ImageCHD: A 3D Computed Tomography Image Dataset for Classification of Congenital Heart Disease
Xiaowei Xu, Tianchen Wang, Jian Zhuang, Haiyun Yuan, Meiping Huang, Jianzheng Cen, Qianjun Jia, Yuhao Dong, Yiyu Shi
IImageCHD: A 3D Computed TomographyImage Dataset for Classification of CongenitalHeart Disease
Xiaowei Xu , Tianchen Wang , Jian Zhuang , Haiyun Yuan , Meiping Huang ,Jianzheng Cen , Qianjun Jia , Yuhao Dong , and Yiyu Shi Guangdong Provincial People’s Hospital, [email protected], [email protected], [email protected] University of Notre Dame { twang9, yshi4 } @nd.edu Abstract.
Congenital heart disease (CHD) is the most common typeof birth defects, which occurs 1 in every 110 births in the United States.CHD usually comes with severe variations in heart structure and greatartery connections that can be classified into many types. Thus highlyspecialized domain knowledge and time-consuming human process isneeded to analyze the associated medical images. On the other hand,due to the complexity of CHD and the lack of dataset, little has beenexplored on the automatic diagnosis (classification) of CHDs. In this pa-per, we present ImageCHD, the first medical image dataset for CHDclassification. ImageCHD contains 110 3D Computed Tomography (CT)images covering most types of CHD, which is of decent size comparedwith existing medical imaging datasets. Classification of CHDs requiresthe identification of large structural changes without any local tissuechanges, with limited data. It is an example of a larger class of problemsthat are quite difficult for current machine-learning based vision methodsto solve. To demonstrate this, we further present a baseline frameworkfor automatic classification of CHD, based on a state-of-the-art CHD seg-mentation method. Experimental results show that the baseline frame-work can only achieve a classification accuracy of 82.0% under selectiveprediction scheme with 88.4% coverage, leaving big room for further im-provement. We hope that ImageCHD can stimulate further research andlead to innovative and generic solutions that would have an impact inmultiple domains. Our dataset is released to the public [1].
Keywords:
Dataset · Congenital Heart Disease · Automatic Diagnosis · Computed Tomography.
Congenital heart disease (CHD) is the problem with the heart structure that ispresent at birth, which is the most common type of birth defects [3]. In recentyears, noninvasive imaging techniques such as computed tomography (CT) have a r X i v : . [ ee ss . I V ] J a n X. Xu, et al. prevailed in comprehensive diagnosis, intervention decision-making, and regu-lar follow-up for CHD. However, analysis (e.g., segmentation or classification)of these medical images are usually performed manually by experienced cardio-vascular radiologists, which is time-consuming and requires highly specializeddomain knowledge.
RA LA
RV LV PA AO PA RA LA RV MyoPA
AO RA LARV LVPA AO PA (a) Normal heart anatomy (b) Pulmonary atresia (c) Common arterial trunk LV Myo Myo
Fig. 1.
Examples of large heart structure and great artery connection variations inCHD (LV-left ventricle, RV-right ventricle, LA-left atrium, RA-right atrium, Myo-myocardium, AO-aorta and PA-pulmonary artery). Best viewed in color.
