Development and Validation of a Novel Deep Learning-Based Model for Detection of Diabetic Kidney Disease from Retinal...

Area of Science:

• Ophthalmology
• Nephrology
• Artificial Intelligence in Medicine

Background:

• Retinal photographs offer a non-invasive method for early detection of systemic diseases in diabetic patients.
• Diabetes complications, such as Chronic Kidney Disease (CKD), can manifest subtle changes in the retina.
• Early detection of CKD in diabetic individuals is crucial for timely intervention and management.

Purpose of the Study:

• To develop and validate a novel deep learning (DL) model for detecting CKD in diabetic patients using retinal images.
• To differentiate between healthy controls, type 2 diabetes mellitus (T2DM) patients without CKD, and T2DM patients with stage 3 CKD.

Main Methods:

• An EfficientNet-B2 DL model was developed with a weighted cross-entropy loss function to handle class imbalance.
• A dataset of 225 participants was used, with a strict 80/20 patient-level split to prevent data leakage.
• Model performance was assessed using sensitivity, specificity, AUC, and Grad-CAM for interpretability.

Main Results:

• The EfficientNet-B2 model achieved a high overall AUC of 0.96, with specific AUCs of 0.99 (controls), 0.90 (T2DM), and 0.90 (T2DM with CKD stage 3).
• The model demonstrated strong performance with 82% sensitivity, 94% specificity, 81% precision, and an F1-score of 0.80.
• Grad-CAM visualizations indicated the model focused on the peripapillary and macular regions for predictions.

Conclusions:

• A three-class fundus-based DL model effectively differentiates controls, isolated T2DM, and T2DM with CKD stage 3.
• The model's performance suggests potential for screening and triage, but further external and prospective validation is required.