ViT-MAD: A novel generative-AI based architecture for diabetic retinopathy detection

Retinal imaging, like fundus photographs, is generally used to analyze eye conditions, where professionals like ophthalmologists review the images to identify signs of nerve damage or leakage, helping determine the severity of the disease. However, manual evaluation of such images can lead to errors due to human factors like fatigue and work stress. With advancements in technology, especially in artificial intelligence (AI), it is possible to reduce human-driven errors. AI-based solutions allow professionals to re-evaluate their analyses and facilitate early and accurate predictions. Computer vision-based AI can detect minute signs of diabetic retinopathy (DR) that humans may miss, thereby improving overall accuracy and guiding the treatment in the right direction. Extensive research has been conducted, and various AI-based solutions have been developed to detect DR using various retinal imaging modalities. Since obtaining data from unhealthy individuals is challenging and expensive, an anomaly detection approach using healthy fundus images has been explored. To address these challenges, proposing a novel solution ViT-MAD (Vision Transformer Masked Anomaly Detector), a Vision Transformer-based method that leverages a similarity score model to accurately detect diabetic retinopathy in fundus images. ViT-MAD has achieved over 94% accuracy across diverse fundus image datasets.