Get 20M+ Full-Text Papers For Less Than $1.50/day. Subscribe now for You or Your Team.

Learn More →

Automated Quantitative Assessment of Coronary Calcification Using Intravascular Ultrasound

Automated Quantitative Assessment of Coronary Calcification Using Intravascular Ultrasound Ultrasound in Med. & Biol., Vol. 46, No. 10, pp. 28012809, 2020 Copyright © 2020 The Author(s). Published by Elsevier Inc. on behalf of World Federation for Ultrasound in Medicine & Biology. This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/) Printed in the USA. All rights reserved. 0301-5629/$ - see front matter https://doi.org/10.1016/j.ultrasmedbio.2020.04.032 Original Contribution AUTOMATED QUANTITATIVE ASSESSMENT OF CORONARY CALCIFICATION USING INTRAVASCULAR ULTRASOUND TAGGEDPSHENGNAN LIU,* TARA NELEMAN,* ELINE M.J. HARTMAN,* JURGEN M.R. LIGTHART,* ,y,z KAREN T. WITBERG,* ANTONIUS F.W. VAN DER STEEN,* JOLANDA J. WENTZEL,* JOOST DAEMEN,* and GIJS VAN SOEST*TAGGEDEND * Department of Cardiology, Erasmus University Medical Center, Rotterdam, The Netherlands; Department of Imaging Physics, Faculty of Applied Sciences, Delft University of Technology, The Netherlands; and Shenzhen Institutes of Advanced Technologies, Shenzhen, China Abstract—Coronary calcification represents a challenge in the treatment of coronary artery disease by stent place- ment. It negatively affects stent expansion and has been related to future adverse cardiac events. Intravascular ultrasound (IVUS) is known for its high sensitivity in detecting coronary calcification. At present, automated quan- tification of calcium as detected by IVUS is not available. For this reason, we developed and validated an optimized framework for accurate automated detection and quantification of calcified plaque in coronary atherosclerosis as seen by IVUS. Calcified lesions were detected by training a supported vector classifier per IVUS A-line on manually annotated IVUS images, followed by post-processing using regional information. We applied our framework to 35 IVUS pullbacks from each of the three commonly used IVUS systems. Cross-validation accuracy for each system was >0.9, and the testing accuracy was 0.87, 0.89 and 0.89 for the three systems. Using the detection result, we pro- pose an IVUS calcium score, based on the fraction of calcium-positive A-lines in a pullback segment, to quantify the extent of calcified plaque. The high accuracy of the proposed classifier suggests that it may provide a robust and accurate tool to assess the presence and amount of coronary calcification and, thus, may play a role in image- guided coronary interventions. (E-mail: g.vansoest@erasmusmc.nl) © 2020 The Author(s). Published by Elsevier Inc. on behalf of World Federation for Ultrasound in Medicine & Biology. This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/). Key Words: Calcified plaque, Intravascular ultrasound, Automated quantification, Intravascular imaging, Coro- nary artery disease. INTRODUCTION its rigidity, circumferential calcification may prevent full dilation of the stent, which may lead to stent underexpan- Coronary artery disease, the most common heart disease, sion, associated with increased risk of target vessel failure is caused by a long-term accumulation of atherosclerotic (Witzenbichler et al. 2014). Specialized techniques, such plaque in the intima of the arterial vessel wall as cutting balloons, rotational atherectomy or intracoro- (Falk 2006; Hong 2010). As plaques grow, they may nary lithotripsy (Sharma et al. 2019), can be used to pre- impinge on the free coronary lumen, causing a stenosis pare calcified plaques to enable complete stent expansion. that limits blood flow to the myocardial territory served The beneficial effects of these plaque modification techni- by the coronary artery. Severe or acute coronary artery ques depend directly on the extent and severity of calcifi- disease is frequently treated by stent implantation in a cation, so detailed knowledge is needed to guide the procedure called percutaneous coronary intervention choice of lesion preparation method (Wijns et al. 2015). (PCI). Atherosclerotic plaques are frequently heteroge- Large calcified plaques can often be unambiguously neous in their composition, and typically consist of identified in intravascular ultrasound (IVUS) images fibrous, lipid-rich and calcified tissue. because of their high reflection of and low penetration by The presence of calcium, in particular, can hamper ultrasound signals (Pu et al. 2014; Mintz and Guagliumi the feasibility of PCI (Hoffmann et al. 1998). Because of 2017). As illustrated in Figure 1, coronary calcium is characterized as a narrow band with high echo intensities, with a dark shadow behind it. Current IVUS software Address correspondence to: Gijs van Soest, PO Box 2040, 3000 does not allow automated calcium quantification. CA Rotterdam, The Netherlands. E-mail: g.vansoest@erasmusmc.nl 2801 2802 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 Fig. 1. Three images in gray scale from Infraredx (a), Volcano (b) and Boston Scientific (c). The calcified plaque is marked with a red arch. Because of the presence of inherent speckle, various calcified plaque load and compare it with the calcium system- and anatomy-dependent artifacts, limited resolu- score based on manual labels. tion and image contrast, automated structure detection in IVUS remains a challenging task (Katouzian et al. 2012). METHODS Simple approaches, using thresholding (Kim et al. 2014) and adaptive thresholding, have been reported IVUS data (Dos Santos Filho et al. 2007). A more complex approach IVUS pullback data sets, acquired in native coro- to segmentation of the leading edge of calcified plaque nary arteries, were extracted from the clinical database combining Rayleigh mixture models, Markov random of the Department of Cardiology, Erasmus MC. The fields, graph searching and prior knowledge, has been data sets were anonymized and contained no identify- reported (Gao et al. 2014). After detection of the intima ing information; all selected patients consented to the and the mediaadventitia borders, calcified regions were use of their data in retrospective studies. All data in detected using a Bayesian classifier (Taki et al. 2008). this study were collected as part of routine clinical Two recent studies have reported the detection of calcifi- care. Consequently, institutional review board cation per frame using a deep learning network approval and individual patient consent are not (Balocco et al. 2018; Sofian et al. 2018). required under Dutch law. We selected 105 pullbacks, In this work, we present a framework for accurate 35 each from three commonly used systems: Infraredx automated detection and quantification of the extent (40 MHz, TVC NIRS Catheter System, Infraredx Inc., of calcification in coronary atherosclerosis as seen by Burlington, MA, USA), Volcano (20 MHz, Eagle Eye IVUS. Without multistep pre-processing, extraction Platinum Rx, and ST Rx, Digital IVUS Catheters, Vol- of complex features or design of a deep learning cano Corp., Rancho Cordova, CA, USA) and Boston approach to perform the task, we found that accurate Scientific (40 MHz, Atlatis SR Pro and OptiCross Cor- classification can be achieved by applying a kernel- onary Imaging Catheter, Boston Scientific Corp., based support vector classifier on simple statistical Natick, MA, USA). Example images from the three features, originating from the imaging physics, systems are provided in Figure 1. extracted directly from un-processed images. Using An overview is given in Table 1. Pullbacks, stored data from three commonly used IVUS systems, we in DICOM format, were manually annotated by trained trained the classifier for recognition of calcified pla- experts (E.M.J.H. and T.N.) at 1-mm intervals. Centering ques per A-line. Using the detection result, we pro- at the imaging catheter, we divided and labeled each pose an IVUS calcium score (ICS) to evaluate the frame as a calcified or non-calcified pie sector (Fig. 2). Table 1. Data description Vendor Population Pullback Frame rate (fs/s) Pullback speed (mm/s) Training Test Infraredx 34 35 16 0.5 31 4 Volcano 35 35 30 0.5 31 4 Boston Scientific 35 35 30 0.5 31 4 Automated IVUS assessment of coronary calcification S. LIU et al. 2803 Fig. 2. Pullback stacks were labeled in Cartesian coordinates and converted to polar coordinates. ‘P’ marks the pull- backs, and ‘L’ marks the label stack. In the zoomed-in red rectangular box, a scale bar is given indicating the pullbacks were labeled every 1 mm. The labeled stacks in the right panels are shown with transparency. Frames (in Cartesian coordinates) and their correspond- hyperplane, which can separate only linearly