Development and Validation of a Deep Learning Model for Macular Hole Staging from Optical Coherence Tomography

The experimental results showed that the training set validation accuracy of the Swin Transformer Tiny (Swin-T) model built on the ViT architecture was 99%, but the highest accuracy on the test set did not exceed 90%. Therefore, the ResNet-18 model based on CNN architecture was ultimately evaluated and tested under five fold cross validation. The overall accuracy of each fold was 97.84%, the average accuracy was 97.11%, the average recall or sensitivity was 96.08%, the average F1 score was 96.54%, the average specificity was 99.29%, and the average AUC value was 99.78%.

Conclusion: The macular hole staging method based on ResNet-18 model proposed in this article achieves accurate and stable judgment of macular hole staging in fundus OCT images while ensuring lightweight model structure and training efficiency, and has strong potential for clinical auxiliary diagnostic applications.