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

Learn More →

Synthetic contrast-enhanced computed tomography generation using a deep convolutional neural network for cardiac substructure delineation in breast cancer radiation therapy: a feasibility study

Synthetic contrast-enhanced computed tomography generation using a deep convolutional neural... Background: Adjuvant radiation therapy improves the overall survival and loco‑regional control in patients with breast cancer. However, radiation‑induced heart disease, which occurs after treatment from incidental radiation exposure to the cardiac organ, is an emerging challenge. This study aimed to generate synthetic contrast‑ enhanced computed tomography (SCECT ) from non‑ contrast CT (NCT ) using deep learning (DL) and investigate its role in con‑ touring cardiac substructures. We also aimed to determine its applicability for a retrospective study on the substruc‑ ture volume‑ dose relationship for predicting radiation‑induced heart disease. Methods: We prepared NCT‑ CECT cardiac scan pairs of 59 patients. Of these, 35, 4, and 20 pairs were used for train‑ ing, validation, and testing, respectively. We adopted conditional generative adversarial network as a framework to generate SCECT. SCECT was validated in the following three stages: (1) The similarity between SCECT and CECT was evaluated; (2) Manual contouring was performed on SCECT and CECT with sufficient intervals and based on this, the geometric similarity of cardiac substructures was measured between them; (3) The treatment plan was quantitatively analyzed based on the contours of SCECT and CECT. Results: While the mean values (± standard deviation) of the mean absolute error, peak signal‑to ‑noise ratio, and structural similarity index measure between SCECT and CECT were 20.66 ± 5.29, 21.57 ± 1.85, and 0.77 ± 0.06, those were 23.95 ± 6.98, 20.67 ± 2.34, and 0.76 ± 0.07 between NCT and CECT, respectively. The Dice similarity coefficients and mean surface distance between the contours of SCECT and CECT were 0.81 ± 0.06 and 2.44 ± 0.72, respectively. The dosimetry analysis displayed error rates of 0.13 ± 0.27 Gy and 0.71 ± 1.34% for the mean heart dose and V5Gy, respectively. Conclusion: Our findings displayed the feasibility of SCECT generation from NCT and its potential for cardiac sub ‑ structure delineation in patients who underwent breast radiation therapy. *Correspondence: JINSUNG@yuhs.ac Oncosoft Inc, Seoul, South Korea 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, visithttp:// 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. Chun et al. Radiation Oncology (2022) 17:83 Page 2 of 9 Keywords: Contrast‑ enhanced computed tomography, Deep learning, Radiation therapy, Breast cancer, Radiation‑ induced heart disease Background auto-contouring (AC) pipelines leveraging MRI’s soft tis- Adjuvant radiation therapy (RT) improves the over- sue contrast, coupled with NCT for cardiac substructure all survival and loco-regional control in patients with delineation [9, 10]. However, it has a limitation in that breast cancer [1]. However, radiation-induced heart dis- additional images must be obtained, which eventually ease (RIHD), which occurs years after treatment from increases the workflow burden of breast RT. More recent incidental radiation exposure to the cardiac organ, is studies developed DL models for the AC of the heart and an emerging challenge of utmost importance [2]. Darby its substructures directly on NCT [11–13]. However, it is et  al. demonstrated that the risk of an acute coronary mentioned that the manual contours itself, which should event increases linearly with the mean heart dose (MHD) be the ground truth of the DL model, may be inaccurate without a safe threshold of exposure [3]. These dose– due to the poor visibility of NCT [13]. Moreover, manual response relationships were corroborated by subsequent modification of prediction from the DL model should be studies on modern 3D data [4]. performed on NCT with nothing visible in their studies. A recent study suggested that the dose-volume data To enable contouring and modification of cardiac sub - of cardiac substructure units along with the heart might structure without additional image acquisition, it would provide a more accurate prediction of RIHD than MHD be advantageous to provide a basis for them rather than [5, 6]. In breast cancer, van den Bogaard et  al. identified the contours themselves. In this study, we aimed to gen- the volume of the left ventricle receiving 5  Gy (LV-V5) erate synthetic contrast-enhanced CT (SCECT) from as a more important prognostic dose-volume param- NCT using a deep convolutional neural network (DCNN) eter than MHD for predicting an acute coronary event for cardiac substructure delineation in breast cancer RT. [6]. In this context, the ongoing MEDIRAD-BRACE study (NCT03211442) is recruiting 7000 participants to Methods and materials validate these findings. Moreover, the ongoing Radio - Study design therapy Comparative Effectiveness randomized clinical We intended to leverage DL to generate SCECT from trial (NCT02603341) aims to compare the role of proton NCT and verify its manifold utility. SCECT was vali- beam therapy with photon beam therapy in breast cancer dated in four stages, as shown in Fig.  1. (1) The similar - and intends to study the dose-volume profiles of cardiac ity between SCECT generated from the DL model and substructures. In contrast, the dose to the left anterior CECT was evaluated in terms of image quality. (2) The descending (LAD) coronary artery was also suggested cardiac substructures were manually contoured based as a more reliable surrogate for the risk of major cardiac on SCECT and CECT with sufficient intervals. Then, the events than MHD [7]. geometric similarity between contours in each image was Considering that cardiac structures are appropriately measured. (3) We quantitatively compared and analyzed visualized in an electrocardiography-gated contrast- the treatment plan based on the contours of SCECT enhanced CT (CECT) or magnetic resonance imaging and CECT. This study was approved by the institutional (MRI) scan, a bottleneck for more profound and active review board of our institution. studies on cardiac structure-dose relationships comprises difficulties in contouring the heart substructure on plan - Data preparation ning CT scans for breast RT, which are generally obtained We prepared NCT-CECT cardiac scan pairs of without the intravenous (IV) administration of a contrast 59 patients comprising an average resolution agent. The use of CECT in the breast cancer staging is of ~ 0.8 × 0.8 × 1.0  mm . For all patients, NCT and CECT not indicated for patients with early breast cancer in the were scanned by using one of the following multidetec- absence of signs/symptoms of metastatic disease accord- tor row scanners: Somatom Sensation 16, Somatom Sen- ing to NCCN guidelines version 2.2022 [8]. The use of sation 64, Definition Flash (Siemens Medical Solutions, breast MRI also is optional and is not universally recom- Forchheim, Germany), Discovery CT 750 HD, Revolu- mended by experts in the field [8]. tion (GE Medical Systems, Milwaukee, Wisconsin, USA), To overcome these limitations, some studies have tried or iCT (Philips Medical Systems, the Netherlands). All to distinguish the heart substructure retrospectively in an cardiac CT scans were acquired in a volumetric mode in environment consisting of non-contrast CT (NCT). Mor- full inspiration. A bolus of 50–90  mL (1.5  mL/kg body ris et  al. developed atlas- and deep learning (DL)-based weight) of iopamidol (300  mg I/mL, Radisense, Taejoon Chun  et al. Radiation Oncology (2022) 17:83 Page 3 of 9 G =argminmax E logD x, y x,y G D (1) +E log(1 − D(G(x))) + �y − G(x)� where G tries to minimize this objective against an adver- sarial D that tries to maximize it. Unlike the original cGAN learned from noise z, in this study, it is learned in an environment similar to almost supervised learning with L1 loss function [17]. λ is arbitrary constant vari- able, and it was set to 100 in this study. The architecture diagram and further details of our models are illustrated in Fig. 2 and Additional file  1: Figure S1. The modified 2D fully convolutional DenseNet (FC-DenseNet) was used as a generator. As a discriminator, that of PatchGAN model was borrowed, and it was modified and used according to the environment of this study. The input and ground truth image of the prepared 384 × 384 size were randomly cropped to 352 × 352 size and used for training of G and D, and other data augmenta- Fig. 1 The overall workflow for the validation of synthetic tions were not applied. G and D were learned from scratch. contrast‑ enhanced computed tomography (SCECT ). The validation Out of 59 datasets, 35 were used for training, 4 for valida- was conducted in the following three stages: (1) The similarity tion, and 20 for testing. DL model training was conducted between SCECT and contrast‑ enhanced CT (CECT ) was evaluated; (2) by monitoring training and validation datasets, and it was Manual contouring was performed on SCECT and CECT with interval stopped at 200 epochs where the loss values for the vali- and based on it, the geometric similarity of cardiac substructures