separable ing labels were converted into polar coordinates, such clusters. To deal with data that are not linearly separable, that each A-line was binary labeled as 1 (calcified) or 0 a non-linear kernel needs to be introduced. When little is (non-calcified). known about the structure of data, a Gaussian RBF ker- nel is a robust choice, assuming only general smoothness (Smola and Sch€olkopf 2004). Identifying features Each A-line is characterized by a set of statistical Calcified plaque can be confidently classified by features x ,and has a binary label y identifying it as calci- i i human observers. The A-line profile itself is quite vari- fied or not. For M labeled A-line distributions, {(x ,y )| i i able though, within each (calcified or non-calcified) cate- y 2 {0, 1}, i 2 {0, , M}}, a classifier gory. By comparing the A-line amplitude statistics, we "# observed that the typical combination of a thin reflection fxðÞ ¼ sgn a y KxðÞ ; x þ b and dorsal shadow of calcium in IVUS may be used as a i i i i¼1 discriminating characteristic, whereas an A-line scan of soft tissue contains many gray values and most ampli- was trained to optimize the problem (Guyon et al. 1993) tude values in a line sampling calcium are low (lumen T T T and shadow), with a few very high ones (calcium bor- min a Qae a such that y a ¼ 0; ð1Þ der). To build an unbiased classifier, we decided to train 0a C; i ¼ 1;⋯;n it upon a rich feature set, including 10 distribution densi- ties, 10 distribution quantiles and the mean value (21 in Here, e is a vector of ones, and C is a parameter balanc- total; examples for calcified and non-calcified A-lines ing the complexity and training error to be tolerated. Q is are illustrated in Fig. 3). a n £ n positive semidefinite matrix, and Q  y y K ij i j 2 2 (x,x ) and Kðx ; x Þ¼ expð k x x k =2s Þ is the i j i j i j commonly used Gaussian RBF kernel. s governs the lin- Training the detection model earity of the classifier: larger s values allow greater non- We trained a radial basis function (RBF) support linearity. The hyperparameters C and s were determined vector classification (SVC) model to classify IVUS A- by an exhaustive grid search optimizing overall classifi- lines. An SVC is a flexible structure used to classify cation performance. high-dimensional data. A basic trained SVC is a linear 2804 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 Fig. 3. Features extracted from image intensities, depicting an example of a calcified A-line (blue) and a non-calcified A-line (red) from the manually annotated set. The mean value and the distribution densities are given in the left panel, and the quantiles, in the right panel. Fig. 4. Flowchart for training the support vector machine. The numbered steps are explained in the text under Training the Detection Model. Table 2. Overview of performance of support vector machine Vendor Experiment Measurement Accuracy* Precision Recall F1 score Infraredx Training 0.9170 § 0.0003 0.95 0.91 0.93 Testing 0.87 0.96 0.77 0.86 Volcano Training 0.9113 § 0.0002 0.92 0.93 0.92 Testing 0.89 0.89 0.90 0.89 Boston Scientific Training 0.9084 § 0.0004 0.92 0.91 0.92 Testing 0.89 0.92 0.85 0.88 * For training we report the cross-validation accuracy. Automated IVUS assessment of coronary calcification S. LIU et al. 2805 Figure 4 is the flowchart for model training and test- empirically chosen to be 10 frames and 21 A-lines. ing; the numbered steps are as follows: Followed by an in-frame morphological closing Step 1. The pullbacks in polar coordinates were ran- (1 £ 21) and opening (1 £ 51), the post-processing domly split 9:1; 90% were used for training and 10% steps remove isolated positive and negative classifica- were used for testing. The splitting numbers of pullbacks tions and integrate labels with small gaps. for each system are given in Table 2. Step 2. After generating the training and testing data Validation and reporting set, we extracted the image gray values for each A-line The model was trained for each system separately. in the labeled frames. Performance in the testing set was evaluated in the num- Step 3. The distribution densities were calculated by ber of true/false positive/negative (TP, TN, FP, FN) A- generating a histogram of the image gray values in [0; line classifications, and reported in precision, recall and 255] in 10 bins. They were normalized such that the sum F1 score (Fawcett 2006): is 1. The 10 quantiles were linearly sampled in the range TP 0.5%99.5%. precision ¼ Step 4. Because there were more non-calcified than TP þ FP ð2Þ TP calcified A-lines, we balanced the training categories by recall ¼ randomly downsampling the non-calcified A-lines such TP þ FP that the total number of non-calcified A-lines was equal precision ¢ recall to that of calcified A-lines. F1 score ¼ 2 ¢ Step 5. The input was further normalized using Z- precision þ recall score normalization. The estimated mean ðm ^ Þ and the For further validation, we introduced the ICS, which is standard deviation ðs^Þ were derived from the training set defined as the fraction of detected calcified A-lines in and later applied to Z-score normalize the testing set. the total acquired number. Two ICSs were calculated Step 6. Hyperparameters C and s were selected. using labeled frames; one was estimated using the man- Grid points were evenly chosen on a double log scale, ual labels (denoted as s) and the other was estimated fC ¼ 10 ju 2 ½3; 3; i ¼ 1;⋯;Mgf i i using the detection results (denoted as s): s ¼ 10 jy 2 ½5; 3; j ¼ 1;⋯;Ng; M ¼ N ¼ 21. The j j #ðÞ labeled calcified Alines A-lines were randomized in the ratio 3:7 to training and s ¼ in all labeled frames validation sets for threefold cross-validation. The #ðÞ all Alines parameters with the highest accuracy were chosen for use in the final model. We find that there is a large #ðÞ detected calcified Alines range with nearly optimal performance for ^ s ¼ in all labeled frames 4 1 #ðÞ all Alines s 2 ½10 ; 10 , approximately, with minimal effect of the value of C, indicating limited sensitivity to algo- During the comparison we observed that s and s are rithm or data specifics. linearly related to each other. Therefore, we applied the Step 7. The final trained model was applied to the random sample consensus (RANSAC) regression to high- testing data. Precision, recall and the F1 score, given in light outliers and to fit a linear function ðs ¼ ks þ bÞ with eqn (2), were computed on balanced data where the neg- inliers. Outliers were removed before computing ative examples were downsampled to be equal to the Pearson’s correlation coefficient: amount of positive examples. P P P n x y  x y i i i i r ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; for i 2fg 0; 1;⋯;n xy P P P P 2 2 2 2 n xðÞ x n y ðÞ y i i i i Post-processing After performing the classification, we applied the We further adjusted the detected values by applying the dense fully connected conditional random field (CRF) ^ ^ linear transform f ðsÞ¼ðsbÞ=k. as the post-processing for noise removal Using the estimation result on a whole pullback, we (Kr€ahenbu¨hl andKoltun2011). The method applies a then calculated a total ICS (denoted as s ), which can total Gaussian penalty when two pixels in the defined neigh- give an overall indication of the amount of calcified pla- borhood have different labels. In combination of prior que in the whole pullback. probability, the a posteriori probability was maxi- mized with an optimized labeling solution. Here the #ðÞ detected calcified Alines s ¼ total prior probability was estimated for each pullback using #ðÞ all Alines the SVC-detected label, and the neighborhood was 2806 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 Furthermore, similar to intravascular near-infrared circumferential calcium (s = 1000) occurring in 2mm spectroscopy (Gardner et al. 2008), we present a local Figure 6a and b. ICS (denoted as s ) in short segments, for instance, 2 mm, which can intuitively highlight artery sections DISCUSSION with a heavy calcium burden: #ðÞ detected calcified Alines In the present study, we developed a pipeline for s ¼ in M neighboring frames #ðÞ all Alines automated detection of calcified plaque on IVUS images. Using an SVC classifier and CRF post-processing, we attempted to use simple statistical features of image Again inspired by previous work on IVUS palpogra- intensities for an A-line-based identification of calcium phy (Schaar et al. 2004) and parametric intravascular opti- in the arterial wall. Results indicate that the proposed cal coherence tomography (Gnanadesigan et al. 2016), we framework can be used for a robust estimation of an represent the detection result in a so-called “carpet view”, IVUS-based calcium