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

Retinal photographs offer great opportunity to early detect systemic disorders related to diabetes, including Chronic Kidney Disease (CKD).

To develop and validate a novel deep learning model to detect CKD among diabetic patients.

We developed an EfficientNet-B2 Deep Learning (DL) model utilizing a weighted cross-entropy loss function to address class imbalance and distinguish retinal images among healthy controls, patients with isolated type 2 diabetes mellitus (T2DM), and patients with CKD stage 3 due to T2DM. The dataset was partitioned using a strict 80/20 patient-level split to evaluate bilateral eyes while strictly preventing data leakage. Model performance was evaluated using sensitivity, specificity, and area under the curve (AUC), alongside Grad-CAM visualizations for clinical interpretability.

The study included 225 participants. Among the evaluated DL architectures, the EfficientNet-B2 model demonstrated the best performance, achieving an overall AUC of 0.96. The model exhibited a sensitivity of 82%, specificity of 94%, precision of 81%, and an F1-score of 0.80. Class-specific AUCs were 0.99 for healthy controls, 0.90 for T2DM without CKD, and 0.90 for T2DM with CKD stage 3. Grad-CAM heatmaps indicated that the model primarily focused on the peripapillary and macular regions to make predictions.

This study presents a three-class fundus-based DL model, trained with a weighted-loss strategy, to differentiate controls, isolated T2DM, and T2DM with CKD stage 3. Further external and prospective validation is needed before it can be considered for screening/triage use.