Towards noninvasive blood count using a deep learning pipeline from bulbar conjunctiva videos.

Abstract

Invasive venous blood draws remain the clinical standard for hematology, yet they are invasive, time-consuming, and costly. We introduce Video-to-Vessels, a computer-vision pipeline that converts high-magnification videos of bulbar conjunctiva capillaries into low-dimensional spatiotemporal vessel representations, reducing video dimensionality by ~200-fold while preserving hemodynamic information. These representations feed VesselNet, a multi-instance regression network that encodes each vessel with a modified ConvNeXt backbone, fuses vessel-specific thickness via cross-attention, and predicts blood biomarkers from concatenated embeddings. On a cohort of 224 participants with paired laboratory counts, VesselNet achieves a hemoglobin-based anemia ROC-AUC of 82.8% and a Spearman's ρ of 0.47, while attaining a ρ of 0.46 for red-blood-cell (RBC) count regression. Removing local stabilization and segmentation-denoising lowers ρ by 38% for hemoglobin and 19% for RBC, underscoring their contributions. Our results mark a step toward a fully noninvasive complete blood count, coupling representation learning with ocular imaging.