. u Th s, it considering only the pixels related to the chest. Hence, calculates for each line the distances between the mid- the method is reapplied to the new data to estimate the point and the intersection point and computes their correct threshold for the lung segmentation (l ). Spe- mean value (µ ) and the standard deviation (σ ). We have lung d d cifically, in order to enhance the threshold search, the designed a length filter by defining as invalid the inter - algorithm focuses the analysis on the lower part of H’(x), section points whose distances from the mid-points are since it corresponds to grey values belonging to lungs longer than a specific threshold that we identified as 2* (Additional file 1: Fig. 3). µ − σ . All points, corresponding to longer distances d d After finding appropriate thresholds for each slice with than this value, are deleted and replaced by new ones this strategy, these are used to segment lungs from chest located at the same distances as the previous valid point region. Specifically, a mask is created where pixels with (Fig.  4a). Once all intersection points have been found, intensity above the l are set as white and the remain- the inner curvature is calculated by an interpolating pro- lung ing ones are set as black. Other segmented elements cess. Initially, the algorithm prepares intersection points besides the lungs are removed and morphological opera- by separating them in two subsets: the ones related to the tion of closing are applied to smooth edges and fill holes upper half of the inner contour and those belonging to inside the lungs. the lower part. Additionally, it performs an initial correc- tion, by deleting points whose y locations are in discon- tinuity with y positions of neighboring points, in order Selection of appropriate slice for indices computation to avoid the eventual errors made by previous opera- Finally, before proceeding to inner contour detection, we tions. Then it applies a shape-preserving piecewise cubic created an automatic technique to exclusively select the interpolation method (‘pchip’) with a high sampling rate. slices where lungs are clearly visible. Indeed, we wanted Finally, we obtain a group of closely spaced points both to exclude from further analysis those slices where inner for superior and inferior half of inner contour. After re- chest region segmentation could be complex, due to the combining them in a unique set of points, we can define Trò et al. BMC Medical Imaging (2022) 22:30 Page 8 of 16 holes. Finally, we obtain a binary image representing the inner chest region from which to extract boundary pixel locations for each slice. However, a further correction may be necessary both around the vertebral body, having grey values close both to lungs and to cardiac structure intensities, and around inferior lung area, being often difficult separating pixels belonging to lungs and to thoracic tissue ones (Fig. 5b). We thus designed a method to correct the inner chest Fig. 4 Algorithm for preliminary inner wall contour a Binary image representing lung region. Yellow line represents outer chest contour by comparing consecutive slices thanks to high curvature. In blue there are the perpendicular lines, generated similarity in boundary pixel positions belonging to the every 12th step. In red there are the intersection points resulting lower half of inner contour of our interest. At this step, from recursive algorithm. b Gray-scale image on which inner mask user intervention is required such that correction pro- boundary points are indicated in green, while intersection points cess starts from a slice where inner contour detection found by recursive algorithm in red. This is clearly a rough contour of inner chest, including vertebral body, and thus further corrections are does not present errors. Once first slice is selected, the required algorithm starts the pair-wise comparison among adja- cent slices in both directions, by taking as reference the points belonging to the contour of slice selected. u Th s, the boundary of a mask that isolates the inner thoracic the algorithm computes the distance among them and all region (Fig.  4b). However, this mask also includes the the points belonging to the contour of the adjacent slice, vertebral body and is not accurate in all the slices, mostly which is assumed as incorrect. Then it creates a vector due to bad lung segmentation. For these reasons, it is that, for each point belonging to the contour to correct, necessary to improve the inner chest segmentation with only keeps the minimum distance among all those just further processing. computed. It also calculates the maximum value (d ) max and the standard deviation (σ ) of all minimum distances. Inner chest contour correction After several tests, we established that the algorithm The first step of inner chest contour correction consists must continue only if σ is greater than 1.8. Additionally, in excluding the vertebral body. For doing so, the algo- we identified d -2* σ as threshold value that sepa- max d rithm applies a thresholding method by using the inner rates correct points from incorrect ones. Thus, for each mask found