Multi-Stage Region-Based Neural Networks for Fine-Grained Glaucoma Screening

Eunjin Kim, Seunghoon Lee, Gitaek Kwon, Hyunwoo Kim, Jaeyoung Kim

Abstract: Glaucoma, the leading cause of irreversible blindness globally, presents substantial challenges due to its reliance on subjective decision-making processes. Recent advancements in deep learning (DL) have shown promising potential to automate screening tools. However, most DL-based methods primarily assess the need for referral without shedding light on specific glaucomatous features or abnormalities. This paper introduces a robust DL pipeline capable of not only determining the need for referral in glaucoma cases (referable vs. non-referable) but also accurately identifying 10 specific glaucoma-related characteristics. Our framework consists of three models: (1) a segmentation model to force the glaucoma classification model to focus on the optic disc region, (2) a glaucoma classification model, and (3) a classification model to classify the 10 features involved in glaucoma. We also apply ensemble techniques to improve the generalization performance of the models. We demonstrate the effectiveness of our proposed algorithm by participating in the JustRAIGS challenge as a team called vuno. The proposed method achieves a recall of 0.9090 (at 95% specificity) for the glaucoma classification task and a Hamming loss of 0.1240 for classifying glaucoma-related characteristics. Code is available at https://github.com/seunghoonlee-fundus/JustRAIGS-vuno.