in each image group was measured; (3) The treatment plan was dation data were saturated. Additional hyper-parameters quantitatively analyzed based on the contours of SCECT and CECT for training the SCECT generation model are detailed in Table  1. The 20 testing datasets were independently man - aged during the training process, and the model applica- tion to the testing datasets was executed only once after the Pharmaceutical, Seoul, South Korea) was injected intra- model was fully trained. venously at a flow rate of 3 mL/s for CECT images. We used the mean absolute error (MAE), peak signal-to- noise ratio (PSNR), and structural similarity index measure Deep learning‑based SCECT generation (SSIM) [18] for evaluating the image quality of the SCECT. The 59 NCT-CECT cardiac scan pairs underwent mul - The formulas used are as follows: tiple pre-processing steps for use in the SCECT gen- y − x i i i=1 eration model. First, image resolution was resampled (3) MAE = to 0.9 × 0.9 × 1.0  mm . After matching the structure of paired NCT-CECT scans based on the deformable image registration algorithm [14, 15], we cropped the 2 PSNR = 10 log MAX /MSE (3) 10 I 384 × 84 × 150 area near the heart. Subsequently, the contrast window was consistently set to [− 150, 500] Hounsfield units  (HU) for a better analysis of important 2μ μ + c 2σ + c x y 1 xy 2 SSIM = . features to be by the DL model. (4) 2 2 2 2 μ + μ + c σ + σ + c 1 2 x y x y We used conditional generative adversarial network (cGAN) [16, 17] as our SCECT generation framework. Here, x denotes the true value (CECT); y , the prediction cGAN learns a mapping from observed image x to y (SCECT), MAX, the maximum possible pixel value of the G : x → y . The generator G is trained to produce out- image; MSE, the mean squared error of two images; μ , the puts that cannot be distinguished from “real” images by average; σ , the variance or covariance; and c , the variable an adversarially trained discriminator, D , which is trained to stabilize the division with weak denominator. The lower to detect the generator’s “fakes.” The objective of a cGAN the MAE and the higher the PSNR and SSIM, the better the can be expressed as values. Chun et al. Radiation Oncology (2022) 17:83 Page 4 of 9 Manual contouring of cardiac substructures of more than a month. According to a recent study [20], a SCECT generation was aimed at contouring the cardiac 5 mm expansion was applied to the two blood vessels: the structures. Accordingly, manual contouring was performed LAD coronary artery and RCA. on SCECT of 20 testing patients by referring to the Univer- sity of Michigan cardiac atlas [19], and geometric similarity evaluation with the contours of CECT was conducted. The SCECT predicted from the deep learning model is restored to another DICOM file that shares most of the information Table 1 Hyper‑parameters for training deep learning‑based with the existing DICOM file of the NCT. Image informa - synthetic contrast‑ enhanced computed tomography generation tion on areas other than prediction ROI (352 × 352 × 150) model shares that of NCT. Parameter Value MIM Maestro (MIM Software, Inc.) was used for con- No. of parameters G : 5.4 M/D : 1.6 M touring organs-at-risk. There were seven target substruc - Batch size 4 tures for delineation, including the heart, left ventricle, left Loss function Adversarial + L1 loss atrium, right ventricle, right atrium, LAD, and right coro- Optimizers G : Adam/D : SGD nary artery (RCA). A physician with more than 10 years of Starting learning rate G : 0.0002/D : 0.00002 experience in breast RT delineated the cardiac substructure Number of epochs 200 of testing datasets. In order to prevent possible bias, the contouring of each image group was conducted at intervals G, generator; D, discriminator; SGD, stochastic gradient descent Fig. 2 Our architecture diagram for the deep learning‑based synthetic contrast ‑ enhanced computed tomography (SCECT ) generation model. The generator model has five Transition Down ( TD ) and Up ( TU) structures, and image features are analyzed in depth through Dense Block (DB) at each stage. Information of low and high‑level features initially extracted from input image is preserved until the end through skip connection and concatenation. The inpuf of generator is NCT, ground truth is CECT, and predicted output is SCECT. The Discriminator model has four transition down (TD ) structures that are slightly different from TD . The input of the discriminator is a two‑ channel image in which NCT is concatenated with D G CECT or SCECT, respectively. The ground truth is 0 or 1, and the predicted output is a decimal value between [0, 1] Chun  et al. Radiation Oncology (2022) 17:83 Page 5 of 9 The geometric similarity between SCECT and CECT structures (Fig.  2e) that were absent in NCT (Fig.  2d), contour groups was evaluated using the Dice similarity similar to those of real CECT (Fig. 2f ). coefficient (DSC) and mean surface distance (MSD): Additional file  1: Figure S2 depicts the MAE, SSIM, and PSNR of NCT and SCECT compared to CECT, respec- 2|X ∩ Y| tively (Additional file  1). A total of 3000 slice images of DSC = (5) |X| + |Y| 20 testing patients were used for the quantitative evalu- ation. While mean values (± standard deviation) of   n ′ S S MAE, PSNR, and SSIM between SCECT and CECT � � � � � � ′ ′   MSD = d p, S + d p , S were 20.66 ± 5.29, 21.57 ± 1.85, and 0.77 ± 0.06, those (6) n + n p=1 p =1 were 23.95 ± 6.98, 20.67 ± 2.34, and 0.76 ± 0.07 NCT and CECT, respectively. In the two-sample t-test, the results Here, X denotes the true volume; Y, the predicted vol- of all three indicators demonstrated statistically signifi - ume; and | |, the number of elements (voxels). S and S′ cant differences (p ≪ 0.05). indicate the outer surfaces of the volume X and Y, respec- tively, n denotes the number of voxels of surface S, and Geometric evaluation of SCECT d(p,S)′ is the distance between a point p on surface S and Figure  4 depicts the results of manual contouring based the surface S is given by the minimum of the Euclidean on SCECT and CECT. The images of the first and sec - ′ ′ norm: d p, S = min p − p . A higher DSC and lower ond rows are examples of two test datasets. As a result ′ ′ p ∈S of the quantitative analysis of contouring (Table 2), DSCs MSD were associated with better values. of atrial and ventricular structures showed a value higher than 0.80, and those of vascular structures showed values Dosimetry analysis higher than 0.55. Overall, MSD was shorter than 3  mm A dosimetric evaluation was additionally performed except for the RCA structure. for 20 testing datasets to determine the extent to which the difference in the contouring structure affected the final dose. Most patients included in the dataset did Dosimetric evaluation of SCECT not receive RT; therefore, the treatment plan was trans- Additional file  1: Figure S3 shows overlaid images of rep- ferred from those who underwent RT with similar resentative test datasets with CT, contours of cardiac physical geometry for this feasibility study. RayStation substructures, and the dose distribution for dosimet- (RaySearch ) was used for the creation and transfer of ric evaluation (Additional file  1). Additional file  1: Fig- the plan. We selected 20 plans from patients with breast ure S4 depicts an averaged DVH for all 20 test datasets cancer treated with RT. RT sites include left, right, or (Additional file  1), and Table  3 summarizes a dosimet- both breasts. All treatments were planned at 40  Gy/15 ric analysis. The values in the table denote the absolute fractions. Each plan was aligned with CT-based rigid difference between dosimetric results of SCECT and registration. For more intuitive visualization, instead of CECT at important clinical points in the DVHs, aver- transferring the contours of SCECT and CECT to NCT, aged over the test datasets. Differences are displayed up the dose distribution was transferred contrary to SCECT to 1.56 Gy in D and 6.64% in the fractional volume in max and CECT (Additional file  1: Figure S3). Then, we ana - V5Gy. The MHD (D of the heart) and V5Gy of the mean lyzed and compared the doses irradiated in the contours left ventricle (bold text in the table) demonstrated differ - of the cardiac substructures of SCECT and CECT. In the ences of 0.13 ± 0.27 Gy and 0.71 ± 1.34%, respectively. In overall dose-volume histogram (DVH), the D , D , the two-sample t-test for all distinct values in Table 3, the max mean V5Gy, V10Gy, V20Gy, V30Gy, and V40Gy were quanti- results did not show statistically significant differences tatively analyzed. All the analyses and evaluations in this (p ≫ 0.05). study were conducted on MATLAB (The MathWorks, Inc.) based on DICOM file information. Discussion We synthesized CECT based on DL to improve the poor Results visualization ability of NCT. We intended to distinguish Image quality evaluation of SCECT cardiac substructures through SCECT and analyze the Figure  3 depicts the inference results of the trained dose-volume relationship for each substructure to enable SCECT generation model. The image in the first and sec - a more meaningful retrospective cardiac toxicity study. ond rows are examples of different slices of test datasets. In terms of image quality, SCECT