On the other hand, automatic segmentation and classification of medicalimages in CHD is rather challenging. Patients with CHD typically suffer fromsevere variation in heart structures and connections between different parts ofthe anatomy. Two examples are shown in Fig. 1: the disappearance of the maintrunk of pulmonary artery (PA) in (b)(c) introduces much difficulty in the cor-rect segmentation of PA and AO. In addition, CHD does not necessarily causelocal tissue changes, as in lesions. As such, hearts with CHD have similar localstatistics as normal hearts but with global structural changes. Automatic algo-rithms to detect the disorders need to be able to capture such changes, whichrequire excellent usage of the contextual information. CHD classification is fur-ther complicated by the fact that a patient’s CT image may exhibit more thanone type of CHD, and the number of types is more than 20 [3].Various works exist in segmentation and classification of heart with normalanatomy, e.g., [13,8,19,9,27,4,22,25,6,26,24,5,14], most of which are based ondeep neural networks (DNNs) [17,15]. Recently, researchers started to exploreheart segmentation in CHD. The works [23,16,21,20,7] adopt DNNs for bloodpool and myocardium segmentation only. The only automatic whole heart andgreat artery segmentation method in CHD [18] in the literature uses a deeplearning and shape similarity analysis based method. A 3D CT dataset for CHDsegmentation is also released there. In addition to segmentation, there are alsosome works about classification of adult heart diseases [2] but not CHD. Theautomatic classification of CHD still remains a missing piece in the literaturedue to the complexity of CHD and the lack of dataset.In this paper, we present ImageCHD, the first medical image dataset for CHDclassification. ImageCHD contains 110 3D Computed Tomography (CT) imageswhich covers 16 types of CHD. CT images are labelled by a team of four expe-rienced cardiovascular radiologists with 7-substructure segmentation and CHD itle Suppressed Due to Excessive Length 3 type classification. The dataset is of decent size compared with other medicalimaging datasets [23][19]. We also present a baseline method for automatic CHDclassification based on the state-of-the-art CHD segmentation framework [18],which is the first automatic CHD classification method in the literature. Resultsshow that the baseline framework can achieve a classification accuracy of 82.0%under selective prediction scheme with 88.4% coverage, and there is still bigroom for further improvement. (a) AAH (b) DAA (c) IAA (d) PuA (e) CA (f) ToF (g) TGA (h) DORV LA PAAO RA RAPA LAAO RA RARA RA RA RAPA PAPA PA PA PAAO AOAO AO AO AO
Myo
Myo
Myo
Myo Myo Myo
Fig. 2.
Examples of CT images in the ImageCHD dataset with its types of CHD.
The ImageCHD dataset consists of 3D CT images captured by a Siemens bio-graph 64 machine from 110 patients, with age between 1 month and 40 years(mostly between 1 month and 2 years). The size of the images is 512 × × (129 − × × mm . The dataset covers 16 typesof CHD, which include eight common types (atrial septal defect (ASD), atrio-ventricular septal defect (AVSD), patent ductus arteriosus (PDA), pulmonaryatresia (PuA), ventricular septal defect (VSD), co-arctation (CA), tetrology offallot (TOF), and transposition of great arteries (TGA)) plus eight less com-mon ones (pulmonary artery sling (PAS), double outlet right ventricle (DORV),common arterial trunk (CAT), double aortic arch (DAA), anomalous pulmonaryvenous drainage (APVC), aortic arch hypoplasia (AAH), interrupted aortic arch(IAA), double superior vena cava (DSVC)). The number of images associatedwith each is summarized in Table 1. Several examples of images in the datasetare shown in Figure 2. Due to the structure complexities, the labeling includingsegmentation and classification is performed by a team of four cardiovascularradiologists who have extensive experience with CHD. The segmentation labelof each image is fulfilled by only one radiologist, and its diagnosis is performedby four. The time to label each image is around 1-1.5 hours on average. Thesegmentation include seven substructures: LV, RV, LA, RA, Myo, AO and PA. X. Xu, et al.
Table 1.
The types of CHD in the ImageCHD dataset (containing 110 3D CT images)and the associated number of images. Note that some images may correspond to morethan one type of CHD. Common CHDASD AVSD VSD TOF PDA TGA CA PuA26 18 44 12 14 7 6 16Less Common CHD NormalPAS DORV CAT DAA APVC AAH IAA DSVC3 8 4 5 6 3 3 8 6
RoI cropping (3D U-net, 64×64×64)Chambers and initial parts of great arteries segmentation (3D U-net, 64×64×64)
Blood pool segmentation (2D U-net, 512×512)
Input
3D medical images
Shape similarity calculation Connection analysis
Similarity based shape analysis
Vessel extraction
Skeleton extraction
Segmentation based connection analysis
Shape analysis
Output classification result
Final determination
Fig. 3.
