DualPath-DRNet: A Self-Annotating Dual-Path Networks for End-To-End Diabetic Retinopathy Diagnosis

One of the challenges in detecting Diabetic Retinopathy (DR) is the detection of subtle early-stage microaneurysms. DualPath-DRNet is an end-to-end deep learning pipeline for 5-class DR prediction. The novelty of this method lies in combining lesion detection using YOLOv8, segmentation using Segment Anything Model (SAM), and dual-path classification using C and Swin Transformer in a single pipeline. The clinician's feedback is used to implement the self-annotation loop that iteratively improves the pseudo-labels. This reduces the dependency on manual annotation by 70%. The model was trained on the APTOS-2019 dataset, achieving 99.40% accuracy, and then validated on the EyePACS and IDRiD datasets, yielding 92% and 89% accuracy, respectively. A comparative study established the superiority of the model over the other state-of-the-art technologies, thus outperforming 26 existing DL models. The interpretability of the model is enhanced by the Grad-CAM heatmap. This research allows a high-precision DR screening that is scalable and requires minimal interactions.