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

To develop and validate a deep learning model for automated macular hole (MH) staging using optical coherence tomography (OCT) images.

This retrospective study included OCT-confirmed MH images collected between July 2019 and October 2024. After quality control and redundancy screening, 6,243 representative OCT images were retained. Highly similar consecutive scans were excluded through combined manual review and program-assisted similarity assessment before model development. Models were developed using standardized preprocessing and five-fold cross-validation. ResNet-based and Swin Transformer-based architectures were constructed and compared. ResNet-18 was further optimized by integrating deformable convolution (DCNv2) and efficient channel attention (ECA).

Swin Transformer Tiny achieved high training accuracy but showed unstable performance on the test set. In contrast, the optimized ResNet-18 demonstrated more consistent classification results across validation folds. The optimized ResNet-18 achieved accuracy, precision, recall, F1-score, specificity, and AUC values of 98.23%, 97.17%, 96.90%, 97.00%, 99.44%, and 99.76%, respectively. External validation showed consistent classification accuracy on independent data.

The optimized ResNet-18 achieved accurate and stable MH staging using OCT images under a standardized training framework. Compared with deeper CNN and Transformer-based architectures, the proposed model demonstrated more reliable performance under limited-data conditions. These findings support the feasibility of automated MH staging using OCT images.