Authors: Mr. S. Kaushik Raj, Mrs. B. Shyamala Devi
Abstract: Diabetic retinopathy (DR) is one of the major causes of vision loss worldwide, making early and reliable detection extremely important. This work presents an advanced transformer-driven framework combined with a Multiple Instance Learning (MIL) strategy to classify DR using retinal fundus images. The transformer model effectively learns long-range relationships and contextual patterns, while the MIL approach analyzes image patches to highlight clinically significant areas. Together, this hybrid system delivers strong feature representation and stable classification performance, even with variations in image quality and resolution. Trained on a large and diverse dataset, the proposed model achieves higher sensitivity and specificity than many existing deep learning techniques. The system is designed to assist eye-care professionals by enabling accurate, timely assessments and providing a scalable solution suitable for extensive DR screening initiatives.
