Deep Learning Based Estimation of Blood Glucose Levels from Multidirectional Scleral Blood Vessel Imaging

A team of researchers, including Muhammad Ahmed Khan and Saif Ur Rehman Khan, has published a groundbreaking paper on arXiv detailing ScleraGluNet. This is a novel deep-learning framework designed to estimate blood glucose levels non-invasively by analyzing the microvasculature in the sclera—the white part of the eye. The system works by capturing five images of a patient's eye from different gaze directions. After applying vascular enhancement techniques, the model uses parallel convolutional neural network (CNN) branches to extract features, refines them with a bio-inspired Manta Ray Foraging Optimization (MRFO) algorithm, and fuses the multi-angle data using a transformer-based cross-view attention mechanism. This sophisticated architecture allows it to detect diabetes-related alterations in the readily visible superficial blood vessels.

The model was trained and validated on a dataset of 2,225 anterior-segment images from 445 participants, categorized as having normal glucose, controlled diabetes, or high-glucose diabetes. Using a strict subject-wise five-fold cross-validation to prevent data leakage, ScleraGluNet demonstrated impressive performance. It achieved a 93.8% overall accuracy for metabolic status classification, with high one-vs-rest AUC scores (Area Under the Curve) ranging from 0.956 to 0.982. For the continuous task of estimating fasting plasma glucose (FPG), the model's mean absolute error (MAE) was 6.42 mg/dL, with a strong correlation (r = 0.983) to standard laboratory blood measurements. Bland-Altman analysis showed minimal mean bias (+1.45 mg/dL). The authors conclude that this multiview scleral imaging approach is a highly promising non-invasive tool for glycemic assessment, though they note the necessity for future multicenter clinical trials before potential real-world deployment.