in previous section as a tool to improve seg- incorrect point, the algorithm finds the range that must mentation. After masking out the external chest area, a be deleted, by identifying its extremities in the near- slice-wise threshold is defined to exclude out heart and est points to the correct curve. Then, it replaces them other cardiac elements through histogram partitioning with the points belonging to the correct curve by using method presented in previous section. Indeed, correction a shape-preserving piecewise cubic interpolation method is not possible without masking out cardiac district, due (‘pchip’) (Fig. 5c). Once a new curve is obtained, the algo- to high similarity in grey values between heart and tho- rithm proceeds to the next slice, by taking as reference racic tissue. Thus, the algorithm returns to original chest the just corrected contour. Such an algorithm allows to image, before the application of inner mask, and assigns satisfactorily correct errors in segmentations (Fig. 5d). to background the just segmented pixels belonging to cardiac structures (Fig. 5a). Then, it is able to separate the Thoracic indexes computation inner chest region from the outer chest one, by applying This module aims at computing PE indices used by phy - as threshold the same value found for lung segmentation sicians to classify the severity of patients’ malformation. (l ). To have the inner region as foreground, the com- As mentioned above, among multiple thoracic markers, lung plementary image is computed, and some morphological we focused on the severity (Haller index and Correc- operations are applied to smooth the edges and fill the tion index) (Fig.  6a, b) and deformity (Asymmetry index (See figure on next page.) Fig. 5 Algorithm for inner wall contour correction. a Grey-scale image after masking out cardiac structures in order to create a mask of inner thoracic area. b Grey-scale image with inner chest boundary pixel locations in red. The error appears around inferior lung area, that has grey values close to thoracic tissue ones. c Plot of x and y coordinates corresponding to inner chest boundary pixels. In red there is the reference curve used for correction, while in green the curve that need to be corrected by algorithm. Between two blue arrows there are the points resulting from interpolation process that substitute incorrect ones. d Inner chest boundary pixel locations after correction are represented in red while those before correction in green; in blue there is inner contour of the reference curve, belonging to previous slice Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 9 of 16 Fig. 5 (See legend on previous page.) Trò et al. BMC Medical Imaging (2022) 22:30 Page 10 of 16 minimum x coordinate and the second one as the point at its same y coordinate. Conversely, the first extremity for measuring min APd corresponds to sternum position. We approximate it as the point with maximum y coordi- nate (y values decrease toward the bottom of image) by only considering the range of inner chest contour points between x coordinate positions of two maximum points of outer chest contour upper half. Second extremity cor- responds to the vertebral body position. It is taken as the point with minimum y coordinate by only considering the range of inner chest contour points between position of x coordinates of two maximum points of outer chest contour lower half. Correction index Correction index (iCorrection) is calculated by dividing the amount of defect, measured as the difference between the maximum anteroposterior distance, i.e. the maxi- mum distance between the anterior spine and the anterior Fig. 6 Inner thoracic distances overlaid on slice of maximal sternal portion of the chest (max APd) and the minimum anter- depression. a Distances useful for iHaller computation: transverse oposterior diameter (min APd), to the maximum anter- diameter in red, min APd in yellow. In cyan there is the horizontal line at same y position of vertebral body. b Distances useful for oposterior distance (max APd), multiplied by 100 [37]. iCorrection computation: min APd in yellow, max APd in green. In (max APd − min APd) cyan there is the horizontal line at same y position of vertebral body. iCorrection = ∗ 100 c Distances useful for iAsymmetry computation: right hemithorax min APd APd in blue and left hemithorax APd in magenta. d Distances useful for iFlatness computation: right hemithorax APd in blue and For max APd computation, firstly, the algorithm draws transverse diameter in red a horizontal line at the same y coordinate of vertebral body position, that is assumed as the anterior spine posi- tion. Then, it identifies two points on inner chest contour and Flatness index) ones (Fig.  6c, d). The algorithm only at the same x coordinate of two maximum points of outer works on the first slice of images processed in the previ - chest contour. The latter are assumed as the positions of ous module. Indeed, it corresponds to the slice selected right and left anterior portion of the chest. Thus, for each by the user or to the first following one where inner chest point it computes the distances between them and the