Overview of the baseline method for CHD classification.
Overview:
Due to the lack of baseline method for CHD classification, alongwith the dataset we establish one as shown in Fig. 3, which modifies and extendsthe whole heart and great artery segmentation method in CHD [18]. It includestwo subtasks: segmentation based connection analysis and similarity based shapeanalysis. Accordingly, the parts and connections most critical to the classificationare extracted.
Segmentation based connection analysis : Segmentation is performed withmultiple U-Nets [11]. There are two steps in segmentation: blood pool segmenta-tion, and chambers and initial parts of great arteries segmentation. The formeris fulfilled by a high-resolution (input size 512 ×
512 ) 2D U-net, while the lat- (d) Ground truth(c) Result of (b) with removed blood boundary(a) Blood pool with boundary Initial part A LV LARV(b) Chambers and initialparts of great arteriesBlood pool Blood boundary
Blood boundaryis removed
Separated two initial parts
Initial part B LV Fig. 4.
Connection analysis between LV/RV and great arteries (AO and PA).itle Suppressed Due to Excessive Length 5 ter is performed with a 3D low-resolution (input size 64 × ×
64 ) 3D U-net.A Region of Interest (RoI) cropping is also included with a 3D U-net beforethe 3D segmentation. With the segmentation results, connection analysis canbe processed, which mainly extracts the connection features between great ar-teries (AO and PA) and LV/RV, and between LV/LA and RV/RA. With thesegmentation results, two connection analyses between chambers, AO and PAare then performed by the connection analysis module. The first one analyzesthe connections between LV/RV and great arteries. We remove high resolutionboundary from low resolution substructures as shown in Fig. 4(a)-(c). Comparedwith the ground truth in Fig. 4(d), Fig. 4(c) shows that the two initial parts arecorrectly separated (but not in (b) where they will be treated as connected).The second one has a similar process as the first one.
Similarity based shape analysis : The flow of this subtask is shown in Fig.5. With the segmentation results, vessel extraction removes the blood pool cor-responding to chambers, and vessel refinement removes any remaining small is-lands in the image, and smooths it with erosion. Then, the skeleton of the vesselsare extracted, sampled, normalized, and fed to the shape similarity calculationmodule to obtain its similarity with all the templates in a pre-defined library.Similarity module is performed using earth mover’s distance (EMD) which is awidely used similarity metric for distributions [12]. Two factors need to be mod-eled: the weight of each bin in the distribution, and the distance between bins.We model each sampled point in the sampled skeleton as a bin, the Euclideandistance between the points as the distance between bins, and the volume ofblood pool around the sampled point as the weight of its corresponding bin.Particularly, the weight is defined as r where r is the radius of the inscribedsphere in the blood pool centered at the sampled point. The template library ismanually created in advance and contains six categories of templates correspond-ing to five types of CHDs and the normal anatomy as shown in Fig. 5, coveringall the possible shapes of great arteries in our dataset. Each category containsmultiple templates. Finally, the shape analysis module takes the skeleton and itssimilarities to obtain two kinds of features. The type of the template with thehighest similarity is extracted as the first kind. The second kind includes twoskeleton features: whether a circle exists in the skeleton, and how the r of thesample points varies. These two features are desired because if there is a circlein the skeleton, the test image is with high possibility to be classified as DAA; Ifa sampled point with a small r is connected to two sampled points with a muchlarger r , narrow vessel happens, which is a possible indication of CA and PuA. Final determination : With the extracted connection and shape features, theclassification can be finally determined using a rule-based automatic approach.Specifically, ASD and VSD have unexpected connection between LA and RA,and LV and RV, respectively. AVSD is a combination of ASD and VSD, andthe three can be classified according to the connection features between LA/LVand RA/RV. DORV has two initial parts of great arteries, both of which areconnected to RV. TOF has connected LV and RV, as well as connected LV,RV and the initial part of AO. CHD with specific shapes including CAT, DAA,
