A Spatial Attention and Multi-Scale Fusion Model for Enhanced Eye Disease Classification with Explainable AI

Eye diseases lead to vision impairment if not detected early. Fundus imaging provides a non-invasive method for diagnosis, but manual interpretation requires specialized expertise and is time-consuming. This project presents an automated classification pipeline using public fundus images across four disease categories. Two novel hybrid models are proposed. LightEyeNetTransformer integrates a Swin Transformer backbone for global contextual processing, Atrous Spatial Pyramid Pooling for multi-scale lesion extraction, and Efficient Channel Attention for channel-wise feature refinement. SANO-Net is a Transformer-based architecture employing cross-scale gating and dynamic feature fusion to capture lesions of varying dimensions. Class imbalance is addressed through stratified train-validation-test splits and class-weighted loss functions. For interpretability, Grad-CAM++ generates visual heatmaps highlighting regions that influence model decisions. A web-based interface built with HTML, CSS, JavaScript, Python, and Flask includes modules for home, about, registration, login, prediction, and logout. The system provides classification outputs alongside explainability maps.

Keywords: Eye disease classification, fundus images, Swin Transformer, ASPP, ECA attention, SANO-Net, cross-scale gating, dynamic feature fusion, Grad-CAM++, explainable AI.