Adaptive Contrastive Learning in Vision Transformer for Non-Invasive Hemoglobin Estimation From Conjunctival Images

Fuentes-Beingolea J.H.; Palomino-Quispe F. · IEEE Access (2025) · DOI: 10.1109/access.2025.3638640

Resumen

The invasive nature and logistical challenges of standard blood tests impede the early detection of anemia, particularly in resource-limited settings. To address this, we introduce Hemo-ConViT, a novel deep learning framework for non-invasive hemoglobin estimation using conjunctival images. Our framework is built upon a data-efficient Vision Transformer (DeiT) and incorporates a purpose-built adaptive contrastive loss function. This unique loss guides the model to learn a feature space sensitive to the continuous nature of hemoglobin (Hgb) by dynamically weighting sample pairs based on their clinical similarity, thereby enhancing differentiation between borderline cases. Evaluated on a public dataset (N=218), Hemo-ConViT demonstrates robust regression performance, achieving a Mean Absolute Error (MAE) of 0.870 g/dL and a coefficient of determination (R2) of 0.761. In the parallel classification task (anemia threshold: Hgb < 12.0 g/dL), the model achieves an overall accuracy of 95.0% (95% CI). Interpretability analysis via Attention Rollout confirms that the model’s predictions are based on physiologically relevant vascular tissue in the conjunctiva, building trust in its clinical application. The model’s architecture (22M parameters) is designed for efficiency, enabling potential deployment on mobile and edge devices for real-time analysis. These findings establish Hemo-ConViT as a promising and interpretable proof-of-concept, laying the groundwork for developing accessible, low-cost, and portable point-of-care (POC) screening systems that leverage ubiquitous sensor technology to address a critical global health challenge.

Instituciones: Universidad Nacional de San Antonio Abad del Cusco