X. Xu, et al.
Initial part A LV LARVChambers and initialparts of great arteries
Initial part B
Blood pool with removed chambersVessel extraction Skeleton extractionSegmentation results Vessel refinement Skeleton samplingSampled skeleton
Extracted skeleton (red line)
Shape similarity calculation
Distribution of test image … (𝑥 , 𝑦 , 𝑧 ) (𝑥 𝑘 , 𝑦 𝑘 , 𝑧 𝑘 ) … (𝑥 𝑛 , 𝑦 𝑛 , 𝑧 𝑛 ) Distribution of template … (𝑥 , 𝑦 , 𝑧 ) (𝑥 𝑘 , 𝑦 𝑘 , 𝑧 𝑘 ) … (𝑥 𝑛 , 𝑦 𝑛 , 𝑧 𝑛 ) Earth mover’s distance
Blood pool with boundary
Blood pool Blood boundary
Skeleton sampling
Normali -zationcontaining LV and RV 𝒙 𝑵 𝒚 𝑵 𝒛 𝑵 Bounding box PA AO Shape analysis
CAT
DAA
PuA IAA …… Normal AO and PA … … PAS … … Template library
Fig. 5.
Similarity based shape analysis of great arteries. Best viewed in color.
PuA, PAS and IAA as shown in Fig. 5 can be classified by their shape features.PDA and CA are determined by analyzing the shapes and skeletons such as thevariety of r along the skeleton. DSVC can be easily classified by analyzing theskeleton of RV, and APVC is determined by the number of islands that the LAhas. Note that if the connection and shape features do not fit any of the aboverules, the classifier outputs uncertain indicating that the test image cannot behandled and manual classification is needed. Experiment setup : All the experiments run on a Nvidia GTX 1080Ti GPUwith 11 GB memory. We implement the 3D U-net using PyTorch based on [8]. For2D U-net, most configurations remain the same with those of the 3D U-net exceptthat it adopts 5 levels and the number of filters in the initial level is 16. BothDice loss and cross entropy loss are used, and the training epochs are 2 and 480for 2D U-net and 3D U-net, respectively. Data augmentation and normalizationare also adopted with the same configuration as in [8] for 3D U-net. For bothnetworks and all the analyses, three-fold cross validation is performed (about 37images for testing, and 73 images for training). We split the dataset such that alltypes of CHD are present in each subset. The classification considers a total of17 classes, including 16 types of CHD and the normal anatomy. The templatesin the template library are randomly selected from the annotated training set.In the evaluation, we use selective prediction scheme [10] and report a case asuncertain if at least one chamber is missing (which does not correspond to any itle Suppressed Due to Excessive Length 7 type in our dataset) in the segmentation results, or in the similarity calculationthe minimum EMD is larger than 0.01. For these cases, manual classification byradiologists is needed. To further evaluate how the baseline method performsagainst human experts, we also extract manual CT classification from the elec-tronic health records (the manual results can still be wrong).
Results and analysis : The CHD classification result is shown in Table 2. Eachentry (X, Y) in the table corresponds to the number of cases with ground truthclass suggested by its row header and predicted class by its column header, whereX, and Y are the results from the baseline, and those from radiologists respec-tively. Again, an image can contribute to multiple cases if it contains more thanone types of CHD. From the table we can see that for the baseline method,due to segmentation error or feature extraction failure, 22 cases are classified asuncertain, yielding a 88.4% coverage; Out of the remaining 167 cases, 137 arecorrect. Thus, for the baseline the overall classification accuracy is 72.5% forfull prediction, and 82.0% for selective prediction. For the modified baseline, theoverall classification accuracy is 39.2% for full prediction and 50.3% for selectiveprediction. On the other hand, the manual classification from experienced radi-ologist can achieve an overall accuracy of 90.5%. It is interesting to note thatout of the 17 classes, the baseline method achieves higher accuracy in one (PuA)and breaks even in four (VSD, CAT, DAA, and AAH) compared with manualclassification. In addition, Out of the 110 cases, the five radiologists only unan-imously agreed on 78 cases, which further reflects the difficulty of the problemand the value of an automated tool.The mean and standard deviation of Dice score of our baseline method forsix substructures of chambers and initial parts of great vessels segmentation,and blood pool segmentation are shown in Table 3. We can notice that bloodpool has the highest score, and initial parts of great vessels has the lowest, andthe overall segmentation performance is moderate. Though the segmentationperformance of initial parts is low, its related types of CHDs (e.g., ToF, TGA)still achieve high classification accuracy which is due to the fact that only thecritical segmentation determines the types of CHDs. Comparing the performanceof segmentation and classification, we can also notice that accurate segmentationusually helps classification, but not necessarily.