demonstrated bet- The SCECT generation model predicted not only the ter similarity to CECT than NCT by quantitative met- image contrast but also the virtual interventional septum rics such as MAE, PSNR, and SSIM. Furthermore, it Chun et al. Radiation Oncology (2022) 17:83 Page 6 of 9 Fig. 3 Representative images of non‑ contrast computed tomography (NCT ) (a, d), synthetic contrast‑ enhanced CT (SCECT ) (b, e), and contrast‑ enhanced CT (CECT ) (c, f) of testing datasets. The upper and lower rows indicate different slices of the test datasets provided clinicians the basis for the manual contouring and modification of the cardiac substructures. A dose analysis based on the contours drawn on SCECT did not reveal a considerable difference from the actual CECT. In particular, the small error rates of 0.13 ± 0.27  Gy and 0.71 ± 1.34% were observed in MHD and V5Gy, respec- tively, confirming the clinical utilization of SCECT. The final judgment on the clinical feasibility of SCECT will be made through the accuracy evaluation of the manually drawn cardiac structures. However, the SCECT generation is the ‘image translation’ task from the per- spective of deep learning. The training of the model pro - ceeds to reduce the difference between the preset SCECT and the label CECT. Metrics (MAE, PSNR, and SSIM) for image quality evaluation are used to evaluate whether the SCECT generation model has been properly learned. We did not set up a control group such as inter-rater variation in the ‘geometric evaluation’ part of this study. Instead, in the previous study by Duane et  al., [21] it was reported that inter-rater contour overlap (mean DSC) was 0.60–0.73 for left ventricular segments and 0.10–0.53 for coronary arterial segments. Interobserver contour separation was 1.5–2.2 mm for left ventricular Fig. 4 Representative results of manual contouring on synthetic segments and 1.3–5.1 mm for coronary artery segments contrast‑ enhanced CT (SCECT ) (a, c) and contrast‑ enhanced CT (CECT ) (b, d) images in terms of directed Hausdorff average distance. This Chun  et al. Radiation Oncology (2022) 17:83 Page 7 of 9 spatial variation resulted in <1 Gy dose variation for most relatively reliable because it is the actual CECT. (3) It also segments but 1.2–21.8 Gy variation for segments close to provides clinicians with manual contouring and modifi - a field edge. In cardiac CT images, even if we don’t con - cation because it creates detailed cardiac substructures sider inter-rater variation in contouring, there are intrin- instead of contours themselves. sic intra-fractional variations of cardiac structures, which SCECT, in turn, allowed each institution to conduct a are attributed by continuous cardiac motion and respi- retrospective dose assessment of cardiac substructures ration motion. Cardiac-gated (ECG-gated) and respir- by applying it to the NCT of abundantly existing breast atory-gated CT images can eliminate the major motion RT data. The implementation of SCECT generation in component, but these techniques may not be routinely actual practice would not only benefit physicians but also used in most centers for breast RT planning or treatment. patients. Patients would be spared from the potential side Recently, Nicolas et  al. studied the heart movements in effects of IV contrast agents. Moreover, physicians would 45 patients using a cardiac-gated CT scan and suggested receive additional pseudo-images similar to CECT, gen- using 5  mm of margin surrounding the coronary artery erated from NCT without additional image acquisition. to account for the movements in the breast RT planning To the best of our knowledge, this is the first study to [20]. We speculated that the additional 5 mm margin for synthesize CECT for patients undergoing breast RT using RT planning might encompass the inter-rater margin. DL, besides analyzing it quantitatively and qualitatively. In the absence of treatment plans for testing datasets, This study has several limitations. First, the number of there can be three methods for performing dosimet- evaluation datasets is relatively small in terms of assess- ric evaluation: (a) Based on the respective contours of ment in contouring and dosimetry. Since there were only SCECT and CECT, two optimized plans are generated 20 datasets on a patient basis, it may be challenging to and analyzed for each patient.; (b) single optimized plan extract meaningful information in evaluations performed. is created for each patient based on NCT (or CECT) However, SCECT generation is ultimately an image trans- and evaluated by changing the contours only.; (c) plan lation task, and 3,000 slice images seem to be sufficient to transfer is performed, and dose differences are evalu - evaluate the performance of the image translation model. ated based on the respective contours. Since this study Second, CECT acts as the ground truth of the SCECT attempted to report the dose difference according to the generation model and can have different features depend - different cardiac substructures in an actual situation in ing on the contrast projection protocol or the elapsed which only NCT (or CECT) exists (where only one plan time following the injection. While the overall HU value exists), (a) is not appropriate. In the case of (b, c), at least of the whole heart was high in some cases, only specific one of SCECT and CECT is not optimized for the struc- structures, such as the right ventricle, appeared excep- ture. Considering additional time investment (additional tionally bright in others. We did not distinguish these contouring and re-optimization), we supposed that (b) properties in this study. Therefore, the SCECT genera - instead of (c) would not show a significant difference in tion model was trained to make an averaged contrast fea- analysis. In the plans of 40  Gy/15 fractions, target (hot) ture. If more data can be acquired in the future study, it is dose and heart are physically far enough to have an MHD necessary to develop models according to characteristics of less than 5  Gy in general, so we think that the plan such as protocol and elapsed time separately. Third, exter - transfer is sufficient for this feasibility study. nal validation using a different dataset and/or population Compared with the previous studies [9–13], the novel is required to ensure the reliable performance and gener- points of SCECT generation from NCT were as follows: alizability of the SCECT generation model. (1) It does not cause any burden on the clinical workflow Cardiac structures of SCECT, such as the interven- because it operates without additional medical imag- tional septum, are predicted artificially. Thus, they may ing scans, such as CECT or MRI. (2) Previous AC stud- differ from those of actual CECT. However, in terms of ies without additional image acquisition have an intrinsic dosimetric analysis for cardiac toxicity studies, there was error because ground truth was created on NCT; in a slight difference in the important clinical points on each contrast, the ground truth of our model seems to be image group (Table  3). Considering CECT is a snapshot Table 2 DSC and MSD statistics between manual contours of SCECT and CECT in 20 patients Structure Heart Lt ventricle Lt atrium Rt ventricle Rt atrium LAD RCA Average DSC 0.95 ± 0.03 0.91 ± 0.04 0.86 ± 0.08 0.85 ± 0.06 0.80 ± 0.07 0.74 ± 0.14 0.55 ± 0.20 0.81 ± 0.06 MSD (mm) 2.01 ± 0.95 1.85 ± 1.01 2.25 ± 1.03 2.41 ± 1.02 2.72 ± 0.83 2.19 ± 1.21 3.68 ± 1.60 2.44 ± 0.72 DSC, dice similarity coefficient; LAD, left anterior descending; MSD, mean surface distance; RCA, right coronary artery Chun et al. Radiation Oncology (2022) 17:83 Page 8 of 9 Table 3 Absolute dosimetric differences between SCECT and CECT of each cardiac substructure averaged over 20 patients Structure D D V5Gy V10Gy V20Gy V30Gy V40Gy max mean (Unit) Gy Gy % % % % % Heart 0.65 ± 1.50 0.13 ± 0.27 0.72 ± 1.34 0.41 ± 0.94 0.23 ± 0.66 0.16 ± 0.50 0.06 ± 0.21 Lt ventricle 0.47 ± 0.93 0.15 ± 0.33 0.71 ± 1.34 0.40 ± 1.16 0.21 ± 0.91 0.19 ± 0.76 0.07 ± 0.31 Lt atrium 0.25 ± 0.47 0.10 ± 0.12 0.84 ± 1.60 0.28 ± 0.85 0.00 ± 0.01 0.00 ± 0.00 0.00 ± 0.00 Rt ventricle 0.93 ± 1.60 0.21 ± 0.21 1.95 ± 2.68 0.75 ± 1.24 0.16 ± 0.47 0.08 ± 0.32 0.01 ± 0.06 Rt atrium 1.56 ± 2.90 0.28 ± 0.38 2.90 ± 4.46 0.76 ± 1.68 0.39 ± 0.97 0.14 ± 0.52 0.02 ± 0.09 LAD 0.38 ± 0.39 0.49 ± 1.10 2.81 ± 5.47 1.24 ± 3.45 1.20 ± 3.67 1.04 ± 4.02 0.03 ± 0.11 RCA 1.23 ± 1.63 0.59 ± 0.80 6.64 ± 8.57 3.61 ± 6.81 1.08 ± 3.57 1.11 ± 4.95 0.18 ± 0.82 V50Gy or higher is not displayed because most values are zero or converge to zero CECT, contrast-enhanced computed tomography; SCECT, synthetic contrast-enhanced computed tomography; LAD, left anterior descending; RCA, right coronary artery captured at a particular time point and averaged within Supplementary Information the variability of patient’s breathing, motion, and heart The online version contains supplementary material available at https:// doi. org/ 10. 