score, which can be used as an depicting the classification result in a display with dimen- objective evaluation of the presence and amount of calci- sions of circumferential angle and pullback length. fied plaque in the vessel, overall and locally. RESULTS Table 2 summarizes the overall performance of A- ICS overestimates calcium in specific situations line-based calcium detection by the support vector In total, 16 pullbacks were detected as outliers machine (SVM) trained on the data. We observe that an (Infraredx: 9, Volcano: 6, Boston Scientific: 1), 15 of average accuracy >0.9 was achieved, with small varia- which are overestimations comparing with the manual tions across validation experiments and similarly high in scores. All outliers were part of the training set. If the out- testing sets. The manual ICS s and the detected ICS s are liers had arisen in testing data, they could be indicative of compared in scatterplots in Figure 5. Despite a moderate overfitting. Rather, post hoc examination suggested that overestimation, the two numbers are highly correlated the overestimation occurred mainly in three scenarios. (Infraredx: r = 0:94, Volcano: r = 0:88, Boston Scientific: First, some non-calcium image features cannot be r = 0:97). The Wilcoxon tests suggest that the manual distinguished from calcium features, based on A-line and automated measurements are sampled from the intensity statistics only. When the pericardial cavity is vis- same distribution (Infraredx: p = 0.6138, Volcano: ible in the IVUS image, it appears as a thin bright band p = 0.9426, Boston Scientific: p = 0.8370). (visceral pericardium) followed by an abrupt dark cavity. A carpet view representation of the calcium detec- This structure usually appears in large series of consecu- tion results is provided in Figure 6. White areas indicate tive frames and can lead to massive overestimation. This calcification. Vertical lines represent manually labeled was observed in three arteries and, in one case, was frames, where red and blue designate calcium-positive observed in 4264 of 5281 frames in one pullback. Occa- and calcium-negative A-lines, respectively. The color sionally, when large arteries are imaged with an eccentric bar above the carpet view displays the local ICS calcu- catheter, leaving only a bright band of signals on the far lated every 2 mm. The local ICS provides an intuitive wall, the statistical features can be similar to those from overview of the distribution of calcified plaques, with calcified regions. This was observed in five arteries. Fig. 5. Scatterplot of ground-truth intravascular ultrasound calcium score (ICS) calculated using manual labels (s) and the ICS calculated using the detected labels ðsÞ, acquired with the three different intravascular ultrasound systems: (a) Infraredx, (b) Volcano, (c) Boston Scientific. RANSAC = random sample consensus. Automated IVUS assessment of coronary calcification S. LIU et al. 2807 Fig. 6. Detection results in carpet views (circumferential angle £ frame number), where detected calcified regions are in white and non-calcified regions are in black. Positive manual labels are represented by red lines, and the negative coun- terparts, by blue lines. The colored strip above each carpet view represents the local intravascular ultrasound calcium score (2 mm windows), ranging from 0 to 1000 (colorbar on the right). Examples from three different systems: (a) Infrar- edx, (b) Volcano, (c) Boston Scientific. Second, post-processing is designed to remove only in the weights of the SVM. Second, our analysis did small positive regions (which are likely to be false). not require pre-processing (motion correction, gating, False positives that are adjacent to a lesion are difficult conversion to polar coordinates). This formulation of the to rule out, however. This may cause structures such as calcium detection problem, which respects the indepen- the guidewire to be classified as calcium. dence of A-lines and relies on statistical features in the Third, calcified regions appearing as a bright band, data that result directly from the imaging physics, out- with a dorsal shadow, are relatively easy to identify for performs the convolutional neural network-based classi- the experts. However, some calcified plaques exhibit fier described by Balocco et al. (2018) as measured by normal brightness with dark shadows (in the case of a the F1 score. Sofian et al. (2018) analyzed only isolated, directional reflection from a non-normal surface not selected frames, which are not necessarily representative received by the transducer). The correct classification of of clinical data. this appearance of calcified plaque requires the observa- tion of several neighboring frames, which were not avail- Outlook on application: Clinical research able to the experts in this study. Compared with the For analysis of large intravascular imaging data experts’ labels, the detection framework performs more sets, algorithmic quantification of plaque features can consistently for the detection of ’’dark’’ calcified lesions. accelerate quantification studies by eliminating the time- However, in our reporting this is counted as an overesti- consuming manual annotation of thousands of images, mation when compared with the experts’ labels. while simultaneously improving reproducibility and reducing inter-observer variation. Future work, including Comparison with previous work prospective studies, will be needed to evaluate the value The present approach differs in a number of ways of the ICS for stratification of the risk of follow-up from recent work, which employed deep learning frame- events after the index PCI. works (Balocco et al. 2018; Sofian et al. 2018). First, we For asymptomatic populations, the relation between employed data from three different IVUS systems and calcified atherosclerosis and cardiovascular events has developed a universally applicable analysis that differs been quantified in the coronary artery calcium score, an 2808 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 important non-invasive diagnostic metric used to predict we proposed an ICS that comprehensively characterizes cardiac risk with computed tomography (CT) (Shah and the extent of coronary calcification in a vessel examined Coulter 2012). CT detects calcified plaque with high by IVUS. accuracy, with a systematic overestimation of the plaque Conflict of interest disclosure—The authors have no conflicts of inter- volume (Leber et al. 2006; Sun et al. 2008; est to declare. Voros et al. 2011). CT lacks sensitivity to small calcifi- cations compared with IVUS (Van Der Giessen REFERENCES et al. 2011). In our study, no such independent assess- Balocco S, Gonzalez M, Nanculef R, Radeva P, Thomas G. Calcified ment of coronary calcium was available. As a result, the plaque detection in IVUS sequences: Preliminary results using con- ICS has not been validated as a measure of the true volutional nets. In: Hernandez Heredia Y, Milian Nunez V, Ruiz Shulcloper J, (eds). Progress in Artificial Intelligence and Pattern amount of calcification in a vessel. Recognition, IWAIPR 2018. Lecture Notes in Computer Science. 11047, Cham: Springer; 2018. p. 34–42. Di Mario C, Koskinas KC, R€aber L. Clinical benefit of IVUS guidance Outlook on application: Intervention guidance for coronary stenting. J Am Coll Cardiol 2018;72:3138–3141. Operator experience plays an important role in the Dos Santos Filho E, Saijo Y, Tanaka A, Yambe T, Li S, Yoshizawa M. Automated calcification detection and quantification in intravascu- practical use and utility of IVUS in PCI. Automated lar ultrasound images by adaptive thresholding. In: Magjarevic R, image analysis can be helpful for the development of Nagel JH, (eds). World Congress on Medical Physics and Biomedi- objective and fast IVUS-guided PCI strategies. Many cal Engineering 2006. IFMBE Proceedings, Vol. 14. Berlin/Heidel- berg. : Springer Verlag; 2007. p. 1421–1425. studies have found that imaging techniques can effec- Falk E. Pathogenesis of atherosclerosis. J Am Coll Cardiol 2006;47: tively reduce periprocedural (Witzenbichler et al. 2014; C7–C12. Zhang et al. 2018) and late ischemic events (Di Mario Fawcett T. An introduction to ROC analysis. Pattern Recognit Lett 2006;27:861–874. et al. 2018). Robust, real-time automated detection of Gao Z, Guo W, Liu X, Huang W, Zhang H, Tan N, Hau WK, Zhang plaque features can make the technology easier to use YT, Liu H. Automated detection framework of the calcified plaque and thus more accessible. with acoustic shadowing in IVUS images. PLoS One 2014;9 e109997. Observational studies reported that calcified lesions Gardner CM, Tan H, Hull EL, Lisauskas JB, Sum ST, Meese TM, can be associated with post-PCI adverse events, including Jiang C, Madden SP, Caplan JD, Burke AP, Virmani R, Goldstein restenosis and stent thrombosis (Witzenbichler et al. 2014). J, Muller JE. Detection of lipid core coronary plaques in autopsy specimens with a novel catheter-based near-infrared spectroscopy Moderate and severe calcifications were associated with system. JACC Cardiovasc Imaging 2008;1:638–648. major adverse events and revascularization in the target Gnanadesigan M, Kameyama T, Karanasos A, Van Ditzhuijzen NS, artery observed in the 3 years after the implantation of a Van Der Sijde JN, Van Geuns RJ, Ligthart J, Witberg K, Ughi GJ, Van Der Steen AF, Regar E, Van Soest G. Automated characterisa- stent (Shiode et al. 2018). Untreated calcified lesions may tion of lipid core plaques in vivo by quantitative optical coherence also trigger later adverse cardiac events (Mintz 2015). tomography tissue type imaging. EuroIntervention 2016;12: Exposed calcified nodules have been identified as a sub- 1490–1497. Guyon I, Boser BE, Vapnik V. Automatic capacity tuning of very large strate for thrombus formation, leading to acute coronary VC-dimension classifiers. In: Hanson SJ, Cowan JD, Giles CL, syndrome (Virmanietal. 