Classification success:
Six types of CHD including TGA, CAT, DAA, AAH,PAS and PuA achieve relatively high accuracy, which is due to their clear andstable features that distinguish them from normal anatomy. Such features can beeasily captured by either connection or shape features extracted by the baselinemethod. For example, CAT has a main trunk that AO and PA are both connectedto; DAA has a circular vessel which is composed of two aortic arches; PAS hasa PA with very different shape; PuA has a very thin PA without main trunk;AAH has a long period of narrow vessels in the arch; and TGA has a reversedconnection to LV and RV.
Classification failure:
Test images are classified as uncertain due to segmenta-tion error. Fig. 6 shows some examples of such error. The test image in Fig. 6(a)has very low contrast, and its blood pool and boundary are not clear compared
X. Xu, et al.
Table 2.
Number of cases (X, Y) with ground truth class and predicted class suggestedby the row and column headers respectively, where X, and Y correspond to automaticclassification by the baseline, and manual classification, respectively. Green numbersalong the diagonal suggest correct cases. (U-Uncertain, 1-ASD, 2-AVSD, 3-VSD, 4-TOF, 5-PDA, 6-TGA, 7-CA, 8-IAA, 9-PAS, 10-DORV, 11-CAT, 12-DAA, 13-APVC,14-AAH, 15-PuA, 16-DSVC, N-Normal)
Type U 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 N1 6,0 18,24 2,22 3,0 1,3 9,14 5,13 1,0 1,2 42,424 1,0 4,2 7,105 7,14 7,06 1,0 6,77 1,0 4,6 1,08 2,3 1,09 1,0 2,310 1,0 3,1 1,1 3,611 4,412 5,513 1,0 3,6 2,014 1,0 2,2 0,115 2,0 0,2 0,1 0,1 14,1016 1,0 5,7 2,1N 2,0 4,6
Table 3.
Mean and standard deviation of Dice score of our baseline method (in %)for six substructures of chambers and initial parts of great vessels segmentation, andblood pool segmentation.LV RV LA RA Initial parts of great vessels Blood pool Average77.7 74.6 77.9 81.5 66.5 86.5 75.6 ± ± ± ± ± ± ± (c) Error: TOF -> NormalSegmentationerror (d) Ground truth of (c) LALV RARV Initial part (a) Poor 3D segmentation due to low data quality (b) Ground truth of (a)
LAAO RARV PA (e) Error: TOF-> NormalNarrow-> Not narrow
Fig. 6.
Examples of classification failure: uncertain classification in (a-b), and wrongclassification of TOF in (c) and (e). Best viewed in color. with other areas, resulting in segmentation error: compared with the groundtruth in Fig. 6(b), only RA and part of the initial parts of great arteries aresegmented. As for the cases where a CHD type is predicted but wrong, we willuse TOF as examples, and leave the comprehensive discussion for all classes inthe supplementary material. Segmentation error around the initial parts of greatarteries is the main reason of the classification failure of TOF as shown in Fig. 6.Compared with the ground truth in Fig. 6(d), the 3D segmentation in Fig. 6(c) itle Suppressed Due to Excessive Length 9 labels part of LV as RV, resulting in the initial part only connected to RV ratherthan RV and LV. As one of the main features of TOF is that one initial part isconnected to both RV and LV, missing such feature leads to misclassification ofTOF as VSD. Another main feature of TOF is the narrow vessels in the initialpart and its connected RV part, which can also lead to wrong classification if notdetected correctly as shown in Fig. 6(e). A precise threshold to decide whetherthe vessels are narrow or not is still missing even in clinical studies.