contour can be detected. horizontal line and gets the maximum diameter between Once inner distances and thoracic indices are com- the two distances. puted, the framework saves their results along with the new pathological marker obtained in Depression quan- Asymmetry index tification module in an Excel file, located in the same Asymmetry index (iAsymmetry) is calculated by divid- folder as patient’s images. Each quantified distance in fol - ing the longest anteroposterior distance of the right chest lowing computations has been multiplied by ‘pixel spac- wall (right hemithorax APd) to the longest anteroposte- ing’ attribute to have measures in mm. rior distance of the left chest wall (left hemithorax APd), multiplied by 100 [38]. Haller index Haller index (iHaller) is calculated by dividing the trans- right hemithorax APd iAsymmetry = ∗ 100 verse diameter, i.e., the widest horizontal distance of the left hemithorax APd inside of the ribcage, to the minimum anteroposterior diameter (min APd), i.e., the shorter distance between Right hemithorax APd’s extremities are identified as the vertebral body and the sternum [12]. the points on inner chest contour at the same x coor- dinate of first maximum point, as it is located in right transverse diameter iHaller = hemithorax. Right hemithorax APd’s bounds are identi- min APd fied as the points on inner chest contour at the same x coordinate of second maximum point, as it is situated in As regards transverse diameter, the algorithm identifies left hemithorax. its first extremity as the point on inner chest contour with Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 11 of 16 Flatness index msec (maximum) and a 32-element cardiac phased- Flatness index (iFlatness) is computed by dividing array coil for signal reception and cardiac synchroniza- the transverse diameter of the thorax to the longer of tion (with “retrospective gating” technique). Our MRI the two maximum anteroposterior diameters of the protocol borrowed cardiac gating and breath-holding right (right hemithorax APd) and left hemithorax (left techniques as well as specific sequences from CMRI, hemithorax APd) [13]. As all the distances have been in order to overcome motion-related artifacts and to already found, the algorithm can proceed with Flatness inspect with further detail cardiovascular morphology. index computation, as follow: The MR acquisition setting thus included scout images and Steady State Free Precession (SSFP) images in transverse diameter iFlatness = axial, coronal and sagittal planes, acquired at the end of max righthemithoraxAPd; lefthemithoraxAPd expiratory phases. Specifically, the SSFP sequence was a Gradient-Echo sequence, named Balanced Turbo Field Echo-Breath Hold (BTFE-BH). Total scanning time was User’s correction 5–8 min approximately. As mentioned above, the algorithm does not always perform indices computation on the same slice selected Results by user, due to its inability to segment images where Out of 50, just three subjects (2 male and 1 female) have different chest areas have similar grey values. In these been excluded from our analysis since characterized by cases, it selects the first following slice, where inner extremely low contrast images that the algorithm could chest contour detection can be performed. We noticed not process. We can thus conclude that, provided suffi - that by going through consecutive slices some inner cient contrast in the input raw image, proposed method distances maintain their value constant, while others, mantains its reliability and accuracy for the whole cohort specifically min APd and max APd are more likely to ander analysis. vary. For this reason, in the algorithm we add the pos- Our image processing pipeline has then been quan- sibility of user’s intervention when the slice is different titatively evaluated through comparison with manual from the one selected. Specifically, the user is asked procedure. As mentioned in previous section, the slice to insert two points on the image, useful for min APd selected by user for indices computation is often dif- and max APd calculation: sternum position and verte- ficult to segment due to similar grey values of different bral body position. Thus, the algorithm recomputes the thoracic regions, so that algorithm automatically picks indices by considering the modifications on these two the first following slice, where inner chest contour detec - inner distances. Finally, two sets of results are obtained: tion can be performed. Specifically, out of the 47 patients the ones calculated on the slice picked by the algorithm processed by algorithm, the latter was able to use the and those obtained on the same slice selected after cor- same slice as the one selected in 24 patients, while in the rection of sternum and vertebral body points. remaining it selected another scan. We thus separated the patients in two groups, depending on whether the Modules validation indices computation was performed on the same slice Current methodological framework has been devel- selected by user (group 1) or it was executed on a differ - oped from a small subset counting 5 subjects. In order ent slice picked