2000; Jia et al. 2013). (eds). Advances in neural information processing systems 5 (NIPS We introduced here a local measure of the calcium Conference). San Francisco. : Morgan Kaufmann; 1993. p. 147–155. burden, by computing the ICS in short (2-mm) segments. Hoffmann R, Mintz GS, Popma JJ, Satler LF, Kent KM, Pichard AD, In the future, the ICS may serve to indicate heavily calci- Leon MB. Treatment of calcified coronary lesions with Palmaz- fied regions that may be likely to benefit from lesion Schatz stents: An intravascular ultrasound study. Eur Heart J 1998;19:1224–1231. preparation during PCI. A threshold value of the ICS for Hong YM. Atherosclerotic cardiovascular disease beginning in child- intervention guidance remains to be determined in fol- hood. Korean Circ J 2010;40:1. low-up research. Jia H, Abtahian F, Aguirre AD, Lee S, Chia S, Lowe H, Kato K, Yonetsu T, Vergallo R, Hu S, Tian J, Lee H, Park SJ, Jang YS, Raf- fel OC, Mizuno K, Uemura S, Itoh T, Kakuta T, Choi SY, Dauer- man HL, Prasad A, Toma C, McNulty I, Zhang S, Yu B, Fuster V, CONCLUSIONS Narula J, Virmani R, Jang IK. In vivo diagnosis of plaque erosion and calcified nodule in patients with acute coronary syndrome by We presented an optimized framework for accurate intravascular optical coherence tomography. J Am Coll Cardiol automated detection and quantification of the presence 2013;62:1748–1758. and extent of calcification in coronary atherosclerosis as Katouzian A, Angelini ED, Carlier SG, Suri JS, Navab N, Laine AF. A state-of-the-art review on segmentation algorithms in intravascular seen by IVUS. Calcified lesions were detected by train- ultrasound (IVUS) images. IEEE Trans Inf Technol Biomed ing a support vector classifier per IVUS A-line on manu- 2012;16:823–834. ally annotated pullback data, followed by post- Kim SW, Mintz GS, Lee WS, Cho JH, Hong SA, Kwon JE, Lee TJ, Park ES, Park KS, Hong JH, Seok JW, Cha YJ, Kim HJ, Kim CJ, processing using regional information. Kim TH. DICOM based intravascular ultrasound signal intensity With manual annotations as a standard, an overall analysis: An echoplaque medical imaging bench study. Coron accuracy of » 0:9 was achieved. Based on this classifier, Artery Dis 2014;25:236–241. Automated IVUS assessment of coronary calcification S. LIU et al. 2809 Kr€ahenbu¨hl P, Koltun V. Efficient inference in fully connected CRFs image for coronary artery disease. 2018 2nd International Confer- with gaussian edge potentials. In: Shawe-Taylor J, Zemel RS, ence on BioSignal Analysis, Processing and Systems (ICBAPS Bartlett PL, Pereira F, Weinberger KQ, (eds). Advances in neural 2018), Kuching, Malaysia. Piscataway, NJ. : IEEE; 2018. p. 47–52. information processing systems 24. Red Hook, NY: Curran Associ- Sun J, Zhang Z, Lu B, Yu W, Yang Y, Zhou Y, Wang Y, Fan Z. Identi- ates; 2011. p. 109–117. fication and quantification of coronary atherosclerotic plaques: A Leber AW, Becker A, Knez A, Von Ziegler F, Sirol M, Nikolaou K, comparison of 64-MDCT and intravascular ultrasound. Am J Ohnesorge B, Fayad ZA, Becker CR, Reiser M, Steinbeck G, Boek- Roentgenol 2008;190:748–754. stegers P. Accuracy of 64-slice computed tomography to classify Taki A, Najafi Z, Roodaki A, Setarehdan SK, Zoroofi RA, Konig A, and quantify plaque volumes in the proximal coronary system: A Navab N. Automatic segmentation of calcified plaques and vessel comparative study using intravascular ultrasound. J Am Coll Car- borders in IVUS images. Int J Comput Assist Radiol Surg diol 2006;47:672–677. 2008;3:347–354. Mintz GS. Intravascular imaging of coronary calcification and its clini- Van Der Giessen AG, Gijsen FJ, Wentzel JJ, Jairam PM, Van Walsum cal implications. JACC Cardiovasc Imaging 2015;8:461–471. T, Neefjes LA, Mollet NR, Niessen WJ, Van De Vosse FN, De Mintz GS, Guagliumi G. Intravascular imaging in coronary artery dis- Feyter PJ, Van Der Steen AF. Small coronary calcifications are not ease. Lancet 2017;390:793–809. detectable by 64-slice contrast enhanced computed tomography. Pu J, Mintz GS, Biro S, Lee JB, Sum ST, Madden SP, Burke AP, Int J Cardiovasc Imaging 2011;27:143–152. Zhang P, He B, Goldstein JA, Stone GW, Muller JE, Virmani R, Virmani R, Kolodgie FD, Burke AP, Farb A, Schwartz SM. Lessons Maehara A. Insights into echo-attenuated plaques, echolucent pla- from sudden coronary death. Arterioscler Thromb Vasc Biol ques, and plaques with spotty calcification: Novel findings from 2000;20:1262–1275. comparisons among intravascular ultrasound, near-infrared spec- Voros S, Rinehart S, Qian Z, Vazquez G, Anderson H, Murrieta L, troscopy, and pathological histology in 2,294 human coronary Wilmer C, Carlson H, Taylor K, Ballard W, Karmpaliotis D, Kaly- artery segments. J Am Coll Cardiol 2014;63:2220–2233. nych A, Brown C. Prospective validation of standardized, 3-dimen- Schaar JA, Regar E, Mastik F, McFadden EP, Saia F, Disco C, de Korte sional, quantitative coronary computed tomographic plaque CL, de Feyter PJ, van der Steen AFW, Serruys PW. Incidence of measurements using radiofrequency backscatter intravascular ultra- high-strain patterns in human coronary arteries: Assessment with sound as reference standard in intermediate coronary arterial three-dimensional intravascular palpography and correlation with lesions. JACC Cardiovasc Interv 2011;4:198–208. clinical presentation. Circulation 2004;109:2716–2719. Wijns W, Shite J, Jones MR, Lee SWL, Price MJ, Fabbiocchi F, Bar- Shah NR, Coulter SA. An evidence-based guide for coronary calcium bato E, Akasaka T, Bezerra H, Holmes D. Optical coherence scoring in asymptomatic patients without coronary heart disease. tomography imaging during percutaneous coronary intervention Tex Hear Inst J 2012;39:240–242. impacts physician decision making: ILUMIEN I study. Eur Heart J Sharma SK, Tomey MI, Teirstein PS, Kini AS, Reitman AB, Lee AC, 2015;36:3346–3355. Genereux P, Chambers JW, Grines CL, Himmelstein SI, Thompson Witzenbichler B, Maehara A, Weisz G, Neumann FJ, Rinaldi MJ, CA, Meredith IT, Bhave A, Moses JW. North American expert Metzger DC, Henry TD, Cox DA, Duffy PL, Brodie BR, Stuckey review of rotational atherectomy. Circ Cardiovasc Interv 2019;12 TD, Mazzaferri EL, Xu K, Parise H, Mehran R, Mintz GS, Stone e007448. GW. Relationship between intravascular ultrasound guidance and Shiode N, Kozuma K, Aoki J, Awata M, Nanasato M, Tanabe K, Yama- clinical outcomes after drug-eluting stents: The assessment of dual guchi J, Kusano H, Nie H, Kimura T. The impact of coronary calcifi- antiplatelet therapy with drug-eluting stents (ADAPT-DES) study. cation on angiographic and 3-year clinical outcomes of everolimus- Circulation 2014;129:463–470. eluting stents: Results of a XIENCE V/PROMUS post-marketing sur- Zhang J, Gao X, Kan J, Ge Z, Han L, Lu S, Tian N, Lin S, Lu Q, Wu X, veillance study. Cardiovasc Interv Ther 2018;33:313–320. Li Q, Liu Z, Chen Y, Qian X, Wang J, Chai D, Chen C, Li X, Smola AJ, Scholkopf B. A tutorial on support vector regression. Stat Gogas BD, Pan T, Shan S, Ye F, Chen SL. Intravascular ultrasound Comput 2004;14:199–222. versus angiography guided drug-eluting stent implantation: The Sofian H, Chia Ming JT, Mohamad S, Noor NM. Calcification detec- ULTIMATE trial. J Am Coll Cardiol 2018;72:3126–3137. tion using deep structured learning in intravascular ultrasound http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Ultrasound in Medicine & Biology Unpaywall

Automated Quantitative Assessment of Coronary Calcification Using Intravascular Ultrasound

Loading next page...
