Feasibility study of deep-learning-based models for automated macular hole segmentation using optical coherence tomography B-scans

We aim to evaluate the feasibility of deep-learning (DL)-based models for automated macular hole (MH) segmentation using optical coherence tomography (OCT) B-scans. Four DL-based OCT B-scan segmentation models with different architectures, including ResNet-50 or EfficientNet backbones, feature pyramid network (FPN) or weighted bi-directional FPN (BiFPN) necks, and FPN head, namely Model 1 (ResNet-50, FPN, FPN), Model 2 (ResNet-50, BiFPN, FPN), Model 3 (EfficientNet, FPN, FPN), and Model 4 (EfficientNet, BiFPN, FPN), were developed, followed by their implementations on 3295 paired OCT B-scans and labels belonging to MH and non-MH. The mean intersection over union (IOU) values (0.969, 0.968, 0.975, and 0.977 for Models 1–4) indicate the best performance of Model 4. The percentage proportion of OCT B-scans with IOU values between 0.8 and 1 reached 94.33%, 95.32%, 100%, and 100% for MH, choroid, retina, and intraretinal cysts (IRCs), respectively, demonstrating the potential of the four DL-based models developed for automated MH segmentation from OCT B-scans.