by the algorithm (group 2). to test the overall quality of our algorithm, we extended its application to other 45 pediatric patients affected by Automatic framework agrees with manual procedure Pectus Excavatum from Gaslini Children’s Hospital, in for indexes computation Genoa, for a total of 50-subjects dataset. In the absence of a ground truth to test performance Additionally, two expert radiologists manually per- against, the accuracy of the thoracic indices resulting formed double-blind thoracic indices computation, from algorithm was evaluated by comparing them to as they routinely do in the clinical setting. The group results obtained by manual measures performed by two of patients consisted in 41 males and 9 females aged expert radiologists (through a double-blind analysis). 13.5 ± 2.78 (mean ± SD), age range 5–18 years. Each of Table  1 shows the results of patients belonging to them underwent MRI examination, in order to estab- group 1. Results of inner thoracic distances show a good lish the severity of malformation and thus the best agreement between measures obtained by the 2 read- treatment strategies. MRI examinations were per- ers and the algorithm. Naturally, the difference is higher formed on a 1.5 Tesla MR scanner (Achieva, Philips by comparing manual results to automatic ones, as it is Healthcare, Cleveland, OH, USA), equipped with 66 shown by a greater mean standard deviation. We can mT/m gradients (maximum), a slew rate of 180 mT/m/ notice that transverse diameter, min APd and max APd Trò et al. BMC Medical Imaging (2022) 22:30 Page 12 of 16 Table 1 Average inner thoracic distances and thoracic indices along with relative mean standard deviation between 2 readers and mean standard deviation (std) among readers and algorithm in case of appropriate user selection of main slice for indices computation Reader 1 Reader 2 Algorithm Std between 2 Std among readers readers and algorithm Thoracic distances (cm) Transverse diameter 24.1 ± 2.6 24.2 ± 2.6 24.4 ± 2.9 0.17 0.30 Min APd 4.9 ± 1.3 5.0 ± 1.4 5.1 ± 1.4 0.18 0.30 Max APd 7.2 ± 0.94 7.6 ± 0.93 7.5 ± 0.95 0.31 0.34 Right hemithorax APd 12.4 ± 1.2 12.6 ± 1.3 12.2 ± 1.2 0.24 0.37 Left hemithorax APd 12.3 ± 1.2 12.6 ± 1.1 11.9 ± 1.2 0.30 0.47 Thoracic indices Haller index 5.3 ± 1.9 5.3 ± 1.9 5.1 ± 1.6 0.31 0.34 Correction index (%) 32.5 ± 13.9 35.2 ± 14.1 32.6 ± 13.6 2.7 3.6 Asymmetry index (%) 101.7 ± 6.7 99.8 ± 5.2 102.8 ± 8.2 2.2 3.2 Flatness index 1.9 ± 0.19 1.9 ± 0.17 2.0 ± 0.20 0.038 0.057 are computed by the algorithm in a comparable way as readers. As we mentioned above, we noticed that some results obtained manually. Contrariwise, right hemitho- distances remained almost constant by measuring them rax APd and left hemithorax APd are characterized by a on consecutive slices. Contrariwise, two distances, spe- higher variability that, however, is stronger also between cifically min APd and max APd, showed more variability the 2 readers. Obviously, inner distances affect the results among consecutive slices. Consequently, we decided to of thoracic indices. Specifically, we notice that Haller apply a correction factor to these measurements, to be index and Flatness index results are comparable, whereas able to compare the algorithm results to the ones com- the differences between manual and automatic computa - puted by the readers. Specifically, as we observed that tion increase by considering Correction index and Asym- algorithm tends to overestimate both min APd and max metry index. APd, we subtracted to them a corrective factor that we identified as the mean standard deviation between read - Inner thoracic distances and thoracic indices in case ers and algorithm (0.50 in both cases). Table  2 shows of matching slice selection the results of patients belonging to group 2, after the Regarding the results belonging to group 2, the measure- just mentioned correction of the two inner distances. ments obtained from the algorithm were performed on The same considerations made for inner thoracic dis - a different slice compared to the one analyzed by the 2 tances results belonging to group 1 apply also in this case. Table 2 Average inner thoracic distances and thoracic indices along with relative mean standard deviation (std) between 2 readers and mean standard deviation among readers and algorithm in case of failed user selection of main slice for indices computation Reader 1 Reader 2 Algorithm Std between 2 Std among readers readers and algorithm Thoracic distances (cm) Transverse diameter 23.2 ± 1.6 23.3 ± 1.6 23.5 ± 1.7 0.06 0.26 Min APd 3.9 ± 1.4 4.0 ± 1.6 4.2 ± 1.6 0.16 0.36 Max APd 6.5 ± 0.85 6.9 ± 0.83 6.8 ± 0.94 0.25 0.35 Right hemithorax APd 11.3 ± 1.1 11.5 ± 1.2 11.2 ± 1.4 0.17 0.39 Left hemithorax APd 11.9 ± 0.96 12.0 ± 1.0 11.7 ± 1.1 0.14 0.33 Thoracic indices Haller index 7.2 ± 4.0 7.2 ± 4.3 7.0 ± 4.6 0.23 0.63 Correction index (%) 42.1 ± 17.9 42.5 ± 19.9 39.1 ± 18.8 2.5 4.7 Asymmetry index (%) 95.1 ± 5.5 96.0 ± 6.8 96.5 ± 8.1 1.4 4.0 Flatness index 2.0 ± 0.16 1.9 ± 0.17 2.0 ± 0.16 0.023 0.050 Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 13 of 16 However, by observing thoracic indices results, we notice the