 
/lp/unpaywall/automated-quantitative-assessment-of-coronary-calcification-using-XiHIHKefQM

References (35)

Publisher
Unpaywall
ISSN
0301-5629
DOI
10.1016/j.ultrasmedbio.2020.04.032
Publisher site
See Article on Publisher Site

Abstract

Ultrasound in Med. & Biol., Vol. 46, No. 10, pp. 28012809, 2020 Copyright © 2020 The Author(s). Published by Elsevier Inc. on behalf of World Federation for Ultrasound in Medicine & Biology. This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/) Printed in the USA. All rights reserved. 0301-5629/$ - see front matter https://doi.org/10.1016/j.ultrasmedbio.2020.04.032 Original Contribution AUTOMATED QUANTITATIVE ASSESSMENT OF CORONARY CALCIFICATION USING INTRAVASCULAR ULTRASOUND TAGGEDPSHENGNAN LIU,* TARA NELEMAN,* ELINE M.J. HARTMAN,* JURGEN M.R. LIGTHART,* ,y,z KAREN T. WITBERG,* ANTONIUS F.W. VAN DER STEEN,* JOLANDA J. WENTZEL,* JOOST DAEMEN,* and GIJS VAN SOEST*TAGGEDEND * Department of Cardiology, Erasmus University Medical Center, Rotterdam, The Netherlands; Department of Imaging Physics, Faculty of Applied Sciences, Delft University of Technology, The Netherlands; and Shenzhen Institutes of Advanced Technologies, Shenzhen, China Abstract—Coronary calcification represents a challenge in the treatment of coronary artery disease by stent place- ment. It negatively affects stent expansion and has been related to future adverse cardiac events. Intravascular ultrasound (IVUS) is known for its high sensitivity in detecting coronary calcification. At present, automated quan- tification of calcium as detected by IVUS is not available. For this reason, we developed and validated an optimized framework for accurate automated detection and quantification of calcified plaque in coronary atherosclerosis as seen by IVUS. Calcified lesions were detected by training a supported vector classifier per IVUS A-line on manually annotated IVUS images, followed by post-processing using regional information. We applied our framework to 35 IVUS pullbacks from each of the three commonly used IVUS systems. Cross-validation accuracy for each system was >0.9, and the testing accuracy was 0.87, 0.89 and 0.89 for the three systems. Using the detection result, we pro- pose an IVUS calcium score, based on the fraction of calcium-positive A-lines in a pullback segment, to quantify the extent of calcified plaque. The high accuracy of the proposed classifier suggests that it may provide a robust and accurate tool to assess the presence and amount of coronary calcification and, thus, may play a role in image- guided coronary interventions. (E-mail: g.vansoest@erasmusmc.nl) © 2020 The Author(s). Published by Elsevier Inc. on behalf of World Federation for Ultrasound in Medicine & Biology. This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/). Key Words: Calcified plaque, Intravascular ultrasound, Automated quantification, Intravascular imaging, Coro- nary artery disease. INTRODUCTION its rigidity, circumferential calcification may prevent full dilation of the stent, which may lead to stent underexpan- Coronary artery disease, the most common heart disease, sion, associated with increased risk of target vessel failure is caused by a long-term accumulation of atherosclerotic (Witzenbichler et al. 2014). Specialized techniques, such plaque in the intima of the arterial vessel wall as cutting balloons, rotational atherectomy or intracoro- (Falk 2006; Hong 2010). As plaques grow, they may nary lithotripsy (Sharma et al. 2019), can be used to pre- impinge on the free coronary lumen, causing a stenosis pare calcified plaques to enable complete stent expansion. that limits blood flow to the myocardial territory served The beneficial effects of these plaque modification techni- by the coronary artery. Severe or acute coronary artery ques depend directly on the extent and severity of calcifi- disease is frequently treated by stent implantation in a cation, so detailed knowledge is needed to guide the procedure called percutaneous coronary intervention choice of lesion preparation method (Wijns et al. 2015). (PCI). Atherosclerotic plaques are frequently heteroge- Large calcified plaques can often be unambiguously neous in their composition, and typically consist of identified in intravascular ultrasound (IVUS) images fibrous, lipid-rich and calcified tissue. because of their high reflection of and low penetration by The presence of calcium, in particular, can hamper ultrasound signals (Pu et al. 2014; Mintz and Guagliumi the feasibility of PCI (Hoffmann et al. 1998). Because of 2017). As illustrated in Figure 1, coronary calcium is characterized as a narrow band with high echo intensities, with a dark shadow behind it. Current IVUS software Address correspondence to: Gijs van Soest, PO Box 2040, 3000 does not allow automated calcium quantification. CA Rotterdam, The Netherlands. E-mail: g.vansoest@erasmusmc.nl 2801 2802 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 Fig. 1. Three images in gray scale from Infraredx (a), Volcano (b) and Boston Scientific (c). The calcified plaque is marked with a red arch. Because of the presence of inherent speckle, various calcified plaque load and compare it with the calcium system- and anatomy-dependent artifacts, limited resolu- score based on manual labels. tion and image contrast, automated structure detection in IVUS remains a challenging task (Katouzian et al. 2012). METHODS Simple approaches, using thresholding (Kim et al. 2014) and adaptive thresholding, have been reported IVUS data (Dos Santos Filho et al. 2007). A more complex approach IVUS pullback data sets, acquired in native coro- to segmentation of the leading edge of calcified plaque nary arteries, were extracted from the clinical database combining Rayleigh mixture models, Markov random of the Department of Cardiology, Erasmus MC. The fields, graph searching and prior knowledge, has been data sets were anonymized and contained no identify- reported (Gao et al. 2014). After detection of the intima ing information; all selected patients consented to the and the mediaadventitia borders, calcified regions were use of their data in retrospective studies. All data in detected using a Bayesian classifier (Taki et al. 2008). this study were collected as part of routine clinical Two recent studies have reported the detection of calcifi- care. Consequently, institutional review board cation per frame using a deep learning network approval and individual patient consent are not (Balocco et al. 2018; Sofian et al. 2018). required under Dutch law. We selected 105 pullbacks, In this work, we present a framework for accurate 35 each from three commonly used systems: Infraredx automated detection and quantification of the extent (40 MHz, TVC NIRS Catheter System, Infraredx Inc., of calcification in coronary atherosclerosis as seen by Burlington, MA, USA), Volcano (20 MHz, Eagle Eye IVUS. Without multistep pre-processing, extraction Platinum Rx, and ST Rx, Digital IVUS Catheters, Vol- of complex features or design of a deep learning cano Corp., Rancho Cordova, CA, USA) and Boston approach to perform the task, we found that accurate Scientific (40 MHz, Atlatis SR Pro and OptiCross Cor- classification can be achieved by applying a kernel- onary Imaging Catheter, Boston Scientific Corp., based support vector classifier on simple statistical Natick, MA, USA). Example images from the three features, originating from the imaging physics, systems are provided in Figure 1. extracted directly from un-processed images. Using An overview is given in Table 1. Pullbacks, stored data from three commonly used IVUS systems, we in DICOM format, were manually annotated by trained trained the classifier for recognition of calcified pla- experts (E.M.J.H. and T.N.) at 1-mm intervals. Centering ques per A-line. Using the detection result, we pro- at the imaging catheter, we divided and labeled each pose an IVUS calcium score (ICS) to evaluate the frame as a calcified or non-calcified pie sector (Fig. 2). Table 1. Data description Vendor Population Pullback Frame rate (fs/s) Pullback speed (mm/s) Training Test Infraredx 34 35 16 0.5 31 4 Volcano 35 35 30 0.5 31 4 Boston Scientific 35 35 30 0.5 31 4 Automated IVUS assessment of coronary calcification S. LIU et al. 2803 Fig. 2. Pullback stacks were labeled in Cartesian coordinates and converted to polar coordinates. ‘P’ marks the pull- backs, and ‘L’ marks the label stack. In the zoomed-in red rectangular box, a scale bar is given indicating the pullbacks were labeled every 1 mm. The labeled stacks in the right panels are shown with transparency. Frames (in Cartesian coordinates) and their correspond- hyperplane, which can separate only linearly separable