A Novel Transformer Model with Multiple Instances Learning for Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a leading cause of vision impairment among diabetic patients, and itstimely detection is crucial for preventing irreversible vision loss. This work presents a novel hybrid model thatcombines Convolutional Neural Networks (CNN), EfficientNetB0, Transformer architectures, and MultipleInstance Learning (MIL) for accurate DR classification and recommendation. EfficientNetB0 serves as alightweight yet powerful backbone for extracting high-quality features from retinal fundus images, while theTransformer module captures global context and spatial relationships across image regions. MIL enhances modelrobustness by treating each image as a collection of instances, allowing effective learning even with limited orweak labels. Additionally, the system provides severity-based recommendations to assist clinicians in prioritizingpatient care. Experimental results across multiple DR datasets demonstrate superior accuracy, generalization,and clinical relevance, highlighting the model’s potential for integration into real-world diabetic screeningworkflows.