Deep Ensemble Learning to Detect Retinal Vascular Leakage on Ultrawide-Field Fundus Photographs of Patients With Uveitis.

Purpose: The purpose of this study was to develop a novel deep learning (DL) algorithm to detect retinal vascular leakage (RVL) on ultra-widefield fundus (UWF) images in patients with posterior segment uveitis. Methods: Ultra-widefield fluorescein angiography (UWFFA) and corresponding UWF images of patients, who were evaluated at the uveitis clinic at the National Eye Institute, National Institutes of Health, were collected for this study. UWFFA images were used for the assessment of RVL, and the corresponding UWF images were used to train the algorithms. Several DL algorithms with different backbone architectures were trained and tested, and the ensemble learning (EL) method was adopted to enhance classification accuracy. Results: A total of 405 eyes were included in the study. Two different datasets were generated, wMildRVL (405 eyes) and woMildRVL (excluding mild RVL eyes). EL based on 3 DL models showed superior performance with an accuracy of 0.7704, a sensitivity of 0.7699, a specificity of 0.7713, and an area under the curve (AUC) of 0.8018 for the dataset wMildRVL, and an accuracy of 0.7900, a sensitivity of 0.7819, a specificity of 0.8000, and an AUC of 0.8344 for the woMildRVL dataset. Conclusions: The proposed EL model demonstrated the potential in distinguishing those with and without RVL on UWF images from eyes with posterior segment uveitis. Translational Relevance: This algorithm model can be a potential screening tool to detect the presence of RVL on UWF images, thus determining the need for UWFFA, as this would be especially helpful in resource-limited settings or in patients with known adverse effects to the fluorescein dye.