Nam Yeo Kang1, Ho Ra1, Kook Lee2, Jun Hyuk Lee1, Won Ki Lee3, Jiwon Baek4. 1. Department of Ophthalmology, Bucheon St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Gyeonggi-do, Republic of Korea. 2. Department of Ophthalmology, Seoul St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea. 3. Retina Division, Nune Eye Center, Seoul, Republic of Korea. 4. Department of Ophthalmology, Bucheon St. Mary's Hospital, College of Medicine, The Catholic University of Korea, Gyeonggi-do, Republic of Korea. md.jiwon@gmail.com.
Abstract
PURPOSE: Pachychoroid is characterized by dilated Haller vessels and choriocapillaris attenuation that are seen on optical coherence tomography (OCT) B-scans. This study investigated the feasibility of using deep learning (DL) models to classify pachychoroid and non-pachychoroid eyes from OCT B-scan images. METHODS: In total, 1898 OCT B-scan images were collected from eyes with macular diseases. Images were labeled as pachychoroid or non-pachychoroid based on strict quantitative and qualitative criteria for multimodal imaging analysis by two retina specialists. DL models were trained (80%) and validated (20%) using pretrained convolutional neural networks (CNNs). Model performance was assessed using an independent test set of 50 non-pachychoroid and 50 pachychoroid images. RESULTS: The final accuracy of AlexNet and VGG-16 was 57.52% for both models. ResNet50, Inception-v3, Inception-ResNet-v2, and Xception showed a final accuracy of 96.31%, 95.25%, 93.40%, and 92.61%, respectively, for the validation set. These models demonstrated accuracy on an independent test set of 78.00%, 86.00%, 90.00%, and 92.00%, and an F1 score of 0.718, 0.841, 0.894, and 0.920, respectively. CONCLUSION: DL models classified pachychoroid and non-pachychoroid images with good performance. Accurate classification can be achieved using CNN models with deep rather than shallow neural networks.
PURPOSE: Pachychoroid is characterized by dilated Haller vessels and choriocapillaris attenuation that are seen on optical coherence tomography (OCT) B-scans. This study investigated the feasibility of using deep learning (DL) models to classify pachychoroid and non-pachychoroid eyes from OCT B-scan images. METHODS: In total, 1898 OCT B-scan images were collected from eyes with macular diseases. Images were labeled as pachychoroid or non-pachychoroid based on strict quantitative and qualitative criteria for multimodal imaging analysis by two retina specialists. DL models were trained (80%) and validated (20%) using pretrained convolutional neural networks (CNNs). Model performance was assessed using an independent test set of 50 non-pachychoroid and 50 pachychoroid images. RESULTS: The final accuracy of AlexNet and VGG-16 was 57.52% for both models. ResNet50, Inception-v3, Inception-ResNet-v2, and Xception showed a final accuracy of 96.31%, 95.25%, 93.40%, and 92.61%, respectively, for the validation set. These models demonstrated accuracy on an independent test set of 78.00%, 86.00%, 90.00%, and 92.00%, and an F1 score of 0.718, 0.841, 0.894, and 0.920, respectively. CONCLUSION: DL models classified pachychoroid and non-pachychoroid images with good performance. Accurate classification can be achieved using CNN models with deep rather than shallow neural networks.
Authors: Philippe Valmaggia; Philipp Friedli; Beat Hörmann; Pascal Kaiser; Hendrik P N Scholl; Philippe C Cattin; Robin Sandkühler; Peter M Maloca Journal: Transl Vis Sci Technol Date: 2022-09-01 Impact factor: 3.048