Discussion:
We can notice that segmentation accuracy is important for suc-cessful classification of CHD. Higher segmentation accuracy can lead to betterconnection and shape feature extraction. In addition, so far we have only con-sidered the connection features in the blood pool and the shapes of the vessels.More structural features associated with classification should be considered toimprove the performance, which due to the lack of local tissue changes, need in-novations from the deep learning community and deeper collaboration betweencomputer scientists and radiologists.
We introduce to the community the ImageCHD dataset [1] in hopes of encourag-ing new research into unique, difficult and meaningful datasets. We also present abaseline method for comparison on this new dataset, based on a state-of-the-artwhole-heart and great artery segmentation method for CHD images. Experimen-tal results show that under selective prediction scheme the baseline method canachieve a classification accuracy of 82.0%, leaving big room for improvement. Wehope that the dataset and the baseline method can encourage new research thatbe used to better address not only the CHD classification but also a wider class ofproblems that have large global structural change but little local texture/featurechange.
References
1. https://github.com/XiaoweiXu/ImageCHD-A-3D-Computed-Tomography-Image-Dataset-for-Classification-of-Congenital-Heart-Disease2. Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin,I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques forautomatic mri cardiac multi-structures segmentation and diagnosis: Is the problemsolved? IEEE transactions on medical imaging (11), 2514–2525 (2018)3. Bhat, V., BeLaVaL, V., Gadabanahalli, K., Raj, V., Shah, S.: Illustrated imagingessay on congenital heart diseases: multimodality approach part i: clinical perspec-tive, anatomy and imaging techniques. Journal of clinical and diagnostic research:JCDR (5), TE01 (2016)4. Dou, Q., Ouyang, C., Chen, C., Chen, H., Glocker, B., Zhuang, X., Heng, P.A.:Pnp-adanet: Plug-and-play adversarial domain adaptation network at unpairedcross-modality cardiac segmentation. IEEE Access , 99065–99076 (2019)5. Habijan, M., Leventi´c, H., Gali´c, I., Babin, D.: Whole heart segmentation from ctimages using 3d u-net architecture. In: 2019 International Conference on Systems,Signals and Image Processing (IWSSIP). pp. 121–126. IEEE (2019)0 X. Xu, et al.6. Liu, T., Tian, Y., Zhao, S., Huang, X., Wang, Q.: Automatic whole heart segmen-tation using a two-stage u-net framework and an adaptive threshold window. IEEEAccess , 83628–83636 (2019)7. Pace, D.F., Dalca, A.V., Brosch, T., Geva, T., Powell, A.J., Weese, J., Moghari,M.H., Golland, P.: Iterative segmentation from limited training data: Applicationsto congenital heart disease. In: Deep Learning in Medical Image Analysis andMultimodal Learning for Clinical Decision Support, pp. 334–342. Springer (2018)8. Payer, C., ˇStern, D., Bischof, H., Urschler, M.: Multi-label whole heart segmenta-tion using cnns and anatomical label configurations. In: International Workshop onStatistical Atlases and Computational Models of the Heart. pp. 190–198. Springer(2017)9. Piccini, D., Littmann, A., Nielles-Vallespin, S., Zenge, M.O.: Respiratory self-navigation for whole-heart bright-blood coronary mri: methods for robust isolationand automatic segmentation of the blood pool. Magnetic resonance in medicine (2), 571–579 (2012)10. Pidan, D., El-Yaniv, R.: Selective prediction of financial trends with hidden markovmodels. In: Advances in Neural Information Processing Systems. pp. 855–863(2011)11. