AETUnet: Enhancing Retinal Segmentation With Parameter-Efficient UNet Architecture and Lightweight Attention Mechanism

Diabetic retinopathy (DR) is a leading cause of vision loss in working-age adults, making early and accurate detection crucial. This paper presents AETUnet (Attention Enhanced Transformer UNet), a new lightweight architecture to improve on-response retinal lesion segmentation. By combining expanded convolutions with a lightweight attension mechanism, AETUnet improved segmentation accuracy while remaining computationally efficient. Evaluated on the DRIVE and IDRiD datasets, AETUnet has outperformed existing models like Unet++ and ResUnet. On the DRIVE dataset, it achieved a mean Intersection over Union (MIoU) of 86.5, a Dice Score of 0.884, and an Accuracy of 97.64%. For haemorrhage segmentation on the IDRiD dataset, AETUnet recorded an MIoU of 89.33, a Dice Score of 0.882, and Accuracy of 98.46%, showcasing strong performance across all lesion types. AETUnet, with its low number of model parameters and excellent results is clinically practical for real-time applications in resource-constrained settings and provides important foundational steps towards more wide-ranging imaging tasks in other domains.