correlation increases with severity indices, i.e. Haller a higher variability among readers and algorithm than the index, and slightly more with Correction index, that are one found in Table 1. The reason is mainly due to the use the most used by physicians to assess the PE malforma- of a different slice for indices computation. Furthermore, tion in a quantitative way. As it is shown in Fig. 7c, small there are more severe cases of PE among patients belong- depressions (low VCI) correspond to low Haller indices. ing to this group. This aspect could be another cause for However, if Haller index begins to increase the linear the higher variability in indices, specifically Haller index. relation tends to disappear, as the same iHaller corre- Indeed, we noticed that differences among reader and sponds to different degrees of depression. Same behavior algorithm results increase when the min APd assumes is visible in Fig.  7d that shows the correlation between low values, as it is placed at the denominator in the index VCI and iCorrection. However, it should be noted that calculation formula. u Th s, variability is higher for high severe cases of PE with high iHaller and iCorrection are Haller indices rather than lower ones. few among all the patients analyzed. Thus, the lack of linear tendency could be caused by a limited number of Inner thoracic distances and thoracic indices in case cases with high degree of PE severity. of not matching slice selection Additional file  1: Fig. 4 displays scatter charts represent- Person correlation between Volumetric Correction Index ing comparison among results obtained by readers and and other traditional indices algorithm for each thoracic index, belonging to patients See Table 3. of group 1. Finally, the average time necessary to perform tra- Discussion ditional indices computation on a single patient was We introduce a set of tools to aid the pre- and post-sur- 50  s and 3  min 45  s for automatic (on a standard Win- gery assessment of PE patients. We opted for developing dows workstation with i7-core and 8  GB RAM) and this algorithm within an existing software rather than a manual processing, respectively. However, if radiologists new stand-alone tool in order to ensure later extensibility are beginners, the time could significatively increase, across different centers. even rising to twice the value indicated for manual The set of algorithms have been tested both qualita - computation. tively and quantitatively through a cohort of 50 pediat- ric patients with varying age, sex and disease severity. In New volumetric index as a promising marker of PE severity our study, we were able to show that automatic results Finally, the new index calculated by our algorithm exclu- obtained by our algorithm are comparable with the ones sively on male subjects (n = 39), named VCI, was com- manually computed by expert radiologist. The proposed pared to all the thoracic indices, in order to evaluate its algorithm offers different advantages First of all, it gives feasibility for quantitative evaluation of PE. Specifically, physicians an accurate tool not subjected to individual we calculated statistical Pearson correlation between VCI interpretation or errors and represents a useful sup- and other traditional indices, as shown in Table  3. Fur- port in establishing proper treatment decision, includ- thermore, scatter plot analysis between new pathological ing the need for surgical correction of malformation. marker and other indices are shown in Fig. 7. Moreover, it ensures a faster processing time compared The results of correlation show a very low correlation to manual measurements, which gets relevant in case of between VCI marker and Asymmetry index and Flatness large datasets and radiologists with limited experience. index (Fig.  7a, b). We expected this behavior since both Furthermore, for now limited to the male subset of our indices do not quantify the severity of depression but the cohort, we suggested a new pathological marker to better degree of chest asymmetry and flatness. Contrariwise, quantify the depression caused by PE: the VCI. Indeed, indexes used so far are based on linear measurements of chest diameters, but they do not evaluate the chest in the Table 3 Result of Pearson correlation between traditional tridimensional aspect of the deformity, which has clinical indices and new pathological marker computed by algorithm implications. A patient with a deep but very localized PE Indices comparison Pearson could have a worse Haller index or Correction index than correlation coefficient another patient with a less severe but more extended PE, even if the real impact of the deformity and the compres- iHaller – VCI (%) 0.79 sion on lung and heart could be globally worse in the sec- iCorrection (%) – VCI (%) 0.81 ond patient, due to the diffuse PE. Therefore, an index iAsymmetry (%) – VCI (%) 0.062 which considers all the missing volume of the thorax iFlatness – VCI (%) 0.22 and not only measures the severity of PE at a single level Trò et al. BMC Medical Imaging (2022) 22:30 Page 14 of 16 Fig. 7 Linear relationship of new volumetric index with existing clinical markers