1186/ s13014‑ 022‑ 02051‑0. beating, the dose differences can be interpreted to be much smaller. The assessment of contouring should be Additional file 1. Fig.: S1. Building blocks and architecture details of designed to meet the established endpoints. Therefore, generator and discriminator. DB stands for Dense Block, TD stands for Tran‑ the results of the dosimetry analysis were considered sition Down, TU stands for Transition Up and m corresponds to the total appropriate to demonstrate the feasibility of SCECT number of feature maps at the end of a block. Fig. S2. Mean absolute error (MAE), peak signal‑to ‑noise ratio (PSNR), and structural similarity [22]. This was similar to a case in which the synthetic index measure (SSIM) of non‑ contrast computed tomography (NCT ) and CT used to realize MRI-only radiotherapy demonstrated synthetic contrast‑ enhanced CT (SCECT ) compared to contrast‑ enhanced an error < 1% in dose calculation, despite the absence of CT (CECT ), respectively. Each boxplot represents the statistics for 3,000 val‑ ues of 20 testing patients in each image group. Fig. S3. Overlaid images 100% consistency with the actual CT [23]. We intend to of two representative patients with computed tomography, manual con‑ continue cardiac toxicity studies according to the irradi- tours (MC), and transferred dose distribution. While the upper row (a, b) ated dose of the cardiac substructures using the present is an example of applying dose distribution of a patient with right breast cancer, the lower row (c, d) is an example of a patient with left breast results. cancer.; Figure S4. The averaged dose‑ volume histogram (DVH) over 20 patients based on synthetic contrast‑ enhanced computed tomography Conclusions (SCECT ) (dashed lines) and contrast‑ enhanced computed tomography (CECT ) (solid lines). Our findings demonstrated the feasibility of SCECT generation from NCT and the potential for cardiac sub- Acknowledgements structure delineation in target substructures, such as the This work was supported by the National Research Foundation of Korea (NRF) ventricles, atriums, and arteries, utilizing SCECT infor- grant. mation for breast RT. Future retrospective studies are Author contributions likely to pave the way for deducing meaningful informa- Study conception and design: JC, JSC, JSK; provision of study materials: YJS, tion from numerous NCTs of patients undergoing RT, CSH, HK, GY, JYM, SYC, JSC, JSK; collection and assembly of data: JC, IKP, JSC; which could not be utilized in the past. Moreover, this data analysis and interpretation: JC, JSC, MSC, JSK; manuscript writing: JC, CO, JSC; All authors read and approved the final manuscript. technology can not only be applied to the heart but also to various regions, such as the abdomen, for studies other Funding than radiation toxicity. The work was funded by the Korea government (MSIT ) (No. 2020R1A4A101661911). Availability of data and materials Abbreviations Data availability is limited due to institutional data protection law and confi‑ SCECT: Synthetic contrast‑ enhanced computed tomography; NCT: Non‑ dentiality of patient data. contrast; DL: Deep learning; AC: Auto‑ contouring; RT: Radiation therapy; RIHD: Radiation‑induced heart disease; MHD: Mean heart dose; LAD: Left anterior descending; CECT: Contrast‑ enhanced CT; MRI: Magnetic resonance imaging; Declarations IV: Intravenous; DCNN: Deep convolutional neural network; HU: Houns‑ field units; PSNR: Peak signal‑to ‑noise ratio; SSIM: Structural similarity index Ethical approval and consent to participate measure; RCA : Right coronary artery; DSC: Dice similarity coefficient; DVH: The study was approved by the Institutional Review Board of Yonsei Cancer Dose‑ volume histogram. Center (4‑2020‑0429). Informed consent was waived. Chun  et al. Radiation Oncology (2022) 17:83 Page 9 of 9 Consent for publication 16. Goodfellow I, Pouget‑Abadie J, Mirza M, et al. Generative adversarial nets. All authors have approved the manuscript and agree with submission to In: Advances in neural information processing systems. 2014. p. 27. Radiation Oncology. 17. Isola P, Zhu J‑ Y, Zhou T, Efros AA. Image‑to ‑image translation with condi‑ tional adversarial networks. 2017:1125–34. Competing interests 18. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: The authors declare that they have no competing interests. from error visibility to structural similarity. IEEE Trans Image Process. 2004;13(4):600–12. Author details 19. Feng M, Moran JM, Koelling T, et al. Development and validation of a Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University heart atlas to study cardiac exposure to radiation following treatment for College of Medicine, Seoul, South Korea. Medical Physics and Biomedical breast cancer. Int J Radiat Oncol Biol Phys. 2011;79(1):10–8. Engineering Lab (MPBEL), Yonsei University College of Medicine, Seoul, South 20. Nicolas E, Khalifa N, Laporte C, Bouhroum S, Kirova Y. Safety margins for 3 4 Korea. Oncosoft Inc, Seoul, South Korea. Department of Radiology, Yonsei the delineation of the left anterior descending artery in patients treated University College of Medicine, Seoul, South Korea. for breast cancer. Int J Radiat Oncol Biol Phys. 2021;109(1):267–72. 21. Duane F, Aznar MC, Bartlett F, et al. A cardiac contouring atlas for radio‑ Received: 27 October 2021 Accepted: 11 April 2022 therapy. Radiother Oncol. 2017;122(3):416–22. 22. Sherer MV, Lin D, Elguindi S, et al. Metrics to evaluate the performance of auto‑segmentation for radiation treatment planning: a critical review. Radiother Oncol. 2021;160:191. 23. Olberg S, Zhang H, Kennedy WR, et al. Synthetic CT reconstruction using a deep spatial pyramid convolutional framework for MR‑ only breast References radiotherapy. Med Phys. 2019;46(9):4135–47. 1. Group EBCTC. Eec ff ts of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15‑ year survival: an overview of the randomised trials. The Lancet. Publisher’s Note 2005;366(9503):2087–2106. Springer Nature remains neutral with regard to jurisdictional claims in pub‑ 2. Taylor C, Kirby A. Cardiac side‑ effects from breast cancer radiotherapy. lished maps and institutional affiliations. Clin Oncol. 2015;27(11):621–9. 3. Darby SC, Ewertz M, McGale P, et al. Risk of ischemic heart dis‑ ease in women after radiotherapy for breast cancer. N Engl J Med. 2013;368(11):987–98. 4. Chung SY, Oh J, Chang JS, et al. Risk of cardiac disease in breast cancer patients: impact of patient‑specific factors and individual heart dose from three‑ dimensional radiotherapy planning. Int J Radiat Oncol Biol Phys. 5. Nieder C, Schill S, Kneschaurek P, Molls M. Influence of different treatment techniques on radiation dose to the LAD coronary artery. Radiat Oncol. 2007;2(1):1–7. 6. van den Bogaard VA, Ta BD, van der Schaaf A, et al. Validation and modification of a prediction model for acute cardiac events in patients with breast cancer treated with radiotherapy based on three‑ dimensional dose distributions to cardiac substructures. J Clin Oncol. 2017;35(11):1171. 7. Atkins KM, Bitterman DS, Chaunzwa TL, et al. Mean heart dose is an inad‑ equate surrogate for left anterior descending coronary artery dose and the risk of major adverse cardiac events in lung cancer radiation therapy. Int J Radiat Oncol Biol Phys. 2021;110:1473. 8. Network NCC. Breast Cancer Version 2.2022. Accessed 8 Feb 2022. https:// www. nccn. org/ profe ssion als/ physi cian_ gls/ pdf/ breast. pdf. 9. Morris ED, Ghanem AI, Dong M, Pantelic MV, Walker EM, Glide‑Hurst CK. Cardiac substructure segmentation with deep learning for improved cardiac sparing. Med Phys. 2020;47(2):576–86. 10. Morris ED, Ghanem AI, Pantelic MV, Walker EM, Han X, Glide‑Hurst CK. Cardiac substructure segmentation and dosimetry using a novel hybrid magnetic resonance and computed tomography cardiac atlas. Int J Radiat Oncol Biol Phys. 2019;103(4):985–93. 11. Harms J, Lei Y, Tian S, et al. Automatic delineation of cardiac substruc‑ tures using a region‑based fully convolutional network. Med Phys. 2021;48:2867. Re Read ady y to to submit y submit your our re researc search h ? Choose BMC and benefit fr ? Choose BMC and benefit from om: : 12. Jin X, Thomas MA, Dise J, et al. Robustness of deep learning segmenta‑ tion of cardiac substructures in non‑ contrast computed tomography for fast, convenient online submission breast cancer radiotherapy. Med Phys. 2021;48:7172. thorough peer review by experienced researchers in your field 13. Jung JW, Mille MM, Ky B, et al. Application of an automatic segmenta‑ tion method for evaluating cardiac structure doses received by breast rapid publication on acceptance radiotherapy patients. Phys Imaging Radiat Oncol. 2021;19:138–44. support for research data, including large and complex data types 14. Pennec X, Cachier P, Ayache N. Understanding the “demon’s algorithm”: • gold Open Access which fosters wider collaboration and increased citations 3D non‑rigid registration by gradient descent. Berlin: Springer; 1999. p. maximum visibility for your research: over 100M website views per year 597–605. • 15. Sharp GC, Li R, Wolfgang J, et al. Plastimatch: an open source software suite for radiotherapy image processing. 2010. At BMC, research is always in progress. Learn more biomedcentral.com/submissions http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Radiation Oncology Springer Journals

Synthetic contrast-enhanced computed tomography generation using a deep convolutional neural network for cardiac substructure delineation in breast cancer radiation therapy: a feasibility study

Loading next page...
