Artificial intelligence for epiretinal membrane detection using ultrawidefield scanning laser ophthalmoscopy

There is currently no validated artificial intelligence (AI) model for detecting epiretinal membrane (ERM) using ultrawidefield scanning laser ophthalmoscopy (UWF-SLO). To address this gap, we aimed to develop and validate a dedicated deep learning (DL) model for automated ERM detection on UWF-SLO images.

In this retrospective study, an optimized Inception-v3 model was developed using 920 original UWF-SLO images. We systematically compared models trained on two distinct fields of view: the macular and posterior pole regions. The models were evaluated on an internal testing set and an independent, multicentre external validation set. Their performance was benchmarked against that of three ophthalmologists with different levels of expertise. Furthermore, an AI-assisted reader study was conducted in which the same ophthalmologists re-evaluated the images using model-generated heatmaps.

The macula AI model achieved a specificity of 93.1%, a sensitivity of 84.0%, and an AUC of 0.943, while the posterior pole AI model achieved corresponding values of 88.6%, 86.0%, and 0.931, respectively. Both models maintained good performance in the external validation set, achieving AUCs of 0.937 and 0.949, respectively. In the reader study, AI assistance improved the average diagnostic accuracy of the three ophthalmologists from 75.4 to 83.3%, with the greatest improvement observed in the resident ophthalmologist.

We present the first well-validated DL model for ERM detection on UWF-SLO images. The model performed well in both internal and external validation and could be used to support the clinical screening and detection of ERMs.