MVGL-Net: A generalizable multi-view convolutional network for anterior segment OCT

Abstract

ObjectiveCorrect diagnosis of primary angle-closure disease (PACD) and classification of plateau iris configuration (PIC) existence is essential for appropriate glaucoma treatment. A fast non-contact method that can automatically detect PACD and PIC would be valuable for improving treatment decisions and reducing the misdiagnosis rate. MethodsIn this work, we propose an automated method to analyze anterior segment optical coherence tomography (AS-OCT) images and test it on two tasks using two multi-institution datasets containing more than 12,000 image slices from more than 400 eyes. Our method, called multi-view glaucoma network (MVGL-Net), leverages the multi-view nature of AS-OCT images for PACD diagnosis and, more importantly, can detect PIC with minimal transfer learning. ResultsThe MVGL-Net achieves comparable prediction results to other published models when used for PACD diagnosis. In the PIC diagnosis task, MVGL-Net significantly outperforms the existing models on PACD and achieves performance comparable to manual measurements on the AS-OCT images reported in the literature with an AUC of 0.71. ConclusionsThe experimental results demonstrate that the MVGL-Net model is a fast, generalizable approach that, when compared with existing models, achieves comparable results for angle-closure detection and improved results for PIC diagnosis, a more challenging medical imaging task.