ing labels were converted into polar coordinates, such clusters. To deal with data that are not linearly separable, that each A-line was binary labeled as 1 (calcified) or 0 a non-linear kernel needs to be introduced. When little is (non-calcified). known about the structure of data, a Gaussian RBF ker- nel is a robust choice, assuming only general smoothness (Smola and Sch€olkopf 2004). Identifying features Each A-line is characterized by a set of statistical Calcified plaque can be confidently classified by features x ,and has a binary label y identifying it as calci- i i human observers. The A-line profile itself is quite vari- fied or not. For M labeled A-line distributions, {(x ,y )| i i able though, within each (calcified or non-calcified) cate- y 2 {0, 1}, i 2 {0, , M}}, a classifier gory. By comparing the A-line amplitude statistics, we "# observed that the typical combination of a thin reflection fxðÞ ¼ sgn a y KxðÞ ; x þ b and dorsal shadow of calcium in IVUS may be used as a i i i i¼1 discriminating characteristic, whereas an A-line scan of soft tissue contains many gray values and most ampli- was trained to optimize the problem (Guyon et al. 1993) tude values in a line sampling calcium are low (lumen T T T and shadow), with a few very high ones (calcium bor- min a Qae a such that y a ¼ 0; ð1Þ der). To build an unbiased classifier, we decided to train 0a C; i ¼ 1;⋯;n it upon a rich feature set, including 10 distribution densi- ties, 10 distribution quantiles and the mean value (21 in Here, e is a vector of ones, and C is a parameter balanc- total; examples for calcified and non-calcified A-lines ing the complexity and training error to be tolerated. Q is are illustrated in Fig. 3). a n £ n positive semidefinite matrix, and Q  y y K ij i j 2 2 (x,x ) and Kðx ; x Þ¼ expð k x x k =2s Þ is the i j i j i j commonly used Gaussian RBF kernel. s governs the lin- Training the detection model earity of the classifier: larger s values allow greater non- We trained a radial basis function (RBF) support linearity. The hyperparameters C and s were determined vector classification (SVC) model to classify IVUS A- by an exhaustive grid search optimizing overall classifi- lines. An SVC is a flexible structure used to classify cation performance. high-dimensional data. A basic trained SVC is a linear 2804 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 Fig. 3. Features extracted from image intensities, depicting an example of a calcified A-line (blue) and a non-calcified A-line (red) from the manually annotated set. The mean value and the distribution densities are given in the left panel, and the quantiles, in the right panel. Fig. 4. Flowchart for training the support vector machine. The numbered steps are explained in the text under Training the Detection Model. Table 2. Overview of performance of support vector machine Vendor Experiment Measurement Accuracy* Precision Recall F1 score Infraredx Training 0.9170 § 0.0003 0.95 0.91 0.93 Testing 0.87 0.96 0.77 0.86 Volcano Training 0.9113 § 0.0002 0.92 0.93 0.92 Testing 0.89 0.89 0.90 0.89 Boston Scientific Training 0.9084 § 0.0004 0.92 0.91 0.92 Testing 0.89 0.92 0.85 0.88 * For training we report the cross-validation accuracy. Automated IVUS assessment of coronary calcification S. LIU et al. 2805 Figure 4 is the flowchart for model training and test- empirically chosen to be 10 frames and 21 A-lines. ing; the numbered steps are as follows: Followed by an in-frame morphological closing Step 1. The pullbacks in polar coordinates were ran- (1 £ 21) and opening (1 £ 51), the post-processing domly split 9:1; 90% were used for training and 10% steps remove isolated positive and negative classifica- were used for testing. The splitting numbers of pullbacks tions and integrate labels with small gaps. for each system are given in Table 2. Step 2. After generating the training and testing data Validation and reporting set, we extracted the image gray values for each A-line The model was trained for each system separately. in the labeled frames. Performance in the testing set was evaluated in the num- Step 3. The distribution densities were calculated by ber of true/false positive/negative (TP, TN, FP, FN) A- generating a histogram of the image gray values in [0; line classifications, and reported in precision, recall and 255] in 10 bins. They were normalized such that the sum F1 score (Fawcett 2006): is 1. The 10 quantiles were linearly sampled in the range TP 0.5%99.5%. precision ¼ Step 4. Because there were more non-calcified than TP þ FP ð2Þ TP calcified A-lines, we balanced the training categories by recall ¼ randomly downsampling the non-calcified A-lines such TP þ FP that the total number of non-calcified A-lines was equal precision ¢ recall to that of calcified A-lines. F1 score ¼ 2 ¢ Step 5. The input was further normalized using Z- precision þ recall score normalization. The estimated mean ðm ^ Þ and the For further validation, we introduced the ICS, which is standard deviation ðs^Þ were derived from the training set defined as the fraction of detected calcified A-lines in and later applied to Z-score normalize the testing set. the total acquired number. Two ICSs were calculated Step 6. Hyperparameters C and s were selected. using labeled frames; one was estimated using the man- Grid points were evenly chosen on a double log scale, ual labels (denoted as s) and the other was estimated fC ¼ 10 ju 2 ½3; 3; i ¼ 1;⋯;Mgf i i using the detection results (denoted as s): s ¼ 10 jy 2 ½5; 3; j ¼ 1;⋯;Ng; M ¼ N ¼ 21. The j j #ðÞ labeled calcified Alines A-lines were randomized in the ratio 3:7 to training and s ¼ in all labeled frames validation sets for threefold cross-validation. The #ðÞ all Alines parameters with the highest accuracy were chosen for use in the final model. We find that there is a large #ðÞ detected calcified Alines range with nearly optimal performance for ^ s ¼ in all labeled frames 4 1 #ðÞ all Alines s 2 ½10 ; 10 , approximately, with minimal effect of the value of C, indicating limited sensitivity to algo- During the comparison we observed that s and s are rithm or data specifics. linearly related to each other. Therefore, we applied the Step 7. The final trained model was applied to the random sample consensus (RANSAC) regression to high- testing data. Precision, recall and the F1 score, given in light outliers and to fit a linear function ðs ¼ ks þ bÞ with eqn (2), were computed on balanced data where the neg- inliers. Outliers were removed before computing ative examples were downsampled to be equal to the Pearson’s correlation coefficient: amount of positive examples. P P P n x y  x y i i i i r ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; for i 2fg 0; 1;⋯;n xy P P P P 2 2 2 2 n xðÞ x n y ðÞ y i i i i Post-processing After performing the classification, we applied the We further adjusted the detected values by applying the dense fully connected conditional random field (CRF) ^ ^ linear transform f ðsÞ¼ðsbÞ=k. as the post-processing for noise removal Using the estimation result on a whole pullback, we (Kr€ahenbu¨hl andKoltun2011). The method applies a then calculated a total ICS (denoted as s ), which can total Gaussian penalty when two pixels in the defined neigh- give an overall indication of the amount of calcified pla- borhood have different labels. In combination of prior que in the whole pullback. probability, the a posteriori probability was maxi- mized with an optimized labeling solution. Here the #ðÞ detected calcified Alines s ¼ total prior probability was estimated for each pullback using #ðÞ all Alines the SVC-detected label, and the neighborhood was 2806 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 Furthermore, similar to intravascular near-infrared circumferential calcium (s = 1000) occurring in 2mm spectroscopy (Gardner et al. 2008), we present a local Figure 6a and b. ICS (denoted as s ) in short segments, for instance, 2 mm, which can intuitively highlight artery sections DISCUSSION with a heavy calcium burden: #ðÞ detected calcified Alines In the present study, we developed a pipeline for s ¼ in M neighboring frames #ðÞ all Alines automated detection of calcified plaque on IVUS images. Using an SVC classifier and CRF post-processing, we attempted to use simple statistical features of image Again inspired by previous work on IVUS palpogra- intensities for an A-line-based identification of calcium phy (Schaar et al. 2004) and parametric intravascular opti- in the arterial wall. Results indicate that the proposed cal coherence tomography (Gnanadesigan et al. 2016), we framework can be used for a robust estimation of an represent the detection result in a so-called “carpet