 
/lp/springer-journals/synthetic-contrast-enhanced-computed-tomography-generation-using-a-t0BTKgd3cq
Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2022
eISSN
1748-717X
DOI
10.1186/s13014-022-02051-0
Publisher site
See Article on Publisher Site

Abstract

Background: Adjuvant radiation therapy improves the overall survival and loco‑regional control in patients with breast cancer. However, radiation‑induced heart disease, which occurs after treatment from incidental radiation exposure to the cardiac organ, is an emerging challenge. This study aimed to generate synthetic contrast‑ enhanced computed tomography (SCECT ) from non‑ contrast CT (NCT ) using deep learning (DL) and investigate its role in con‑ touring cardiac substructures. We also aimed to determine its applicability for a retrospective study on the substruc‑ ture volume‑ dose relationship for predicting radiation‑induced heart disease. Methods: We prepared NCT‑ CECT cardiac scan pairs of 59 patients. Of these, 35, 4, and 20 pairs were used for train‑ ing, validation, and testing, respectively. We adopted conditional generative adversarial network as a framework to generate SCECT. SCECT was validated in the following three stages: (1) The similarity between SCECT and CECT was evaluated; (2) Manual contouring was performed on SCECT and CECT with sufficient intervals and based on this, the geometric similarity of cardiac substructures was measured between them; (3) The treatment plan was quantitatively analyzed based on the contours of SCECT and CECT. Results: While the mean values (± standard deviation) of the mean absolute error, peak signal‑to ‑noise ratio, and structural similarity index measure between SCECT and CECT were 20.66 ± 5.29, 21.57 ± 1.85, and 0.77 ± 0.06, those were 23.95 ± 6.98, 20.67 ± 2.34, and 0.76 ± 0.07 between NCT and CECT, respectively. The Dice similarity coefficients and mean surface distance between the contours of SCECT and CECT were 0.81 ± 0.06 and 2.44 ± 0.72, respectively. The dosimetry analysis displayed error rates of 0.13 ± 0.27 Gy and 0.71 ± 1.34% for the mean heart dose and V5Gy, respectively. Conclusion: Our findings displayed the feasibility of SCECT generation from NCT and its potential for cardiac sub ‑ structure delineation in patients who underwent breast radiation therapy. *Correspondence: JINSUNG@yuhs.ac Oncosoft Inc, Seoul, South Korea 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, visithttp:// 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. Chun et al. Radiation Oncology (2022) 17:83 Page 2 of 9 Keywords: Contrast‑ enhanced computed tomography, Deep learning, Radiation therapy, Breast cancer, Radiation‑ induced heart disease Background auto-contouring (AC) pipelines leveraging MRI’s soft tis- Adjuvant radiation therapy (RT) improves the over- sue contrast, coupled with NCT for cardiac substructure all survival and loco-regional control in patients with delineation [9, 10]. However, it has a limitation in that breast cancer [1]. However, radiation-induced heart dis- additional images must be obtained, which eventually ease (RIHD), which occurs years after treatment from increases the workflow burden of breast RT. More recent incidental radiation exposure to the cardiac organ, is studies developed DL models for the AC of the heart and an emerging challenge of utmost importance [2]. Darby its substructures directly on NCT [11–13]. However, it is et  al. demonstrated that the risk of an acute coronary mentioned that the manual contours itself, which should event increases linearly with the mean heart dose (MHD) be the ground truth of the DL model, may be inaccurate without a safe threshold of exposure [3]. These dose– due to the poor visibility of NCT [13]. Moreover, manual response relationships were corroborated by subsequent modification of prediction from the DL model should be studies on modern 3D data [4]. performed on NCT with nothing visible in their studies. A recent study suggested that the dose-volume data To enable contouring and modification of cardiac sub - of cardiac substructure units along with the heart might structure without additional image acquisition, it would provide a more accurate prediction of RIHD than MHD be advantageous to provide a basis for them rather than [5, 6]. In breast cancer, van den Bogaard et  al. identified the contours themselves. In this study, we aimed to gen- the volume of the left ventricle receiving 5  Gy (LV-V5) erate synthetic contrast-enhanced CT (SCECT) from as a more important prognostic dose-volume param- NCT using a deep convolutional neural network (DCNN) eter than MHD for predicting an acute coronary event for cardiac substructure delineation in breast cancer RT. [6]. In this context, the ongoing MEDIRAD-BRACE study (NCT03211442) is recruiting 7000 participants to Methods and materials validate these findings. Moreover, the ongoing Radio - Study design therapy Comparative Effectiveness randomized clinical We intended to leverage DL to generate SCECT from trial (NCT02603341) aims to compare the role of proton NCT and verify its manifold utility. SCECT was vali- beam therapy with photon beam therapy in breast cancer dated in four stages, as shown in Fig.  1. (1) The similar - and intends to study the dose-volume profiles of cardiac ity between SCECT generated from the DL model and substructures. In contrast, the dose to the left anterior CECT was evaluated in terms of image quality. (2) The descending (LAD) coronary artery was also suggested cardiac substructures were manually contoured based as a more reliable surrogate for the risk of major cardiac on SCECT and CECT with sufficient intervals. Then, the events than MHD [7]. geometric similarity between contours in each image was Considering that cardiac structures are appropriately measured. (3) We quantitatively compared and analyzed visualized in an electrocardiography-gated contrast- the treatment plan based on the contours of SCECT enhanced CT (CECT) or magnetic resonance imaging and CECT. This study was approved by the institutional (MRI) scan, a bottleneck for more profound and active review board of our institution. studies on cardiac structure-dose relationships comprises difficulties in contouring the heart substructure on plan - Data preparation ning CT scans for breast RT, which are generally obtained We prepared NCT-CECT cardiac scan pairs of without the intravenous (IV) administration of a contrast 59 patients comprising an average resolution agent. The use of CECT in the breast cancer staging is of ~ 0.8 × 0.8 × 1.0  mm . For all patients, NCT and CECT not indicated for patients with early breast cancer in the were scanned by using one of the following multidetec- absence of signs/symptoms of metastatic disease accord- tor row scanners: Somatom Sensation 16, Somatom Sen- ing to NCCN guidelines version 2.2022 [8]. The use of sation 64, Definition Flash (Siemens Medical Solutions, breast MRI also is optional and is not universally recom- Forchheim, Germany), Discovery CT 750 HD, Revolu- mended by experts in the field [8]. tion (GE Medical Systems, Milwaukee, Wisconsin, USA), To overcome these limitations, some studies have tried or iCT (Philips Medical Systems, the Netherlands). All to distinguish the heart substructure retrospectively in an cardiac CT scans were acquired in a volumetric mode in environment consisting of non-contrast CT (NCT). Mor- full inspiration. A bolus of 50–90  mL (1.5  mL/kg body ris et  al. developed atlas- and deep learning (DL)-based weight) of iopamidol (300  mg I/mL, Radisense, Taejoon Chun  et al. Radiation Oncology (2022) 17:83 Page 3 of 9 G =argminmax E logD x, y x,y G D (1) +E log(1 − D(G(x))) + �y − G(x)� where G tries to minimize this objective against an adver- sarial D that tries to maximize it. Unlike the original cGAN learned from noise z, in this study, it is learned in an environment similar to almost supervised learning with L1 loss function [17]. λ is arbitrary constant vari- able, and it was set to 100 in this study. The architecture diagram and further details of our models are illustrated in Fig. 2 and Additional file  1: Figure S1. The modified 2D fully convolutional DenseNet (FC-DenseNet) was used as a generator. As a discriminator, that of PatchGAN model was borrowed, and it was modified and used according to the environment of this study. The input and ground truth image of the prepared 384 × 384 size were randomly cropped to 352 × 352 size and used for training of G and D, and other data augmenta- Fig. 1 The overall workflow for the validation of synthetic tions were not applied. G and D were learned from scratch. contrast‑ enhanced computed tomography (SCECT ). The validation Out of 59 datasets, 35 were used for training, 4 for valida- was conducted in the following three stages: (1) The similarity tion, and 20 for testing. DL model training was conducted between SCECT and contrast‑ enhanced CT (CECT ) was evaluated; (2) by monitoring training and validation datasets, and it was Manual contouring was performed on SCECT and CECT with interval stopped at 200 epochs where the loss values for the vali- and based on it, the geometric similarity of cardiac substructures in each image group was measured; (3) The treatment