Clinical-guided deep learning framework for diabetic retinopathy: integrating lesion-aware attention, adversarial...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) prevalence is rising globally with diabetes, creating a significant public health concern.
• Current automated DR screening lacks interpretability, robustness to class imbalance, and reliable uncertainty estimation, hindering clinical application.
• Early detection of DR, especially mild nonproliferative diabetic retinopathy (NPDR), is crucial for preventing vision loss.

Purpose of the Study:

• To introduce a clinical-guided deep learning framework (CG-DRNet) for reliable and explainable DR severity detection.
• To improve early-stage DR detection and mimic clinical diagnostic procedures.
• To address limitations of current automated systems, including class imbalance and lack of uncertainty quantification.

Main Methods:

• A multi-task deep learning framework with a lesion-aware attention network to explicitly identify DR lesions (microaneurysms, hemorrhages, exudates, neovascularization).
• A conditional generative adversarial network (CWGAN-GP) for generating realistic minority class fundus images to combat class imbalance.
• Bayesian uncertainty modeling using Monte Carlo dropout and an uncertainty-informed semi-supervised learning strategy for enhanced data efficiency.

Main Results:

• CG-DRNet achieved high accuracy (93.8% on APTOS 2019, 91.2% on Messidor-2) with minimal generalization difference.
• The framework demonstrated strong performance with a macro F1-score of 0.891 and quadratic weighted kappa of 0.912.
• Achieved high AUC (0.963) for referable DR detection and low expected calibration error (0.034), with 84.7% sensitivity for Grade 2+ DR at 67 ms inference time.

Conclusions:

• The proposed CG-DRNet framework offers a viable solution for reliable, explainable, and efficient DR severity detection in clinical settings.
• The integration of lesion-conscious attention, adversarial data augmentation, and Bayesian uncertainty measurement significantly enhances diagnostic performance.
• The framework's ability to detect early-stage DR and its efficiency suggest potential for widespread clinical adoption and improved patient care.