Using a Deep Learning Model to Predict Postoperative Visual Outcomes of Idiopathic Epiretinal Membrane Surgery

• Published in: American journal of ophthalmology Vol. 272; pp. 67 - 78
• Main Authors: LIN, HSIN-LE, TSENG, PO-CHEN, HSU, MIN-HUEI, PENG, SYU-JYUN
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.04.2025
• Subjects: Aged; Aged, 80 and over; Algorithms; Deep Learning; Epiretinal Membrane - diagnosis; Epiretinal Membrane - diagnostic imaging; Epiretinal Membrane - physiopathology; Epiretinal Membrane - surgery; Female; Humans; Male; Middle Aged; Postoperative Period; Retrospective Studies; ROC Curve; Tomography, Optical Coherence - methods; Visual Acuity - physiology; Vitrectomy - methods; ERM; OCT; GCIPL; CNN; P-IM; IPL; RPE; L-IM; Grad-CAM; INL; DL; AI; AUROC; EIFL; OPL; ILM; ROI; DRIL; COST; BCVA; SGDM
• ISSN: 0002-9394; 1879-1891; 1879-1891
• DOI: 10.1016/j.ajo.2025.01.003

This study assessed the performance of various deep learning models in predicting the postoperative outcomes of idiopathic epiretinal membrane (ERM) surgery based on preoperative optical coherence tomography (OCT) images. Validation of algorithms to predict the outcomes of ERM surgery based on OCT data. Internal training and validation were performed using 1,392 OCT images from 696 eyes. External testing was performed using 152 OCT images from 76 eyes. This study assessed three deep learning models, including Inception-v3, ResNet-101, and VGG-19. Grad-CAM was employed for hotspot analysis. The dataset was split into a training set (80%) and a validation set (20%). Subjects presenting an improvement of ≥2 lines on the Snellen chart at 1-year postsurgery were classified as pronounced visual improvement, whereas those presenting an improvement of <2 lines were classified as limited visual improvement. Using an external test dataset, we compared assessments by seven ophthalmologists with the prediction of deep learning model. The main outcome measures were recall, specificity, precision, F1 score, accuracy, and area under the receiver operating characteristic curve (AUROC). ResNet-101 achieved the best overall performance, as evidenced by the following metrics: recall (0.90), specificity (0.90), precision (0.91), F1-score (0.90), accuracy (0.90), and AUROC (0.97). In Grad-CAM heatmap analysis, the logic of ResNet-101 closely resembled that of clinical physicians. Overall, the performance of this deep learning model was significantly better than that of general ophthalmologists and non-retina specialists and was slightly superior to that of retina specialists. Deep learning based on preoperative OCT images proved highly effective in predicting the outcomes of ERM surgery and elucidating the structural mechanisms underlying the phenomena observed in OCT images.