Abstract
An Attention Dual Transformer with Adaptive Temporal Convolutional (ADT-ATC) model is proposed in this research work for enhanced detection of Diabetic Retinopathy (DR) from retinal fundus images. Unlike traditional methods which evolved so far in DR analysis, the proposed model specifically processes the multi-scale spatial features through dual spatial transformer network and captures the temporal dependencies through adaptive temporal convolutional unit. The fine patterns like microaneurysms, and larger anatomical regions, including hemorrhages are focused on dual spatial transformer block which provides comprehensive and detailed analysis of spatial features. Additionally, a hierarchical cross attention module is included to fuse the spatial and temporal features which is essential to identify the DR. Experimentation of the proposed model using DRIVE and Diabetic Retinopathy datasets demonstrates the better performance of proposed ADTATC model with an accuracy of 98.2% on DRIVE and 97.7% on Diabetic Retinopathy datasets compared to conventional deep learning models.