Machine Learning to Predict Surgical Indications for ERM

Abstract

Purpose : Epiretinal membrane (ERM) is a fibrous membrane on the macula causing reduced visual acuity and metamorphopsia. Although the only treatment for ERM is surgical removal, there are no clear surgical criteria. Our previous work showed a significant correlation between the maximum depth of retinal folds (MDRF) in the parafoveal area and metamorphopsia, suggesting MDRF as an objective parameter for the indication criteria for ERM surgery (Kanzaki, Retina 2022). However, measuring MDRF is time-consuming because a large number of OCT images must be examined. Therefore, the purpose of this study is to investigate whether MDRF can be predicted by machine learning using OCT images.

Methods : We retrospectively enrolled 258 eyes with ERM and 126 healthy eyes. 3D images of the retina obtained with swept source OCT (Triton, Topcon) were used to construct en face images and measure MDRF. We generated two network models for predicting MDRF: one was constructed by transfer learning with EfficientNetV2 by whose input was three en face images per eye, and the other was LSTM for MDRF prediction with seven en face images using. Five-fold cross-validation was used to evaluate the mean absolute error (MAE) and the accuracy of the classification using previously reported surgical indication criteria (MDRF>69um).

Results : The median (minimum and maximum) MDRF of the eyes with ERM was 75.4 (0 and 223.6) μm. The MAE was 13.1 µm for EfficientNetV2 and 11.7 µm for LSTM. The classification accuracy of EfficientNetV2 was 89.8% (sensitivity 87.0%, specificity 91.6%) and that of LSTM was 88.0% (sensitivity 75.1%, specificity 95.8%).

Conclusions : Machine learning using OCT en face images was effective in predicting MDRF and classifying patients suitable for ERM surgery.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.