Foundation Model-Based Haemoglobin Concentration Estimation and Anaemia Screening from Retinal Fundus Images

Introduction The widespread screening for anaemia is currently limited by its reliance on invasive blood tests. This study's primary objective was to evaluate the feasibility and interpretability of Vision Transformer-based foundation models (FMs), including those pre-trained on natural and medical domains, for the noninvasive estimation of haemoglobin (Hb) levels using retinal fundus images. Methods This retrospective cross-sectional study included 34511 individuals from the Health Promotion Center of Seoul National University Hospital (2005–2016). We adapted five Vision Transformer (ViT)-based FMs (DINOv2, OpenCLIP, MAE, RetFound, VisionFM) for haemoglobin concentration regression. Low-Rank Adaptation (LoRA) and partial fine-tuning were employed for model adaptation. For explainability, we applied Grad-CAM saliency mapping and perturbation-based analysis using retinal vessel segmentation. Results Among all tested models, the natural-domain model DINOv2 combined with LoRA achieved the best performance (MAE = 0.703, R-squared = 0.682, accuracy = 0.957, AUROC = 0.917, sensitivity = 0.363, sensitivity for significant anemia = 0.806, negative predictive value = 0.968). Interpretation revealed that the model's predictions predominantly relied on the macula and peripapillary regions. The perturbation analysis confirmed that increased brightness and vividness of the retinal arteries correlated with higher estimated Hb levels. Conclusions Fine-tuned foundation models, particularly those pre-trained on the natural domain (DINOv2), showed high accuracy in estimating haemoglobin concentration noninvasively from retinal fundus images. The model's decision-making process was found to be clinically explainable and physiologically relevant, focusing on known anatomical and vascular indicators. These results suggest that appropriately adapted FMs can serve as powerful tools in clinical setting.