a Linear Regression between VCI and iHaller. b VCI and iCorrection. c VCI and iFlatness. d VCI and iAsymmetry. Estimates for the slope and intercept of the linear equation as well as R are reported for each measure could overcome the limits of traditional thoracic indices, By analyzing patient-wise results, we could notice that such as the dependence on the slice selected for measure- accuracy of algorithm outcomes is strongly dependent ments or chest shape. Theoretically, it could also have a on quality of MR images acquired. Thus, optimization of better clinical correlation than the current indexes. Nev- acquisition setting would lead to higher-quality images ertheless, further investigations are required to prove the and thus improve pipeline’s performance without the clinical relevance of VCI and incorporate it among the need of further corrections beside main modules. This clinical and radiological parameters considered in the improvement may also allow to properly handle with the decision regarding surgical indication. quantification of depression volume for female patients. Our work has made a relevant contribution to the lit- Indeed, one current limitation of this study is exclu- erature. Indeed, the other studies focused on automatic sion of female patients from computation of newly pro- or semi-automatic quantification of the markers of chest- posed VCI marker, given the higher variability in chest wall deformity are exclusively limited to CT scans [14, 15, shape caused by differential breast growth. A similar 36, 39, 40]. situation may apply for overweight patients with rel- The peculiarity of our algorithm is that it works on MR evant gynecomastia. However, both target patients images. The adoption of MRI in the evaluation of this substantially represent outliers for this kind of condi- condition is relatively recent since, despite their non- tion, mainly affecting under- or normal-weight male invasiveness, MRI scans are more complex to process adolescents [41–43]. As a result, current algorithm with automatic segmentation methods than standard CT already performs successfully for most PE candidates, ones [21, 23, 29, 30]. Tr ò et al. BMC Medical Imaging (2022) 22:30 Page 15 of 16 Abbreviations being other cathegories statistically negligible in their CT: Computed tomography; CMR: Cardiac magnetic resonance; DICOM: amount. Digital imaging and communications in medicine; GUI: Graphical user inter- Another possible limitation to the accuracy in calcu- face; iAsymmetry: Asymmetry index; iCorrection: Correction index; iFlatness: Flatness index; iHaller: Haller index; max APd: Maximum anterior–posterior lation of VCI is the thickness of subcutaneous fat tis- diameter; min APd: Minimum anterior–posterior diameter; MRI: Magnetic sue in pectoral region. However, as anticipated, the vast resonance imaging; PE: Pectus excavatum; VCI: Volume Correction Index. majority of PE patients are very slim so the influence of the subcutaneous tissue is in our opinion neglectable in Supplementary Information most patients. Calculating VCI at the level of cartilage The online version contains supplementary material available at https:// doi. and bones instead of the inner skin level could potentally org/ 10. 1186/ s12880- 022- 00754-0. overcome this limitation. One last potential future development of present algo- Additional file 1: Figure 1. Graphical User Interface for slices selection. Figure 2. Image border correction. Figure 3. Histogram partitioning for rithm may include the computation of cardiac indices. lung segmentation. Figure 4. Comparison between double-blind manual Specifically, it may be useful to develop a new method for measurements and automatic algorithm for computation of thoracic quantification of cardiac compression caused by PE mal - indexes for patients of group 1 a. formation, in order to compare it with the new pathologi- cal marker proposed. Acknowledgements The authors thank all the Technicians and Nurses of our Cardiac MRI Unit for their contribution, and in general to their dedication for Cardio-Radiologic Group since the foundation in 2007 Conclusions In this work we present a piece of software specifically Authors’ contributions SM, RT and MF designed, coded, tested the software tools and performed the designed to support radiologists in diagnosis and best analyses. NS, MT, and VS provided the data and validated as expert the results. personalized treatment choice for patients affected by SM, RT, NS, MT, MF wrote the manuscript. All authors read and approved the PE condition. Indeed, our study proved its reliability and final manuscript. robustness in processing a discrete number of MR images Funding ranging across different degrees of PE severity. Our tool None. significantly eases assessment of pathology by improving Availability of data and materials accuracy of thoracic distances and subsequent clinical The dataset supporting the conclusions of this article is not publicly available indexes beyond subjectivity inherent to manual interven- due to privacy restrictions of clinical data imposed by the Gaslini Hospital’s tion, and by reducing the time required for computation