GRIFD: a graded region-wise dissection and cross-pooling RNN framework for precise diabetic retinopathy detection in fundus images

Early and precise detection of diabetic retinopathy (DR) is essential in averting vision impairment. This work introduces graded region wise inspection and feature dissection (GRIFD), a novel deep learning architecture designed explicitly for fine-grained DR detection from fundus images. GRIFD enforces a graded region of interest (ROI) dissection approach, which methodically segments fundus images into diagnostic blocks to promote local lesion visibility. A recurrent neural network (RNN) is used to model sequential dependencies between these ROI blocks, aiming to mimic the progressive development of DR symptoms. We also introduce a cross-pooling refinement procedure to eliminate inter-region inconsistencies and improve feature continuity. Experiments on the Mendeley and EyePACS benchmarks show that GRIFD obtains 94.35% accuracy, 96.05% precision, and 94.91% specificity on Mendeley, all of which surpass baseline models such as convolutional neural network (CNN), long short-term memory (LSTM), Transformer, and some published DR detection models. The introduced framework presents a lightweight and robust method for DR screening and can potentially facilitate real-time clinical practice.