view”, IVUS-based calcium score, which can be used as an depicting the classification result in a display with dimen- objective evaluation of the presence and amount of calci- sions of circumferential angle and pullback length. fied plaque in the vessel, overall and locally. RESULTS Table 2 summarizes the overall performance of A- ICS overestimates calcium in specific situations line-based calcium detection by the support vector In total, 16 pullbacks were detected as outliers machine (SVM) trained on the data. We observe that an (Infraredx: 9, Volcano: 6, Boston Scientific: 1), 15 of average accuracy >0.9 was achieved, with small varia- which are overestimations comparing with the manual tions across validation experiments and similarly high in scores. All outliers were part of the training set. If the out- testing sets. The manual ICS s and the detected ICS s are liers had arisen in testing data, they could be indicative of compared in scatterplots in Figure 5. Despite a moderate overfitting. Rather, post hoc examination suggested that overestimation, the two numbers are highly correlated the overestimation occurred mainly in three scenarios. (Infraredx: r = 0:94, Volcano: r = 0:88, Boston Scientific: First, some non-calcium image features cannot be r = 0:97). The Wilcoxon tests suggest that the manual distinguished from calcium features, based on A-line and automated measurements are sampled from the intensity statistics only. When the pericardial cavity is vis- same distribution (Infraredx: p = 0.6138, Volcano: ible in the IVUS image, it appears as a thin bright band p = 0.9426, Boston Scientific: p = 0.8370). (visceral pericardium) followed by an abrupt dark cavity. A carpet view representation of the calcium detec- This structure usually appears in large series of consecu- tion results is provided in Figure 6. White areas indicate tive frames and can lead to massive overestimation. This calcification. Vertical lines represent manually labeled was observed in three arteries and, in one case, was frames, where red and blue designate calcium-positive observed in 4264 of 5281 frames in one pullback. Occa- and calcium-negative A-lines, respectively. The color sionally, when large arteries are imaged with an eccentric bar above the carpet view displays the local ICS calcu- catheter, leaving only a bright band of signals on the far lated every 2 mm. The local ICS provides an intuitive wall, the statistical features can be similar to those from overview of the distribution of calcified plaques, with calcified regions. This was observed in five arteries. Fig. 5. Scatterplot of ground-truth intravascular ultrasound calcium score (ICS) calculated using manual labels (s) and the ICS calculated using the detected labels ðsÞ, acquired with the three different intravascular ultrasound systems: (a) Infraredx, (b) Volcano, (c) Boston Scientific. RANSAC = random sample consensus. Automated IVUS assessment of coronary calcification S. LIU et al. 2807 Fig. 6. Detection results in carpet views (circumferential angle £ frame number), where detected calcified regions are in white and non-calcified regions are in black. Positive manual labels are represented by red lines, and the negative coun- terparts, by blue lines. The colored strip above each carpet view represents the local intravascular ultrasound calcium score (2 mm windows), ranging from 0 to 1000 (colorbar on the right). Examples from three different systems: (a) Infrar- edx, (b) Volcano, (c) Boston Scientific. Second, post-processing is designed to remove only in the weights of the SVM. Second, our analysis did small positive regions (which are likely to be false). not require pre-processing (motion correction, gating, False positives that are adjacent to a lesion are difficult conversion to polar coordinates). This formulation of the to rule out, however. This may cause structures such as calcium detection problem, which respects the indepen- the guidewire to be classified as calcium. dence of A-lines and relies on statistical features in the Third, calcified regions appearing as a bright band, data that result directly from the imaging physics, out- with a dorsal shadow, are relatively easy to identify for performs the convolutional neural network-based classi- the experts. However, some calcified plaques exhibit fier described by Balocco et al. (2018) as measured by normal brightness with dark shadows (in the case of a the F1 score. Sofian et al. (2018) analyzed only isolated, directional reflection from a non-normal surface not selected frames, which are not necessarily representative received by the transducer). The correct classification of of clinical data. this appearance of calcified plaque requires the observa- tion of several neighboring frames, which were not avail- Outlook on application: Clinical research able to the experts in this study. Compared with the For analysis of large intravascular imaging data experts’ labels, the detection framework performs more sets, algorithmic quantification of plaque features can consistently for the detection of ’’dark’’ calcified lesions. accelerate quantification studies by eliminating the time- However, in our reporting this is counted as an overesti- consuming manual annotation of thousands of images, mation when compared with the experts’ labels. while simultaneously improving reproducibility and reducing inter-observer variation. Future work, including Comparison with previous work prospective studies, will be needed to evaluate the value The present approach differs in a number of ways of the ICS for stratification of the risk of follow-up from recent work, which employed deep learning frame- events after the index PCI. works (Balocco et al. 2018; Sofian et al. 2018). First, we For asymptomatic populations, the relation between employed data from three different IVUS systems and calcified atherosclerosis and cardiovascular events has developed a universally applicable analysis that differs been quantified in the coronary artery calcium score, an 2808 Ultrasound in Medicine & Biology Volume 46, Number 10, 2020 important non-invasive diagnostic metric used to predict we proposed an ICS that comprehensively characterizes cardiac risk with computed tomography (CT) (Shah and the extent of coronary calcification in a vessel examined Coulter 2012). CT detects calcified plaque with high by IVUS. accuracy, with a systematic overestimation of the plaque Conflict of interest disclosure—The authors have no conflicts of inter- volume (Leber et al. 2006; Sun et al. 2008; est to declare. Voros et al. 2011). CT lacks sensitivity to small calcifi- cations compared with IVUS (Van Der Giessen REFERENCES et al. 2011). In our study, no such independent assess- Balocco S, Gonzalez M, Nanculef R, Radeva P, Thomas G. Calcified ment of coronary calcium was available. As a result, the plaque detection in IVUS sequences: Preliminary results using con- ICS has not been validated as a measure of the true volutional nets. In: Hernandez Heredia Y, Milian Nunez V, Ruiz Shulcloper J, (eds). Progress in Artificial Intelligence and Pattern amount of calcification in a vessel. Recognition, IWAIPR 2018. Lecture Notes in Computer Science. 11047, Cham: Springer; 2018. p. 34–42. Di Mario C, Koskinas KC, R€aber L. Clinical benefit of IVUS guidance Outlook on application: Intervention guidance for coronary stenting. J Am Coll Cardiol 2018;72:3138–3141. Operator experience plays an important role in the Dos Santos Filho E, Saijo Y, Tanaka A, Yambe T, Li S, Yoshizawa M. Automated calcification detection and quantification in intravascu- practical use and utility of IVUS in PCI. Automated lar ultrasound images by adaptive thresholding. In: Magjarevic R, image analysis can be helpful for the development of Nagel JH, (eds). World Congress on Medical Physics and Biomedi- objective and fast IVUS-guided PCI strategies. Many cal Engineering 2006. IFMBE Proceedings, Vol. 14. Berlin/Heidel- berg. : Springer Verlag; 2007. p. 1421–1425. studies have found that imaging techniques can effec- Falk E. Pathogenesis of atherosclerosis. J Am Coll Cardiol 2006;47: tively reduce periprocedural (Witzenbichler et al. 2014; C7–C12. Zhang et al. 2018) and late ischemic events (Di Mario Fawcett T. An introduction to ROC analysis. Pattern Recognit Lett 2006;27:861–874. et al. 2018). Robust, real-time automated detection of Gao Z, Guo W, Liu X, Huang W, Zhang H, Tan N, Hau WK, Zhang plaque features can make the technology easier to use YT, Liu H. Automated detection framework of the calcified plaque and thus more accessible. with acoustic shadowing in IVUS images. PLoS One 2014;9 e109997. Observational studies reported that calcified lesions Gardner CM, Tan H, Hull EL, Lisauskas JB, Sum ST, Meese TM, can be associated with post-PCI adverse events, including Jiang C, Madden SP, Caplan JD, Burke AP, Virmani R, Goldstein restenosis and stent thrombosis (Witzenbichler et al. 2014). J, Muller JE. Detection of lipid core coronary plaques in autopsy specimens with a novel catheter-based near-infrared spectroscopy Moderate and severe calcifications were associated with system. JACC Cardiovasc Imaging 2008;1:638–648. major adverse events and revascularization in the target Gnanadesigan M, Kameyama T, Karanasos A, Van Ditzhuijzen NS, artery observed in the 3 years after the implantation of a Van Der Sijde JN, Van Geuns RJ, Ligthart J, Witberg K, Ughi GJ, Van Der Steen AF, Regar E, Van Soest G. Automated characterisa- stent (Shiode et al. 2018). Untreated calcified lesions may tion of lipid core plaques in vivo by quantitative optical coherence also trigger later adverse cardiac events (Mintz 2015). tomography tissue type imaging. EuroIntervention 2016;12: Exposed calcified nodules have been identified as a sub- 1490–1497. Guyon I, Boser BE, Vapnik V. Automatic capacity tuning of very large strate for thrombus formation, leading to acute coronary VC-dimension classifiers. In: Hanson SJ, Cowan JD, Giles CL, syndrome (Virmanietal. 