plan was dation data were saturated. Additional hyper-parameters quantitatively analyzed based on the contours of SCECT and CECT for training the SCECT generation model are detailed in Table  1. The 20 testing datasets were independently man - aged during the training process, and the model applica- tion to the testing datasets was executed only once after the Pharmaceutical, Seoul, South Korea) was injected intra- model was fully trained. venously at a flow rate of 3 mL/s for CECT images. We used the mean absolute error (MAE), peak signal-to- noise ratio (PSNR), and structural similarity index measure Deep learning‑based SCECT generation (SSIM) [18] for evaluating the image quality of the SCECT. The 59 NCT-CECT cardiac scan pairs underwent mul - The formulas used are as follows: tiple pre-processing steps for use in the SCECT gen- y − x i i i=1 eration model. First, image resolution was resampled (3) MAE = to 0.9 × 0.9 × 1.0  mm . After matching the structure of paired NCT-CECT scans based on the deformable image registration algorithm [14, 15], we cropped the 2 PSNR = 10 log MAX /MSE (3) 10 I 384 × 84 × 150 area near the heart. Subsequently, the contrast window was consistently set to [− 150, 500] Hounsfield units  (HU) for a better analysis of important 2μ μ + c 2σ + c x y 1 xy 2 SSIM = . features to be by the DL model. (4) 2 2 2 2 μ + μ + c σ + σ + c 1 2 x y x y We used conditional generative adversarial network (cGAN) [16, 17] as our SCECT generation framework. Here, x denotes the true value (CECT); y , the prediction cGAN learns a mapping from observed image x to y (SCECT), MAX, the maximum possible pixel value of the G : x → y . The generator G is trained to produce out- image; MSE, the mean squared error of two images; μ , the puts that cannot be distinguished from “real” images by average; σ , the variance or covariance; and c , the variable an adversarially trained discriminator, D , which is trained to stabilize the division with weak denominator. The lower to detect the generator’s “fakes.” The objective of a cGAN the MAE and the higher the PSNR and SSIM, the better the can be expressed as values. Chun et al. Radiation Oncology (2022) 17:83 Page 4 of 9 Manual contouring of cardiac substructures of more than a month. According to a recent study [20], a SCECT generation was aimed at contouring the cardiac 5 mm expansion was applied to the two blood vessels: the structures. Accordingly, manual contouring was performed LAD coronary artery and RCA. on SCECT of 20 testing patients by referring to the Univer- sity of Michigan cardiac atlas [19], and geometric similarity evaluation with the contours of CECT was conducted. The SCECT predicted from the deep learning model is restored to another DICOM file that shares most of the information Table 1 Hyper‑parameters for training deep learning‑based with the existing DICOM file of the NCT. Image informa - synthetic contrast‑ enhanced computed tomography generation tion on areas other than prediction ROI (352 × 352 × 150) model shares that of NCT. Parameter Value MIM Maestro (MIM Software, Inc.) was used for con- No. of parameters G : 5.4 M/D : 1.6 M touring organs-at-risk. There were seven target substruc - Batch size 4 tures for delineation, including the heart, left ventricle, left Loss function Adversarial + L1 loss atrium, right ventricle, right atrium, LAD, and right coro- Optimizers G : Adam/D : SGD nary artery (RCA). A physician with more than 10 years of Starting learning rate G : 0.0002/D : 0.00002 experience in breast RT delineated the cardiac substructure Number of epochs 200 of testing datasets. In order to prevent possible bias, the contouring of each image group was conducted at intervals G, generator; D, discriminator; SGD, stochastic gradient descent Fig. 2 Our architecture diagram for the deep learning‑based synthetic contrast ‑ enhanced computed tomography (SCECT ) generation model. The generator model has five Transition Down ( TD ) and Up ( TU) structures, and image features are analyzed in depth through Dense Block (DB) at each stage. Information of low and high‑level features initially extracted from input image is preserved until the end through skip connection and concatenation. The inpuf of generator is NCT, ground truth is CECT, and predicted output is SCECT. The Discriminator model has four transition down (TD ) structures that are slightly different from TD . The input of the discriminator is a two‑ channel image in which NCT is concatenated with D G CECT or SCECT, respectively. The ground truth is 0 or 1, and the predicted output is a decimal value between [0, 1] Chun  et al. Radiation Oncology (2022) 17:83 Page 5 of 9 The geometric similarity between SCECT and CECT structures (Fig.  2e) that were absent in NCT (Fig.  2d), contour groups was evaluated using the Dice similarity similar to those of real CECT (Fig. 2f ). coefficient (DSC) and mean surface distance (MSD): Additional file  1: Figure S2 depicts the MAE, SSIM, and PSNR of NCT and SCECT compared to CECT, respec- 2|X ∩ Y| tively (Additional file  1). A total of 3000 slice images of DSC = (5) |X| + |Y| 20 testing patients were used for the quantitative evalu- ation. While mean values (± standard deviation) of   n ′ S S MAE, PSNR, and SSIM between SCECT and CECT � � � � � � ′ ′   MSD = d p, S + d p , S were 20.66 ± 5.29, 21.57 ± 1.85, and 0.77 ± 0.06, those (6) n + n p=1 p =1 were 23.95 ± 6.98, 20.67 ± 2.34, and 0.76 ± 0.07 NCT and CECT, respectively. In the two-sample t-test, the results Here, X denotes the true volume; Y, the predicted vol- of all three indicators demonstrated statistically signifi - ume; and | |, the number of elements (voxels). S and S′ cant differences (p ≪ 0.05). indicate the outer surfaces of the volume X and Y, respec- tively, n denotes the number of voxels of surface S, and Geometric evaluation of SCECT d(p,S)′ is the distance between a point p on surface S and Figure  4 depicts the results of manual contouring based the surface S is given by the minimum of the Euclidean on SCECT and CECT. The images of the first and sec - ′ ′ norm: d p, S = min p − p . A higher DSC and lower ond rows are examples of two test datasets. As a result ′ ′ p ∈S of the quantitative analysis of contouring (Table 2), DSCs MSD were associated with better values. of atrial and ventricular structures showed a value higher than 0.80, and those of vascular structures showed values Dosimetry analysis higher than 0.55. Overall, MSD was shorter than 3  mm A dosimetric evaluation was additionally performed except for the RCA structure. for 20 testing datasets to determine the extent to which the difference in the contouring structure affected the final dose. Most patients included in the dataset did Dosimetric evaluation of SCECT not receive RT; therefore, the treatment plan was trans- Additional file  1: Figure S3 shows overlaid images of rep- ferred from those who underwent RT with similar resentative test datasets with CT, contours of cardiac physical geometry for this feasibility study. RayStation substructures, and the dose distribution for dosimet- (RaySearch ) was used for the creation and transfer of ric evaluation (Additional file  1). Additional file  1: Fig- the plan. We selected 20 plans from patients with breast ure S4 depicts an averaged DVH for all 20 test datasets cancer treated with RT. RT sites include left, right, or (Additional file  1), and Table  3 summarizes a dosimet- both breasts. All treatments were planned at 40  Gy/15 ric analysis. The values in the table denote the absolute fractions. Each plan was aligned with CT-based rigid difference between dosimetric results of SCECT and registration. For more intuitive visualization, instead of CECT at important clinical points in the DVHs, aver- transferring the contours of SCECT and CECT to NCT, aged over the test datasets. Differences are displayed up the dose distribution was transferred contrary to SCECT to 1.56 Gy in D and 6.64% in the fractional volume in max and CECT (Additional file  1: Figure S3). Then, we ana - V5Gy. The MHD (D of the heart) and V5Gy of the mean lyzed and compared the doses irradiated in the contours left ventricle (bold text in the table) demonstrated differ - of the cardiac substructures of SCECT and CECT. In the ences of 0.13 ± 0.27 Gy and 0.71 ± 1.34%, respectively. In overall dose-volume histogram (DVH), the D , D , the two-sample t-test for all distinct values in Table 3, the max mean V5Gy, V10Gy, V20Gy, V30Gy, and V40Gy were quanti- results did not show statistically significant differences tatively analyzed. All the analyses and evaluations in this (p ≫ 0.05). study were conducted on MATLAB (The MathWorks, Inc.) based on DICOM file information. Discussion We synthesized CECT based on DL to improve the poor Results visualization ability of NCT. We intended to distinguish Image quality evaluation of SCECT cardiac substructures through SCECT and analyze the Figure  3 depicts the inference results of the trained dose-volume relationship for each substructure to enable SCECT generation model. The image in the first and sec - a more meaningful retrospective cardiac toxicity study. ond rows are examples of different slices of test datasets. In terms of image quality, SCECT