Medically Explainable Deep Learning-Based Glaucoma Diagnosis

Recent advances in deep learning-based models have demonstrated great potential as effective, low-cost solutions for glaucoma screening and diagnosis. However, the "black box" nature of these models, resulting from their lack of explainability, undermines trust and poses a barrier to their adoption in clinical practice. This study presents a novel approach to addressing this challenge by introducing a multitask deep learning model designed to enhance medical explainability in glaucoma diagnosis. Unlike traditional single-task learning models, the proposed model simultaneously performs multiple tasks: image-level glaucoma classification, pixel-level segmentation of the optic cup (OC) and optic disc (OD) to compute the vertical cup-to-disc ratio (vCDR), and image-level classification of peripapillary atrophy (PPA). Evaluated using the Retinal Fundus Glaucoma Challenge (REFUGE) database, the proposed model achieves superior performance across all tasks compared to baseline single-task models, demonstrating the model's ability to leverage shared information among related tasks to enhance overall performance. More importantly, the model enhances medical explainability by generating glaucoma-related, clinically relevant biomarkers, such as the vCDR and PPA. These biomarkers provide support and validation for the glaucoma diagnosis, making the model's predictions more interpretable for both healthcare professionals and patients. This study contributes to the development of more transparent, reliable, and trustworthy AI systems for healthcare applications.Clinical Relevance- The proposed model can support clinicians in making more informed decisions by providing interpretable diagnosis supported by clinically relevant biomarkers for glaucoma screening and diagnosis.