Explainable-AI guided Deep Learning for Glaucoma Detection: Feature-Weighting Approach using SHAP Values

Glaucoma is the leading cause of irreversible blindness, affecting over 70 million people. While deep learning models offer great diagnostic performance, their ‘black-box’ nature acts as a barrier to clinical implementations, as clinicians require transparency in the decision-making process before trusting the automated systems entirely. While post-hoc methods such as Grad-Cam, LIME and SHAP can help visualize results and provide explanations, they cannot influence or control the way the model focuses on features. In this study, we propose a model-agnostic SHAP-guided weight adjustment framework that shifts explainability from a retrospective tool to an active training tool. By utilizing the SHAP values of each feature and augmenting the classifier weights with said values, we steer the model to focus on relevant features and ignore noise. Evaluated on the EyePACS-AIROGS-light-V2 dataset, our modified models (DenseNet-121 and ResNet-50) demonstrated consistent improvement across key metrics. Quantitative interpretability analysis confirmed progressive alignment between SHAP values and model weights, demonstrating consistent and stable feature prioritization.