在臨床上，臨床醫師常常需要依賴心臟超音波來判斷病患的心臟結構與功能是否有無正常，加上超音波檢查所特有之無痛與非侵入性的特性，使其常常為臨床上進一步分析心臟功能的第一選擇工具。也因此，由於臨床需求量的大幅增加，使得臨床心臟科醫師判讀的負擔亦相形增加，若每筆檢查的單位判讀時間因而被壓縮，長此以往，容易衍生出漏判或效率下降等情況。
近年來，由於深度學習使用於影像分析技術的持續進步，使得自動化靜態或動態影像分析的效能與準確率逐漸得到信賴，因而本論文將提出以深度學習架構模型，判斷部分心臟超音波檢查的重點：心臟收縮異常功能區塊，用以來協助臨床醫師做判讀，以期望能提升臨床判讀效率以維持足夠的準確度。

Deep neural network–assisted automated interpretation of echocardiography results confirmed by cardiologists is a gradually emerging topic. Unlike other studies, which have used still frame–level spatial relationships, we used a deep neural network model for echocardiographic video analysis; the model applies spatial and temporal information simultaneously for the automated recognition of left ventricle myocardial regional wall motion abnormalities.
We collected echocardiography series obtained from July 2017 to April 2018 from 2 hospitals. Regional wall abnormalities were confirmed by qualified cardiologists. We first developed a 3-dimensional (3D) convolutional neural network (CNN) model for view selection using the view prediction confidence level, thus ensuring stringent image quality control. We developed a second model, U-net, to segment images to annotate the location of each wall. Finally, we developed a third major model, a 3D CNN model; the model inputs comprised 4 views before and after the segmentation of echocardiographic videos, and its outputs comprised final labels representing either the absence or presence of motion abnormalities in each wall. Moreover, to examine model stability, we subjected the major model to 5-fold cross-validation tests and external validation.
We collected 10 638 echocardiography series and acquired 4 main views with sufficient minimum confidence levels, thus generating 6454 echocardiography series as the training dataset. For external validation, we collected an additional 2819 echocardiography series from another hospital. Among these images, we annotated 2740 frames to develop the U-net model, which achieved a Dice similarity coefficient of 0.756. Coupled with the segmentation model, the final model effectively recognized regional wall motion abnormalities; for the cross-validation and external validation datasets, the final model achieved average sensitivity levels of 84.0% ± 3.5% and 81.6% ± 1.5%, specificity levels of 86.8% ± 1.0% and 83.6% ± 7.9%, accuracy levels of 86.6% ± 0.9% and 81.5% ± 1.4%, and average area under the receiver operating characteristic curve values of 0.912 ± 0.008 and 0.891 ± 0.029respectively.
With appropriate image quality confidence levels, deep neural networks constitute a stable and feasible approach for video-level interpretation and recognition of regional wall motion abnormalities. Further investigation is required to improve model performance and clinical applications.

1. McGowan JH, Martin W, Burgess MI, McCurrach G, Ray SG, McDonagh TA and Cleland JG. Validation of an echocardiographic wall motion index in heart failure due to ischaemic heart disease. Eur J Heart Fail. 2001;3:731-7.
2. McGowan JH and Cleland JGF. Reliability of reporting left ventricular systolic function by echocardiography: A systematic review of 3 methods. Am Heart J. 2003;146:388-397.
3. Fihn SD, Blankenship JC, Alexander KP, Bittl JA, Byrne JG, Fletcher BJ, Fonarow GC, Lange RA, Levine GN, Maddox TM, Naidu SS, Ohman EM and Smith PK. 2014 ACC/AHA/AATS/PCNA/SCAI/STS focused update of the guideline for the diagnosis and management of patients with stable ischemic heart disease: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines, and the American Association for Thoracic Surgery, Preventive Cardiovascular Nurses Association, Society for Cardiovascular Angiography and Interventions, and Society of Thoracic Surgeons. J Am Coll Cardiol. 2014;64:1929-49.
4. Choi E, Schuetz A, Stewart WF and Sun J. Using recurrent neural network models for early detection of heart failure onset. J Am Med Inform Assoc. 2017;24:361-370.
5. Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, Liu PJ, Liu X, Marcus J, Sun M, Sundberg P, Yee H, Zhang K, Zhang Y, Flores G, Duggan GE, Irvine J, Le Q, Litsch K, Mossin A, Tansuwan J, Wang D, Wexler J, Wilson J, Ludwig D, Volchenboum SL, Chou K, Pearson M, Madabushi S, Shah NH, Butte AJ, Howell MD, Cui C, Corrado GS and Dean J. Scalable and accurate deep learning with electronic health records. npj Digital Medicine. 2018;1.
6. Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng L and Webster DR. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature Biomedical Engineering. 2018;2:158-164.
7. Madani A, Ong JR, Tibrewal A and Mofrad MRK. Deep echocardiography: data-efficient supervised and semi-supervised deep learning towards automated diagnosis of cardiac disease. npj Digital Medicine. 2018;1.
8. Zhang J, Gajjala S, Agrawal P, Tison GH, Hallock LA, Beussink-Nelson L, Lassen MH, Fan E, Aras MA, Jordan C, Fleischmann KE, Melisko M, Qasim A, Shah SJ, Bajcsy R and Deo RC. Fully Automated Echocardiogram Interpretation in Clinical Practice. Circulation. 2018;138:1623-1635.
9. Tsai WC, Liu YW, Huang YY, Lin CC, Lee CH and Tsai LM. Diagnostic value of segmental longitudinal strain by automated function imaging in coronary artery disease without left ventricular dysfunction. J Am Soc Echocardiogr. 2010;23:1183-9.
10. Shalbaf A, Behnam H, Alizade-Sani Z and Shojaifard M. Automatic assessment of regional and global wall motion abnormalities in echocardiography images by nonlinear dimensionality reduction. Med Phys. 2013;40:052904.
11. Vermes E, Guyon P, Weingrod M, Otmani A, Soussana C, Halphen C, Leroy G and Haiat R. Assessment of left ventricular regional wall motion by color kinesis technique: comparison with angiographic findings. Echocardiography. 2000;17:521-7.
12. Kusunose K, Abe T, Haga A, Fukuda D, Yamada H, Harada M and Sata M. A Deep Learning Approach for Assessment of Regional Wall Motion Abnormality From Echocardiographic Images. JACC Cardiovasc Imaging. 2019.
13. Bryan C. Russell AT, Kevin P. Murphy, William T. Freeman. LabelMe: A Database and Web-Based Tool for Image Annotation | SpringerLink. 2008.
14. Lang RM, Badano LP, Mor-Avi V, Afilalo J, Armstrong A, Ernande L, Flachskampf FA, Foster E, Goldstein SA, Kuznetsova T, Lancellotti P, Muraru D, Picard MH, Rietzschel ER, Rudski L, Spencer KT, Tsang W and Voigt JU. Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. J Am Soc Echocardiogr. 2015;28:1-39 e14.
15. Falk T, Mai D, Bensch R, Çiçek Ö, Abdulkadir A, Marrakchi Y, Böhm A, Deubner J, Jäckel Z, Seiwald K, Dovzhenko A, Tietz O, Dal Bosco C, Walsh S, Saltukoglu D, Tay TL, Prinz M, Palme K, Simons M, Diester I, Brox T and Ronneberger O. U-Net: deep learning for cell counting, detection, and morphometry. Nature Methods. 2019;16:67-70.
16. Huang G, Liu Z, van der Maaten L and Weinberger KQ. Densely Connected Convolutional Networks. 2016.
17. Lee W-H, Liu Y-W, Yang L-T and Tsai W-C. Prognostic value of longitudinal strain of subepicardial myocardium in patients with hypertension. 2016;34:1195-1200.
18. Huang M-s, Lee W-H, Tsai H-R, Liu Y-W, Liu P-Y and Tsai W-C. Value of layer-specific strain distribution patterns in hypertrophied myocardium from different etiologies. Int J Cardiol. 2019;281:69-75.
19. Tan M and Le QV. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv e-prints. 2019.
20. Bianco S, Cadene R, Celona L and Napoletano P. Benchmark Analysis of Representative Deep Neural Network Architectures. arXiv e-prints. 2018.