2000; Jia et al. 2013). (eds). Advances in neural information processing systems 5 (NIPS We introduced here a local measure of the calcium Conference). San Francisco. : Morgan Kaufmann; 1993. p. 147–155. burden, by computing the ICS in short (2-mm) segments. Hoffmann R, Mintz GS, Popma JJ, Satler LF, Kent KM, Pichard AD, In the future, the ICS may serve to indicate heavily calci- Leon MB. Treatment of calcified coronary lesions with Palmaz- fied regions that may be likely to benefit from lesion Schatz stents: An intravascular ultrasound study. Eur Heart J 1998;19:1224–1231. preparation during PCI. A threshold value of the ICS for Hong YM. Atherosclerotic cardiovascular disease beginning in child- intervention guidance remains to be determined in fol- hood. Korean Circ J 2010;40:1. low-up research. Jia H, Abtahian F, Aguirre AD, Lee S, Chia S, Lowe H, Kato K, Yonetsu T, Vergallo R, Hu S, Tian J, Lee H, Park SJ, Jang YS, Raf- fel OC, Mizuno K, Uemura S, Itoh T, Kakuta T, Choi SY, Dauer- man HL, Prasad A, Toma C, McNulty I, Zhang S, Yu B, Fuster V, CONCLUSIONS Narula J, Virmani R, Jang IK. In vivo diagnosis of plaque erosion and calcified nodule in patients with acute coronary syndrome by We presented an optimized framework for accurate intravascular optical coherence tomography. J Am Coll Cardiol automated detection and quantification of the presence 2013;62:1748–1758. and extent of calcification in coronary atherosclerosis as Katouzian A, Angelini ED, Carlier SG, Suri JS, Navab N, Laine AF. A state-of-the-art review on segmentation algorithms in intravascular seen by IVUS. Calcified lesions were detected by train- ultrasound (IVUS) images. IEEE Trans Inf Technol Biomed ing a support vector classifier per IVUS A-line on manu- 2012;16:823–834. ally annotated pullback data, followed by post- Kim SW, Mintz GS, Lee WS, Cho JH, Hong SA, Kwon JE, Lee TJ, Park ES, Park KS, Hong JH, Seok JW, Cha YJ, Kim HJ, Kim CJ, processing using regional information. Kim TH. DICOM based intravascular ultrasound signal intensity With manual annotations as a standard, an overall analysis: An echoplaque medical imaging bench study. Coron accuracy of » 0:9 was achieved. Based on this classifier, Artery Dis 2014;25:236–241. Automated IVUS assessment of coronary calcification S. LIU et al. 2809 Kr€ahenbu¨hl P, Koltun V. Efficient inference in fully connected CRFs image for coronary artery disease. 2018 2nd International Confer- with gaussian edge potentials. In: Shawe-Taylor J, Zemel RS, ence on BioSignal Analysis, Processing and Systems (ICBAPS Bartlett PL, Pereira F, Weinberger KQ, (eds). Advances in neural 2018), Kuching, Malaysia. Piscataway, NJ. : IEEE; 2018. p. 47–52. information processing systems 24. Red Hook, NY: Curran Associ- Sun J, Zhang Z, Lu B, Yu W, Yang Y, Zhou Y, Wang Y, Fan Z. Identi- ates; 2011. p. 109–117. fication and quantification of coronary atherosclerotic plaques: A Leber AW, Becker A, Knez A, Von Ziegler F, Sirol M, Nikolaou K, comparison of 64-MDCT and intravascular ultrasound. Am J Ohnesorge B, Fayad ZA, Becker CR, Reiser M, Steinbeck G, Boek- Roentgenol 2008;190:748–754. stegers P. Accuracy of 64-slice computed tomography to classify Taki A, Najafi Z, Roodaki A, Setarehdan SK, Zoroofi RA, Konig A, and quantify plaque volumes in the proximal coronary system: A Navab N. Automatic segmentation of calcified plaques and vessel comparative study using intravascular ultrasound. J Am Coll Car- borders in IVUS images. Int J Comput Assist Radiol Surg diol 2006;47:672–677. 2008;3:347–354. Mintz GS. Intravascular imaging of coronary calcification and its clini- Van Der Giessen AG, Gijsen FJ, Wentzel JJ, Jairam PM, Van Walsum cal implications. JACC Cardiovasc Imaging 2015;8:461–471. T, Neefjes LA, Mollet NR, Niessen WJ, Van De Vosse FN, De Mintz GS, Guagliumi G. Intravascular imaging in coronary artery dis- Feyter PJ, Van Der Steen AF. Small coronary calcifications are not ease. Lancet 2017;390:793–809. detectable by 64-slice contrast enhanced computed tomography. Pu J, Mintz GS, Biro S, Lee JB, Sum ST, Madden SP, Burke AP, Int J Cardiovasc Imaging 2011;27:143–152. Zhang P, He B, Goldstein JA, Stone GW, Muller JE, Virmani R, Virmani R, Kolodgie FD, Burke AP, Farb A, Schwartz SM. Lessons Maehara A. Insights into echo-attenuated plaques, echolucent pla- from sudden coronary death. Arterioscler Thromb Vasc Biol ques, and plaques with spotty calcification: Novel findings from 2000;20:1262–1275. comparisons among intravascular ultrasound, near-infrared spec- Voros S, Rinehart S, Qian Z, Vazquez G, Anderson H, Murrieta L, troscopy, and pathological histology in 2,294 human coronary Wilmer C, Carlson H, Taylor K, Ballard W, Karmpaliotis D, Kaly- artery segments. J Am Coll Cardiol 2014;63:2220–2233. nych A, Brown C. Prospective validation of standardized, 3-dimen- Schaar JA, Regar E, Mastik F, McFadden EP, Saia F, Disco C, de Korte sional, quantitative coronary computed tomographic plaque CL, de Feyter PJ, van der Steen AFW, Serruys PW. Incidence of measurements using radiofrequency backscatter intravascular ultra- high-strain patterns in human coronary arteries: Assessment with sound as reference standard in intermediate coronary arterial three-dimensional intravascular palpography and correlation with lesions. JACC Cardiovasc Interv 2011;4:198–208. clinical presentation. Circulation 2004;109:2716–2719. Wijns W, Shite J, Jones MR, Lee SWL, Price MJ, Fabbiocchi F, Bar- Shah NR, Coulter SA. An evidence-based guide for coronary calcium bato E, Akasaka T, Bezerra H, Holmes D. Optical coherence scoring in asymptomatic patients without coronary heart disease. tomography imaging during percutaneous coronary intervention Tex Hear Inst J 2012;39:240–242. impacts physician decision making: ILUMIEN I study. Eur Heart J Sharma SK, Tomey MI, Teirstein PS, Kini AS, Reitman AB, Lee AC, 2015;36:3346–3355. Genereux P, Chambers JW, Grines CL, Himmelstein SI, Thompson Witzenbichler B, Maehara A, Weisz G, Neumann FJ, Rinaldi MJ, CA, Meredith IT, Bhave A, Moses JW. North American expert Metzger DC, Henry TD, Cox DA, Duffy PL, Brodie BR, Stuckey review of rotational atherectomy. Circ Cardiovasc Interv 2019;12 TD, Mazzaferri EL, Xu K, Parise H, Mehran R, Mintz GS, Stone e007448. GW. Relationship between intravascular ultrasound guidance and Shiode N, Kozuma K, Aoki J, Awata M, Nanasato M, Tanabe K, Yama- clinical outcomes after drug-eluting stents: The assessment of dual guchi J, Kusano H, Nie H, Kimura T. The impact of coronary calcifi- antiplatelet therapy with drug-eluting stents (ADAPT-DES) study. cation on angiographic and 3-year clinical outcomes of everolimus- Circulation 2014;129:463–470. eluting stents: Results of a XIENCE V/PROMUS post-marketing sur- Zhang J, Gao X, Kan J, Ge Z, Han L, Lu S, Tian N, Lin S, Lu Q, Wu X, veillance study. Cardiovasc Interv Ther 2018;33:313–320. Li Q, Liu Z, Chen Y, Qian X, Wang J, Chai D, Chen C, Li X, Smola AJ, Scholkopf B. A tutorial on support vector regression. Stat Gogas BD, Pan T, Shan S, Ye F, Chen SL. Intravascular ultrasound Comput 2004;14:199–222. versus angiography guided drug-eluting stent implantation: The Sofian H, Chia Ming JT, Mohamad S, Noor NM. Calcification detec- ULTIMATE trial. J Am Coll Cardiol 2018;72:3126–3137. tion using deep structured learning in intravascular ultrasound

Journal

Ultrasound in Medicine & BiologyUnpaywall

Published: Oct 1, 2020

There are no references for this article.