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

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Macular holes (MH) are significant retinal conditions requiring accurate segmentation for diagnosis and treatment.
• Automated segmentation of MH using optical coherence tomography (OCT) B-scans can improve efficiency and consistency in clinical practice.

Purpose of the Study:

• To evaluate the feasibility of deep learning (DL) models for automated macular hole (MH) segmentation.
• To compare the performance of four different DL architectures for MH segmentation on OCT B-scans.

Main Methods:

• Four DL models were developed using ResNet-50 or EfficientNet backbones and FPN or BiFPN necks.
• Models were implemented and evaluated on 3295 paired OCT B-scans and labels for MH and non-MH cases.
• Performance was assessed using mean intersection over union (IOU) values.

Main Results:

• All four DL models demonstrated high performance in automated MH segmentation.
• Model 4 (EfficientNet, BiFPN, FPN) achieved the highest mean IOU of 0.977.
• High IOU values (0.8-1) were achieved for MH, choroid, retina, and intraretinal cysts (IRCs), with Models 3 and 4 reaching 100% for some structures.

Conclusions:

• Deep learning models are feasible for automated MH segmentation from OCT B-scans.
• The developed models, particularly Model 4, show strong potential for clinical application in MH diagnosis and management.
• Further validation on larger datasets is warranted to confirm clinical utility.