Enhanced Diabetic Retinopathy Grading Using a Deep Hybrid Approach

In recent times there have been significant achievements in deep learning methodologies, the task of classifying retinal diseases is still challenging because of their minute pathological regions. The retina’s blood vessels can be compromised by diabetic retinopathy, a complication arising from diabetes falls under the distinct category of retinal diseases. Despite many attempts with classification models such as CNN and other deep learning approaches detection encounters several challenges. These include existing deep learning models are susceptible to biases towards either local or global perspectives. These biases can impact semantic-level understanding of these representations. This work proposes a hybrid framework for Diabetic retinopathy grading. Inspired from the achievements of vision transformers in image classification and its ability to capture spatial relationships in fundus images related to Diabetic retinopathy features. Augmenting this vision transformer with traditional multiple instance learning forming a hybrid model ViT-MIL. By combining ViT’s global learning capabilities with MIL’s effective local feature extraction this hybrid model achieves enhanced performance.