demonstrated bet- The SCECT generation model predicted not only the ter similarity to CECT than NCT by quantitative met- image contrast but also the virtual interventional septum rics such as MAE, PSNR, and SSIM. Furthermore, it Chun et al. Radiation Oncology (2022) 17:83 Page 6 of 9 Fig. 3 Representative images of non‑ contrast computed tomography (NCT ) (a, d), synthetic contrast‑ enhanced CT (SCECT ) (b, e), and contrast‑ enhanced CT (CECT ) (c, f) of testing datasets. The upper and lower rows indicate different slices of the test datasets provided clinicians the basis for the manual contouring and modification of the cardiac substructures. A dose analysis based on the contours drawn on SCECT did not reveal a considerable difference from the actual CECT. In particular, the small error rates of 0.13 ± 0.27  Gy and 0.71 ± 1.34% were observed in MHD and V5Gy, respec- tively, confirming the clinical utilization of SCECT. The final judgment on the clinical feasibility of SCECT will be made through the accuracy evaluation of the manually drawn cardiac structures. However, the SCECT generation is the ‘image translation’ task from the per- spective of deep learning. The training of the model pro - ceeds to reduce the difference between the preset SCECT and the label CECT. Metrics (MAE, PSNR, and SSIM) for image quality evaluation are used to evaluate whether the SCECT generation model has been properly learned. We did not set up a control group such as inter-rater variation in the ‘geometric evaluation’ part of this study. Instead, in the previous study by Duane et  al., [21] it was reported that inter-rater contour overlap (mean DSC) was 0.60–0.73 for left ventricular segments and 0.10–0.53 for coronary arterial segments. Interobserver contour separation was 1.5–2.2 mm for left ventricular Fig. 4 Representative results of manual contouring on synthetic segments and 1.3–5.1 mm for coronary artery segments contrast‑ enhanced CT (SCECT ) (a, c) and contrast‑ enhanced CT (CECT ) (b, d) images in terms of directed Hausdorff average distance. This Chun  et al. Radiation Oncology (2022) 17:83 Page 7 of 9 spatial variation resulted in <1 Gy dose variation for most relatively reliable because it is the actual CECT. (3) It also segments but 1.2–21.8 Gy variation for segments close to provides clinicians with manual contouring and modifi - a field edge. In cardiac CT images, even if we don’t con - cation because it creates detailed cardiac substructures sider inter-rater variation in contouring, there are intrin- instead of contours themselves. sic intra-fractional variations of cardiac structures, which SCECT, in turn, allowed each institution to conduct a are attributed by continuous cardiac motion and respi- retrospective dose assessment of cardiac substructures ration motion. Cardiac-gated (ECG-gated) and respir- by applying it to the NCT of abundantly existing breast atory-gated CT images can eliminate the major motion RT data. The implementation of SCECT generation in component, but these techniques may not be routinely actual practice would not only benefit physicians but also used in most centers for breast RT planning or treatment. patients. Patients would be spared from the potential side Recently, Nicolas et  al. studied the heart movements in effects of IV contrast agents. Moreover, physicians would 45 patients using a cardiac-gated CT scan and suggested receive additional pseudo-images similar to CECT, gen- using 5  mm of margin surrounding the coronary artery erated from NCT without additional image acquisition. to account for the movements in the breast RT planning To the best of our knowledge, this is the first study to [20]. We speculated that the additional 5 mm margin for synthesize CECT for patients undergoing breast RT using RT planning might encompass the inter-rater margin. DL, besides analyzing it quantitatively and qualitatively. In the absence of treatment plans for testing datasets, This study has several limitations. First, the number of there can be three methods for performing dosimet- evaluation datasets is relatively small in terms of assess- ric evaluation: (a) Based on the respective contours of ment in contouring and dosimetry. Since there were only SCECT and CECT, two optimized plans are generated 20 datasets on a patient basis, it may be challenging to and analyzed for each patient.; (b) single optimized plan extract meaningful information in evaluations performed. is created for each patient based on NCT (or CECT) However, SCECT generation is ultimately an image trans- and evaluated by changing the contours only.; (c) plan lation task, and 3,000 slice images seem to be sufficient to transfer is performed, and dose differences are evalu - evaluate the performance of the image translation model. ated based on the respective contours. Since this study Second, CECT acts as the ground truth of the SCECT attempted to report the dose difference according to the generation model and can have different features depend - different cardiac substructures in an actual situation in ing on the contrast projection protocol or the elapsed which only NCT (or CECT) exists (where only one plan time following the injection. While the overall HU value exists), (a) is not appropriate. In the case of (b, c), at least of the whole heart was high in some cases, only specific one of SCECT and CECT is not optimized for the struc- structures, such as the right ventricle, appeared excep- ture. Considering additional time investment (additional tionally bright in others. We did not distinguish these contouring and re-optimization), we supposed that (b) properties in this study. Therefore, the SCECT genera - instead of (c) would not show a significant difference in tion model was trained to make an averaged contrast fea- analysis. In the plans of 40  Gy/15 fractions, target (hot) ture. If more data can be acquired in the future study, it is dose and heart are physically far enough to have an MHD necessary to develop models according to characteristics of less than 5  Gy in general, so we think that the plan such as protocol and elapsed time separately. Third, exter - transfer is sufficient for this feasibility study. nal validation using a different dataset and/or population Compared with the previous studies [9–13], the novel is required to ensure the reliable performance and gener- points of SCECT generation from NCT were as follows: alizability of the SCECT generation model. (1) It does not cause any burden on the clinical workflow Cardiac structures of SCECT, such as the interven- because it operates without additional medical imag- tional septum, are predicted artificially. Thus, they may ing scans, such as CECT or MRI. (2) Previous AC stud- differ from those of actual CECT. However, in terms of ies without additional image acquisition have an intrinsic dosimetric analysis for cardiac toxicity studies, there was error because ground truth was created on NCT; in a slight difference in the important clinical points on each contrast, the ground truth of our model seems to be image group (Table  3). Considering CECT is a snapshot Table 2 DSC and MSD statistics between manual contours of SCECT and CECT in 20 patients Structure Heart Lt ventricle Lt atrium Rt ventricle Rt atrium LAD RCA Average DSC 0.95 ± 0.03 0.91 ± 0.04 0.86 ± 0.08 0.85 ± 0.06 0.80 ± 0.07 0.74 ± 0.14 0.55 ± 0.20 0.81 ± 0.06 MSD (mm) 2.01 ± 0.95 1.85 ± 1.01 2.25 ± 1.03 2.41 ± 1.02 2.72 ± 0.83 2.19 ± 1.21 3.68 ± 1.60 2.44 ± 0.72 DSC, dice similarity coefficient; LAD, left anterior descending; MSD, mean surface distance; RCA, right coronary artery Chun et al. Radiation Oncology (2022) 17:83 Page 8 of 9 Table 3 Absolute dosimetric differences between SCECT and CECT of each cardiac substructure averaged over 20 patients Structure D D V5Gy V10Gy V20Gy V30Gy V40Gy max mean (Unit) Gy Gy % % % % % Heart 0.65 ± 1.50 0.13 ± 0.27 0.72 ± 1.34 0.41 ± 0.94 0.23 ± 0.66 0.16 ± 0.50 0.06 ± 0.21 Lt ventricle 0.47 ± 0.93 0.15 ± 0.33 0.71 ± 1.34 0.40 ± 1.16 0.21 ± 0.91 0.19 ± 0.76 0.07 ± 0.31 Lt atrium 0.25 ± 0.47 0.10 ± 0.12 0.84 ± 1.60 0.28 ± 0.85 0.00 ± 0.01 0.00 ± 0.00 0.00 ± 0.00 Rt ventricle 0.93 ± 1.60 0.21 ± 0.21 1.95 ± 2.68 0.75 ± 1.24 0.16 ± 0.47 0.08 ± 0.32 0.01 ± 0.06 Rt atrium 1.56 ± 2.90 0.28 ± 0.38 2.90 ± 4.46 0.76 ± 1.68 0.39 ± 0.97 0.14 ± 0.52 0.02 ± 0.09 LAD 0.38 ± 0.39 0.49 ± 1.10 2.81 ± 5.47 1.24 ± 3.45 1.20 ± 3.67 1.04 ± 4.02 0.03 ± 0.11 RCA 1.23 ± 1.63 0.59 ± 0.80 6.64 ± 8.57 3.61 ± 6.81 1.08 ± 3.57 1.11 ± 4.95 0.18 ± 0.82 V50Gy or higher is not displayed because most values are zero or converge to zero CECT, contrast-enhanced computed tomography; SCECT, synthetic contrast-enhanced computed tomography; LAD, left anterior descending; RCA, right coronary artery captured at a particular time point and averaged within Supplementary Information the variability of patient’s breathing, motion, and heart The online version contains supplementary material available at https:// doi. org/ 10. 