administration but is available from the corresponding author on reasonable request. The Software has been released under MIT License. Code for our of these markers as well. pipeline is available at https:// github. com/ rosel la1234/ PE_ pipel ine. Current Given the relatively high incidence of this disease software has been designed and tested under MATLAB 2020a. A proprietary (1:400 live births), disposing of a novel semi-automatic license is required for using MATLAB . supportive tool enriched with an easily extensible, user- friendly interface may have a substantial clinical impact. Declarations Finally, formulation of a new relevant marker for PE scor- Ethics approval and consent to participate ing paves the way for exploring new strategies for PE This research study was conducted retrospectively from data obtained for clin- assessment. ical purposes. An IRB official waiver of ethical approval was granted from the IRB of IRCCS Istituto Giannina Gaslini. IRB number: 8/2020. Written informed consent was obtained from all patients or their guardians for each participant prior to examination in accordance with the Declaration of Helsinki. Availability and requirements Consent for publication The authors affirm that human research participants and their parents • Project name: e.g. PE_pipeline provided written informed consent for publication of all the images in the • Project home page: https:// github. com/ rosel la1234/ manuscript. PE_ pipel ine/ Competing interests • Operating system(s): Platform independent The other authors declare no competing interests. • Programming language: MATLAB Author details • Other requirements: none Department of Informatics, Bioengineering Robotics and System Engineering • License: e.g. MIT 2 (DIBRIS), University of Genoa, Viale Causa 13, 16143 Genova, Italy. Complex • Any restrictions to use by non-academics: e.g. licence Operative Radiology Unit, IRCCS Giannina Gaslini, Genova, Italy. Depar tment of Health Sciences (DISSAL), Radiodiagnostics, University of Genoa, Via A. Pas- needed for MATLAB tore 1, 16132 Genova, Italy. Pediatric Thoracic and Airway Surgery Unit, IRCCS Giannina Gaslini, Genova, Italy. Trò et al. BMC Medical Imaging (2022) 22:30 Page 16 of 16 Received: 10 September 2021 Accepted: 8 February 2022 23. Marcovici PA, LoSasso BE, Kruk P, Dwek JR. MRI for the evaluation of pectus excavatum. Pediatr Radiol. 2011. 24. Tseng WYI, Su MYM, Tseng YHE. Introduction to cardiovascular magnetic resonance: Technical principles and clinical applications. Acta Cardiol Sin. 2016. 25. Deviggiano A, Carrascosa P, Vallejos J, Bellia-Munzon G, Vina N, Rodríguez- Granillo GA, et al. Relationship between cardiac MR compression clas- References sification and CT chest wall indexes in patients with pectus excavatum. 1. Brochhausen C, Turial S, Müller FKP, Schmitt VH, Coerdt W, Wihlm JM, et al. Journal of Pediatric Surgery. 2018. Pectus excavatum: history, hypotheses and treatment options. Interact 26. Viña NA, Carrascosa P, Mogensen VC, Deviggiano A, Bellia-Munzon G, Cardiovasc Thoracic Surg. 2012. Martinez-Ferro M, et al. Evaluation of pectus excavatum indexes during 2. Fonkalsrud EW. Current management of pectus excavatum. World J Surg. standard cardiac magnetic resonance: potential for single preoperative tool. Clin Imaging. 2019. 3. Kelly RE, Cash TF, Shamberger RC, Mitchell KK, Mellins RB, Lawson ML, 27. Narayan RL, Vaishnava P, Castellano JM, Fuster V. Quantitative assessment et al. Surgical repair of pectus excavatum markedly improves body image of right ventricular function in pectus excavatum. J Thoracic Cardiovasc and perceived ability for physical activity: multicenter study. Pediatrics. Surg. 2012. 28. Saleh RS, Finn JP, Fenchel M, Moghadam A, Krishnam M, Abrazado M, 4. Rodriguez-Granillo GA, Raggio IM, Deviggiano A, Bellia-Munzon G, Capu- et al. Cardiovascular magnetic resonance in patients with pectus excava- nay C, Nazar M, et al. Impact of pectus excavatum on cardiac morphol- tum compared with normal controls. J Cardiovasc Magn Reson. 2010. ogy and function according to the site of maximum compression: effect 29. Humphries CM, Anderson JL, Flores JH, Doty JR. Cardiac magnetic reso- of physical exertion and respiratory cycle. Eur Heart J Cardiovasc Imaging. nance imaging for perioperative evaluation of sternal eversion for pectus excavatum. Eur J Cardio-thoracic Surg. 2013. 5. Jaroszewski DE. Physiologic implications of pectus excavatum. J Thoracic 30. Lai K, Colen TM, Dicken BJ. Integrating cardiac MRI into pre-operative Cardiovasc Surg. 2017. planning for patients with pectus excavatum and right ventricular dys- 6. Sigalet DL, Montgomery M, Harder J, Wong V, Kravarusic D, Alassiri A. function. J Pediatr Surg Case Rep. 2019. Long term cardiopulmonary effects of closed repair of pectus excavatum. 31. Stagnaro N, Trocchio G, Torre M, Rizzo F, Martucciello G, Marasini M. In: Pediatric Surgery International. 2007. Cardiovascular MRI assessment of pectus excavatum in pediatric patients 7. Kotagal M, Swanson JW, Avansino JR. Correlating Haller Index and cardio- and postoperative simulation using vacuum bell. J Pediatr Surg. 2021. pulmonary disease in pectus excavatum. In: Chest Wall Deformities. 