Deep Multi-Task Learning for Interpretable Glaucoma Detection

Abstract

Glaucoma is one of the leading causes of blindness worldwide. The rising prevalence of glaucoma, with our aging population, increases the need to develop automated systems that can aid physicians in early detection, ultimately preventing vision loss. Clinical interpretability and adequately labeled data present two major challenges for existing deep learning algorithms for automated glaucoma screening. We propose an interpretable multi-task model for glaucoma detection, called Interpretable Glaucoma Detector (InterGD). InterGD is composed of two major complementary components, segmentation and prediction modules. The segmentation module addresses the lack of clinical interpretability by locating the optic disc and optic cup regions in a fundus image. The prediction module utilizes a larger dataset to improve the performance of the segmentation task and thus mitigate the problem of limited labeled data in a segmentation module. The two components are effectively integrated into a unified multi-task framework allowing end-to-end training. To the best of our knowledge, this work is the first to incorporate interpretability into glaucoma screening employing deep learning methods. The experiments on three datasets, two public and one private, demonstrate the effectiveness of InterGD in achieving interpretable results for glaucoma screening.