Automated macular hole staging on optical coherence tomography using an optimized ResNet-18 framework

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Macular holes (MH) require accurate staging for effective treatment.
• Optical coherence tomography (OCT) is crucial for visualizing MH.
• Automated staging can improve efficiency and consistency in clinical practice.

Purpose of the Study:

• To develop and validate a deep learning model for automated macular hole (MH) staging.
• To compare the performance of different deep learning architectures for MH staging.
• To optimize a ResNet-based model for enhanced accuracy and stability.

Main Methods:

• A retrospective study utilized 6,243 OCT images of macular holes.
• ResNet-based and Swin Transformer-based deep learning models were developed and compared.
• ResNet-18 was optimized with deformable convolution (DCNv2) and efficient channel attention (ECA).

Main Results:

• The optimized ResNet-18 achieved high performance metrics: 98.23% accuracy, 97.00% F1-score, and 99.76% AUC.
• The optimized ResNet-18 demonstrated stable and consistent classification across validation folds.
• External validation confirmed the model's consistent classification accuracy on independent datasets.

Conclusions:

• An optimized ResNet-18 model provides accurate and stable automated macular hole staging using OCT images.
• The model shows superior reliability compared to deeper CNN and Transformer architectures, especially with limited data.
• These findings support the clinical feasibility of automated MH staging via OCT imaging.