2017. 32. Silva B, Pessanha I, Correia-Pinto J, Fonseca JC, Queiros S. Automatic 8. Nuss D, Kelly RE. Indications and technique of nuss procedure for pectus assessment of Pectus Excavatum severity from CT images using deep excavatum. Thoracic Surgery Clinics. 2010. learning. IEEE J Biomed Health Inf. 2021. 9. Jayaramakrishnan K, Wotton R, Bradley A, Naidu B. Does repair of pectus 33. Martinez-Ferro M. Indexes for pectus deformities. In: Chest wall deformi- excavatum improve cardiopulmonary function? Interactive Cardiovascu- ties and corrective procedures. 2016. p. 35–60. lar and Thoracic Surgery. 2013; 34. Otsu, N. A threshold selection method from gray-level histograms. IEEE 10. Abid I, Ewais MM, Marranca J, Jaroszewski DE. Pectus excavatum: a review Trans Syst Man Cybern 1996. of diagnosis and current treatment options. Journal of the American 35. Balan AGR, Traina AJM, Ribeiro MX, Marques PMA, Traina C. Smart histo- Osteopathic Association. 2017. gram analysis applied to the skull-stripping problem in T1-weighted MRI. 11. Sarwar ZU, DeFlorio R, O’Connor SC. Pectus excavatum: current imaging Comput Biol Med. 2012. techniques and opportunities for dose reduction. Seminars in Ultrasound, 36. Papp L, Juhasz R, Travar S, Kolli A, Sorantin E. Automatic detection and CT and MRI. 2014. characterization of funnel chest based on spiral CT. J X-ray Sci Technol. 12. Haller JA, Kramer SS, Lietman SA. Use of CT scans in selection of patients for pectusexcavatum surgery: a preliminary report. J Pediatr Surg. 1987. 37. st. Peter SD, Juang D, Garey CL, Laituri CA, Ostlie DJ, Sharp RJ, et al. A 13. Choi JH, Park IK, Kim Y T, Kim WS, Kang CH. Classification of Pectus Excava- novel measure for pectus excavatum: The correction index. In: Journal of tum according to objective parameters from chest computed tomogra- Pediatric Surgery. 2011. phy. Ann Thoracic Surgery. 2016; 38. Lawson ML, Barnes-Eley M, Burke BL, Mitchell K, Katz ME, Dory CL, et al. 14. Kim M, Lee KY, Park HJ, Kim HY, Kang EY, Oh YW, et al. Development of Reliability of a standardized protocol to calculate cross-sectional chest area new cardiac deformity indexes for pectus excavatum on computed and severity indices to evaluate pectus excavatum. J Pediatr Surg. 2006. tomography: feasibility for pre- and post-operative evaluation. Yonsei 39. Kim HC, Park HJ, Nam KW, Kim SM, Choi EJ, Jin S, et al. Fully automatic Med J. 2009. initialization method for quantitative assessment of chest-wall deformity 15. Lai L, Cai S, Huang L, Zhou H, Xie L. Computer-aided diagnosis of pectus in funnel chest patients. Med Biol Eng Comput. 2010. excavatum using CT images and deep learning methods. Sci Rep. 2020. 40. Barbosa D, Queirós S, Rodrigues N, Correia-Pinto J, Vilaça J. Semi- 16. Chu ZG, Yu JQ, Yang ZG, Peng LQ, Bai HL, Li XM. Correlation between automatic 3D segmentation of costal cartilage in CT data from Pectus sternal depression and cardiac rotation in pectus excavatum: evaluation Excavatum patients. In: Medical Imaging 2015: Image Processing. 2015. with helical CT. Am J Roentgenol. 2010. 41. Nguyen TM, Le VT, Nguyen HU, Pham HL, Phung HSD, Vu NT, et al. An 17. Rattan AS, Laor T, Ryckman FC, Brody AS. Pectus excavatum imaging: initial 5-year single-center experience of 365 patients undergoing the enough but not too much. Pediatr Radiol. 2010. video-assisted thoracoscopic surgery for nuss procedure for pectus 18. Dore M, Triana Junco P, Bret M, Gomez Cervantes M, Muñoz Romo M, excavatum in resource-scare setting. Front Surg. 2021. Jimenez Gomez J, et al. Advantages of cardiac magnetic resonance imag- 42. Kim JJ, Park HJ, Park JK, Cho DG, Moon SW. A study about the costoclavic- ing for severe Pectus Excavatum assessment in children. Eur J Pediatr ular space in patients with pectus excavatum. J Cardiothorac Surg. 2014. Surg. 2018. 43. Akkaş Y, Gülay Peri N, Koçer B, Gülbahar G, Baran Aksakal FN. The preva- 19. Lollert A, Funk J, Tietze N, Turial S, Laudemann K, Düber C, et al. Morpho- lence of chest wall deformity in Turkish children. Turk J Med Scie. 2018. logic assessment of thoracic deformities for the preoperative evaluation of pectus excavatum by magnetic resonance imaging. Eur Radiol. 2015. 20. lo Piccolo R, Bongini U, Basile M, Savelli S, Morelli C, Cerra C, et al. Chest Publisher’s Note fast MRI: an imaging alternative on pre-operative evaluation of Pectus Springer Nature remains neutral with regard to jurisdictional claims in pub- Excavatum. J Pediatr Surg. 2012. lished maps and institutional affiliations. 21. Sun J, Chen C, Peng Y, Zhang Y, Tian H, Yu J, et al. Comparison of magnetic resonance imaging and computed tomography to measure preoperative parameters of children with pectus excavatum. Pediatr Investig. 2019. 22. Birkemeier KL, Podberesky DJ, Salisbury S, Serai S. Limited, fast magnetic resonance imaging as an alternative for preoperative evaluation of pec- tus excavatum: a feasibility study. J Thoracic Imaging. 2012.

Journal

BMC Medical Imaging – Springer Journals

Published: Feb 20, 2022

Keywords: Pectus Excavatum; Magnetic resonance imaging; Image processing pipeline

There are no references for this article.