Automated Diabetic Screening via Anterior Segment Ocular Imaging: A Deep Learning and Explainable AI Approach

Diabetic retinopathy screening traditionally relies on fundus photography, requiring specialized equipment and expertise often unavailable in primary care and resource limited settings. We developed and validated a deep learning (DL) system for automated diabetic classification using anterior segment ocular imaging a readily accessible alternative utilizing standard photography equipment. The system leverages visible biomarkers in the iris, sclera, and conjunctiva that correlate with systemic diabetic status. We systematically evaluated five contemporary architectures (EfficientNet-V2-S with self-supervised learning (SSL), Vision Transformer, Swin Transformer, ConvNeXt-Base, and ResNet-50) on 2,640 clinically annotated anterior segment images spanning Normal, Controlled Diabetic, and Uncontrolled Diabetic categories. A tailored preprocessing pipeline combining specular reflection mitigation and contrast limited adaptive histogram equalization (CLAHE) was implemented to enhance subtle vascular and textural patterns critical for classification. SSL using SimCLR on domain specific ocular images substantially improved model performance.EfficientNet-V2-S with SSL achieved optimal performance with an F1-score of 98.21%, precision of 97.90%, and recall of 98.55% a substantial improvement over ImageNet only initialization (94.63% F1). Notably, the model attained near perfect precision (100%) for Normal classification, critical for minimizing unnecessary clinical referrals.

💡 Research Summary

This paper presents a novel approach to diabetes screening that departs from the conventional reliance on retinal fundus photography and instead exploits readily obtainable anterior segment ocular images (iris, sclera, and conjunctiva) captured with standard cameras. Recognizing that systemic diabetic manifestations extend to the anterior eye—e.g., increased scleral vessel tortuosity, conjunctival micro‑angiopathy, and iris structural alterations—the authors assembled a dataset of 2,640 high‑resolution anterior segment photographs from three clinical groups: Normal, Controlled Diabetes, and Uncontrolled Diabetes.

A major obstacle uncovered early in the study was a high rate of labeling errors (≈ 65 % of images). To address this, the team devised an automated quality‑assessment pipeline that quantifies three clinically relevant biomarkers: redness (via HSV analysis), vessel density (via CLAHE‑enhanced Canny edge detection), and scleral whiteness (via LAB L‑channel analysis). By combining these into an inflammation score and applying K‑means clustering, they automatically re‑assigned mislabeled images, improving the baseline F1‑score from 0.31 to 0.8862 and establishing statistically significant separation among the three groups (ANOVA, p < 0.001).

Pre‑processing consists of two stages: (1) specular reflection removal using intensity thresholding, morphological dilation, and Telea in‑painting; (2) contrast‑limited adaptive histogram equalization (CLAHE) applied only to the L‑channel of LAB color space, preserving true color while enhancing subtle vascular and textural cues. Standard augmentation (rotations, flips, color jitter, random crops) further bolsters model robustness.

The core technical contribution is the integration of self‑supervised learning (SSL) via SimCLR on a large pool of unlabeled anterior segment images (≈ 5 k). This SSL phase learns domain‑specific representations of ocular micro‑structures without any expert annotation. The pretrained encoder is then fine‑tuned on the cleaned, labeled dataset. Five contemporary architectures were benchmarked: EfficientNet‑V2‑S, Vision Transformer (ViT‑B/16), Swin‑Transformer‑V2‑Base, ConvNeXt‑Base, and ResNet‑50.

EfficientNet‑V2‑S with SSL achieved the best performance: F1‑score = 98.21 %, precision = 97.90 %, recall = 98.55 %. By contrast, the same architecture initialized only on ImageNet reached an F1‑score of 94.63 %, demonstrating a 3.58 % absolute gain from domain‑specific SSL. Notably, the model attained 100 % precision for the Normal class, minimizing false‑positive referrals. EfficientNet‑V2‑S uses 20.8 M parameters—about four times fewer than the transformer models (86–88 M)—yet outperforms them (93.80 % F1 for ViT/Swin), highlighting its suitability for resource‑constrained deployments.

Explainable AI (XAI) was addressed through Gradient‑Weighted Class Activation Mapping (Grad‑CAM). Heatmaps consistently highlighted medically interpretable regions: scleral vessel morphology, iris structural patterns, and conjunctival hyperemia. Quantitative regional attention analysis (iris vs. sclera vs. peripheral) showed statistically significant differences across disease categories (Kruskal‑Wallis, p < 0.001; Dunn’s post‑hoc). Independent review by three ophthalmologists of 50 randomly selected cases yielded a 94 % agreement with the model’s highlighted regions, confirming that predictions are grounded in genuine pathology rather than spurious artifacts.

The authors also contributed an automated data‑quality assessment pipeline that identified and corrected labeling errors, a systematic architecture benchmark, and a clinically validated XAI framework. Together, these innovations demonstrate that anterior segment imaging combined with SSL can serve as a practical, non‑invasive diabetes screening modality, dramatically lowering equipment costs and expanding access in primary‑care and low‑resource settings.

In conclusion, the study provides compelling evidence that (1) anterior eye structures contain discriminative biomarkers for systemic diabetes, (2) self‑supervised pretraining on domain‑specific unlabeled data markedly improves downstream classification, (3) efficient CNNs can outperform larger transformers while remaining computationally lightweight, and (4) explainability techniques can verify that AI decisions align with clinical knowledge. Future work should focus on multi‑center external validation, integration into mobile or edge devices, and correlation of model outputs with quantitative measures such as HbA1c levels, paving the way for widespread clinical adoption.