Clinical-guided deep learning framework for diabetic retinopathy: integrating lesion-aware attention, adversarial augmentation, and uncertainty quantification

Diabetic retinopathy (DR) is an escalating global health concern, driven by the increasing prevalence of diabetes. Existing automated screening systems often fall short in providing interpretable lesion-level analysis, robustness against class imbalance, and reliable uncertainty estimation, hindering their practical application in real-world clinical settings.

In response, researchers have developed a clinical-guided deep learning framework called CG-DRNet, which aims to improve reliable and explainable DR severity detection while enabling early-stage detection, particularly mild nonproliferative diabetic retinopathy (NPDR).

CG-DRNet incorporates lesion-aware attention, adversarial data augmentation, and Bayesian uncertainty measurement to tackle the challenges of DR diagnosis. By employing a multi-task deep learning framework and a lesion-aware attention network, CG-DRNet explicitly predicts microaneurysms, hemorrhages, exudates, and neovascularization.

To address the extreme imbalance of classes, a conditional generative adversarial network (CWGAN-GP) is utilized to generate clinically realistic images of minority classes in fundus images. Monte Carlo dropout is employed to model Bayesian uncertainty and estimate predictive confidence, while an uncertainty-informed semi-supervised learning strategy enhances data efficiency.

Evaluation of CG-DRNet was performed on publicly available fundus image datasets, including APTOS 2019, Messidor-2, and Clinical metadata. The framework achieved impressive results, reaching 93.8% accuracy on APTOS 2019 and 91.2% on Messidor-2, with only a 2.6% difference between the generalization and original model's performance.

The macro F1-score was 0.891, and the quadratic weighted kappa was 0.912. The referable DR detection AUC was 0.963, and the expected calibration error was only 0.034. CG-DRNet demonstrated an 84.7% sensitivity in detecting Grade 2+ DR, with an inference time of just 67 milliseconds, making it a viable solution for clinical use.

Diabetic retinopathy is a neurovascular complication resulting from processes such as AGE accumulation, PKC activation, and VEGF upregulation, which degrade the inner blood-retinal barrier. Typical appearances include microaneurysms, intraretinal hemorrhages, hard and soft exudates, and neovascular proliferation. Currently, approximately 90 million people worldwide have DR, with one-third likely to develop it by 2040.

DR is a leading cause of vision loss among working-age adults globally and carries an immense burden on healthcare systems, particularly in resource-constrained environments.

Early diagnosis of DR is crucial, as landmark DCCT and UKPDS trials have shown that maintaining HbA1c levels below 7% can decrease the risk of DR by up to 76% in type 1 diabetes and 25% in type 2 diabetes. However, global screening coverage remains low due to labor shortages, geographic differences, and inter-observer variability.

Manually grading DR is subjective, time-consuming, and inaccurate, with consensus among experts at a low 70-80% and throughput limited to 100-120 patients daily, far from the needs of large-scale screening. Deep learning, particularly Convolutional Neural Networks (CNNs), has proven effective in automated DR screening and grading, outperforming expert-level performance.