The Use of Explainable Deep Learning Models for Diabetic Retinopathy Management: From Screening to Severity Grading

Abstract

Background: Diabetic Retinopathy (DR), a leading cause of blindness, is rising globally due to increasing diabetes prevalence, especially in low-resource settings lacking diagnostic tools and specialists. Early DR is often asymptomatic; timely and scalable screening is essential. This study introduces a cost-effective two-stage deep learning framework for DR detection: Stage 1 performs binary classification on Optical Coherence Tomography (OCT) images, while Stage 2 grades DR severity use multi-class classification. Shapley Additive exPlanations (SHAP) enhance model transparency and clinical trust. Trained on public datasets and validated by a CHUK ophthalmologist, the approach provides a cost-effective solution for DR management in the underserved diabetic population.

Methods: The study followed five phases: (1) image dataset preparation using a four-step preprocessing pipeline: blurring, denoising, augmentation, and vector transformation for CNN input; (2) training and evaluating five pre-trained CNNs (MobileNetV3, DenseNet121, InceptionV3, Xception, VGG19) with standard metrics; (3) applying SHAP to interpret predictions of a multi-class model; (4) validating performance on 270 CHUK specialist-annotated images and 270 from the Brazil dataset; and (5) deploying the top models in a web application.

Results: The proposed model achieved strong performance in both DR screening and severity grading. DenseNet121 performed best, reaching 97% accuracy for referable DR detection and class accuracy of 97% (healthy), 92% (mild), 89% (moderate), 94% (severe), and 92% (proliferative). Validation datasets yielded Cohen’s Kappa scores of 0.771 and 0.680, demonstrating substantial agreement with expert grading. SHAP explanations enhanced interpretability across DR stages.

Conclusion: The applied preprocessing techniques improved image clarity and boosted model accuracy in grading DR severity. The proposed CNN framework enables end-to-end DR management. Combining expert validation, quality imaging, and interpretable deep learning offers a promising solution for early detection of DR and vision preservation in diabetic populations.