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi-cal image segmentation. In: International Conference on Medical image computingand computer-assisted intervention. pp. 234–241. Springer (2015)12. Rubner, Y., Tomasi, C., Guibas, L.J.: The earth mover’s distance as a metric forimage retrieval. International journal of computer vision (2), 99–121 (2000)13. Wang, C., MacGillivray, T., Macnaught, G., Yang, G., Newby, D.: A two-stage3d unet framework for multi-class segmentation on full resolution image. arXivpreprint arXiv:1804.04341 (2018)14. Wang, T., Xiong, J., Xu, X., Jiang, M., Yuan, H., Huang, M., Zhuang, J., Shi, Y.:Msu-net: Multiscale statistical u-net for real-time 3d cardiac mri video segmenta-tion. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 614–622. Springer (2019)15. Wang, T., Xiong, J., Xu, X., Shi, Y.: SCNN: A general distribution based statisticalconvolutional neural network with application to video object detection. arXivpreprint arXiv:1903.07663 (2019)16. Wolterink, J.M., Leiner, T., Viergever, M.A., Iˇsgum, I.: Dilated convolutional neu-ral networks for cardiovascular mr segmentation in congenital heart disease. In: Re-construction, segmentation, and analysis of medical images, pp. 95–102. Springer(2016)17. Xu, X., Lu, Q., Yang, L., Hu, S., Chen, D., Hu, Y., Shi, Y.: Quantization offully convolutional networks for accurate biomedical image segmentation. In: Pro-ceedings of the IEEE conference on computer vision and pattern recognition. pp.8300–8308 (2018)18. Xu, X., Wang, T., Shi, Y., Yuan, H., Jia, Q., Huang, M., Zhuang, J.: Whole heartand great vessel segmentation in congenital heart disease using deep neural net-works and graph matching. In: International Conference on Medical Image Com-puting and Computer-Assisted Intervention. pp. 477–485. Springer (2019)19. Xu, Z., Wu, Z., Feng, J.: Cfun: Combining faster r-cnn and u-net network forefficient whole heart segmentation. arXiv preprint arXiv:1812.04914 (2018)20. Yang, X., Bian, C., Yu, L., Ni, D., Heng, P.A.: Class-balanced deep neural networkfor automatic ventricular structure segmentation. In: International Workshop onStatistical Atlases and Computational Models of the Heart. pp. 152–160. Springer(2017)itle Suppressed Due to Excessive Length 1121. Yang, X., Bian, C., Yu, L., Ni, D., Heng, P.A.: Hybrid loss guided convolutionalnetworks for whole heart parsing. In: International Workshop on Statistical Atlasesand Computational Models of the Heart. pp. 215–223. Springer (2017)22. Ye, C., Wang, W., Zhang, S., Wang, K.: Multi-depth fusion network for whole-heartct image segmentation. IEEE Access , 23421–23429 (2019)23. Yu, L., Yang, X., Qin, J., Heng, P.A.: 3d fractalnet: dense volumetric segmentationfor cardiovascular mri volumes. In: Reconstruction, segmentation, and analysis ofmedical images, pp. 103–110. Springer (2016)24. Zhang, R., Chung, A.C.: A fine-grain error map prediction and segmentation qual-ity assessment framework for whole-heart segmentation. In: International Con-ference on Medical Image Computing and Computer-Assisted Intervention. pp.550–558. Springer (2019)25. Zheng, H., Yang, L., Han, J., Zhang, Y., Liang, P., Zhao, Z., Wang, C., Chen,D.Z.: Hfa-net: 3d cardiovascular image segmentation with asymmetrical poolingand content-aware fusion. In: International Conference on Medical Image Comput-ing and Computer-Assisted Intervention. pp. 759–767. Springer (2019)26. Zhou, Z., Guo, X., Yang, W., Shi, Y., Zhou, L., Wang, L., Yang, M.: Cross-modalattention-guided convolutional network for multi-modal cardiac segmentation. In:International Workshop on Machine Learning in Medical Imaging. pp. 601–610.Springer (2019)27. Zhuang, X., Shen, J.: Multi-scale patch and multi-modality atlases for whole heartsegmentation of mri. Medical image analysis31