Multi-Headed CNN and Vision Transformer-Based Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a medical condition that can occur as a complication of diabetes and is characterised by blood leakage into the retinal tissues. The condition initially manifests as no symptoms or minor vision problems, which can progress to severe vision loss. DR is classified into five stages: Normal, Mild, Moderate, Severe, and PDR (Proliferative Diabetic Retinopathy). Detecting DR early is crucial to avoid severe complications, including complete blindness. The presence of certain characteristics, including soft exudates, hard exudates, hemorrhages, and microaneurysms in the affected eye, may indicate the presence of diabetic retinopathy. Early detection and prevention measures can control DR and prevent further damage to the retina. Diagnosis of diabetic retinopathy involves monitoring and assessment of fundus images. However, manual detection of DR is a time consuming process and is highly dependent on the expertise of ophthalmologists. It is also challenging for ophthalmologists to detect DR in its early stages from fundus images.Therefore, there is a need for an automated diagnosis system capable of extracting features from fundus images and classifying DR into different stages. In this regard, we propose a multi-headed CNN and vision transformer based model capable of accurately detecting and classifying all early stages of diabetic retinopathy as early as possible. Our proposed model leverages the CNN’s ability to extract local spatial features and the vision transformer’s ability to extract global contextual information. By combining these local and global features, a rich feature representation is created, which is then used to detect and classify all stages of DR.