1186/ s13014‑ 022‑ 02051‑0. beating, the dose differences can be interpreted to be much smaller. The assessment of contouring should be Additional file 1. Fig.: S1. Building blocks and architecture details of designed to meet the established endpoints. Therefore, generator and discriminator. DB stands for Dense Block, TD stands for Tran‑ the results of the dosimetry analysis were considered sition Down, TU stands for Transition Up and m corresponds to the total appropriate to demonstrate the feasibility of SCECT number of feature maps at the end of a block. Fig. S2. Mean absolute error (MAE), peak signal‑to ‑noise ratio (PSNR), and structural similarity [22]. This was similar to a case in which the synthetic index measure (SSIM) of non‑ contrast computed tomography (NCT ) and CT used to realize MRI-only radiotherapy demonstrated synthetic contrast‑ enhanced CT (SCECT ) compared to contrast‑ enhanced an error < 1% in dose calculation, despite the absence of CT (CECT ), respectively. Each boxplot represents the statistics for 3,000 val‑ ues of 20 testing patients in each image group. Fig. S3. Overlaid images 100% consistency with the actual CT [23]. We intend to of two representative patients with computed tomography, manual con‑ continue cardiac toxicity studies according to the irradi- tours (MC), and transferred dose distribution. While the upper row (a, b) ated dose of the cardiac substructures using the present is an example of applying dose distribution of a patient with right breast cancer, the lower row (c, d) is an example of a patient with left breast results. cancer.; Figure S4. The averaged dose‑ volume histogram (DVH) over 20 patients based on synthetic contrast‑ enhanced computed tomography Conclusions (SCECT ) (dashed lines) and contrast‑ enhanced computed tomography (CECT ) (solid lines). Our findings demonstrated the feasibility of SCECT generation from NCT and the potential for cardiac sub- Acknowledgements structure delineation in target substructures, such as the This work was supported by the National Research Foundation of Korea (NRF) ventricles, atriums, and arteries, utilizing SCECT infor- grant. mation for breast RT. Future retrospective studies are Author contributions likely to pave the way for deducing meaningful informa- Study conception and design: JC, JSC, JSK; provision of study materials: YJS, tion from numerous NCTs of patients undergoing RT, CSH, HK, GY, JYM, SYC, JSC, JSK; collection and assembly of data: JC, IKP, JSC; which could not be utilized in the past. Moreover, this data analysis and interpretation: JC, JSC, MSC, JSK; manuscript writing: JC, CO, JSC; All authors read and approved the final manuscript. technology can not only be applied to the heart but also to various regions, such as the abdomen, for studies other Funding than radiation toxicity. The work was funded by the Korea government (MSIT ) (No. 2020R1A4A101661911). Availability of data and materials Abbreviations Data availability is limited due to institutional data protection law and confi‑ SCECT: Synthetic contrast‑ enhanced computed tomography; NCT: Non‑ dentiality of patient data. contrast; DL: Deep learning; AC: Auto‑ contouring; RT: Radiation therapy; RIHD: Radiation‑induced heart disease; MHD: Mean heart dose; LAD: Left anterior descending; CECT: Contrast‑ enhanced CT; MRI: Magnetic resonance imaging; Declarations IV: Intravenous; DCNN: Deep convolutional neural network; HU: Houns‑ field units; PSNR: Peak signal‑to ‑noise ratio; SSIM: Structural similarity index Ethical approval and consent to participate measure; RCA : Right coronary artery; DSC: Dice similarity coefficient; DVH: The study was approved by the Institutional Review Board of Yonsei Cancer Dose‑ volume histogram. Center (4‑2020‑0429). Informed consent was waived. Chun  et al. Radiation Oncology (2022) 17:83 Page 9 of 9 Consent for publication 16. Goodfellow I, Pouget‑Abadie J, Mirza M, et al. Generative adversarial nets. All authors have approved the manuscript and agree with submission to In: Advances in neural information processing systems. 2014. p. 27. Radiation Oncology. 17. Isola P, Zhu J‑ Y, Zhou T, Efros AA. Image‑to ‑image translation with condi‑ tional adversarial networks. 2017:1125–34. Competing interests 18. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: The authors declare that they have no competing interests. from error visibility to structural similarity. IEEE Trans Image Process. 2004;13(4):600–12. Author details 19. Feng M, Moran JM, Koelling T, et al. Development and validation of a Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University heart atlas to study cardiac exposure to radiation following treatment for College of Medicine, Seoul, South Korea. Medical Physics and Biomedical breast cancer. Int J Radiat Oncol Biol Phys. 2011;79(1):10–8. Engineering Lab (MPBEL), Yonsei University College of Medicine, Seoul, South 20. Nicolas E, Khalifa N, Laporte C, Bouhroum S, Kirova Y. Safety margins for 3 4 Korea. Oncosoft Inc, Seoul, South Korea. Department of Radiology, Yonsei the delineation of the left anterior descending artery in patients treated University College of Medicine, Seoul, South Korea. for breast cancer. Int J Radiat Oncol Biol Phys. 2021;109(1):267–72. 21. Duane F, Aznar MC, Bartlett F, et al. A cardiac contouring atlas for radio‑ Received: 27 October 2021 Accepted: 11 April 2022 therapy. Radiother Oncol. 2017;122(3):416–22. 22. Sherer MV, Lin D, Elguindi S, et al. Metrics to evaluate the performance of auto‑segmentation for radiation treatment planning: a critical review. Radiother Oncol. 2021;160:191. 23. Olberg S, Zhang H, Kennedy WR, et al. Synthetic CT reconstruction using a deep spatial pyramid convolutional framework for MR‑ only breast References radiotherapy. Med Phys. 2019;46(9):4135–47. 1. Group EBCTC. Eec ff ts of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15‑ year survival: an overview of the randomised trials. The Lancet. Publisher’s Note 2005;366(9503):2087–2106. Springer Nature remains neutral with regard to jurisdictional claims in pub‑ 2. Taylor C, Kirby A. Cardiac side‑ effects from breast cancer radiotherapy. lished maps and institutional affiliations. Clin Oncol. 2015;27(11):621–9. 3. Darby SC, Ewertz M, McGale P, et al. Risk of ischemic heart dis‑ ease in women after radiotherapy for breast cancer. N Engl J Med. 2013;368(11):987–98. 4. Chung SY, Oh J, Chang JS, et al. Risk of cardiac disease in breast cancer patients: impact of patient‑specific factors and individual heart dose from three‑ dimensional radiotherapy planning. Int J Radiat Oncol Biol Phys. 5. Nieder C, Schill S, Kneschaurek P, Molls M. Influence of different treatment techniques on radiation dose to the LAD coronary artery. Radiat Oncol. 2007;2(1):1–7. 6. van den Bogaard VA, Ta BD, van der Schaaf A, et al. Validation and modification of a prediction model for acute cardiac events in patients with breast cancer treated with radiotherapy based on three‑ dimensional dose distributions to cardiac substructures. J Clin Oncol. 2017;35(11):1171. 7. Atkins KM, Bitterman DS, Chaunzwa TL, et al. Mean heart dose is an inad‑ equate surrogate for left anterior descending coronary artery dose and the risk of major adverse cardiac events in lung cancer radiation therapy. Int J Radiat Oncol Biol Phys. 2021;110:1473. 8. Network NCC. Breast Cancer Version 2.2022. Accessed 8 Feb 2022. https:// www. nccn. org/ profe ssion als/ physi cian_ gls/ pdf/ breast. pdf. 9. Morris ED, Ghanem AI, Dong M, Pantelic MV, Walker EM, Glide‑Hurst CK. Cardiac substructure segmentation with deep learning for improved cardiac sparing. Med Phys. 2020;47(2):576–86. 10. Morris ED, Ghanem AI, Pantelic MV, Walker EM, Han X, Glide‑Hurst CK. Cardiac substructure segmentation and dosimetry using a novel hybrid magnetic resonance and computed tomography cardiac atlas. Int J Radiat Oncol Biol Phys. 2019;103(4):985–93. 11. Harms J, Lei Y, Tian S, et al. Automatic delineation of cardiac substruc‑ tures using a region‑based fully convolutional network. Med Phys. 2021;48:2867. Re Read ady y to to submit y submit your our re researc search h ? Choose BMC and benefit fr ? Choose BMC and benefit from om: : 12. Jin X, Thomas MA, Dise J, et al. Robustness of deep learning segmenta‑ tion of cardiac substructures in non‑ contrast computed tomography for fast, convenient online submission breast cancer radiotherapy. Med Phys. 2021;48:7172. thorough peer review by experienced researchers in your field 13. Jung JW, Mille MM, Ky B, et al. Application of an automatic segmenta‑ tion method for evaluating cardiac structure doses received by breast rapid publication on acceptance radiotherapy patients. Phys Imaging Radiat Oncol. 2021;19:138–44. support for research data, including large and complex data types 14. Pennec X, Cachier P, Ayache N. Understanding the “demon’s algorithm”: • gold Open Access which fosters wider collaboration and increased citations 3D non‑rigid registration by gradient descent. Berlin: Springer; 1999. p. maximum visibility for your research: over 100M website views per year 597–605. • 15. Sharp GC, Li R, Wolfgang J, et al. Plastimatch: an open source software suite for radiotherapy image processing. 2010. At BMC, research is always in progress. Learn more biomedcentral.com/submissions

Journal

Radiation OncologySpringer Journals

Published: Apr 22, 2022

Keywords: Contrast-enhanced computed tomography; Deep learning; Radiation therapy; Breast cancer; Radiation-induced heart disease

References