Neuro-Symbolic AI System for Logical Reasoning and Decision Making

Abstract: The ABO-Rh blood group system constitutes a critical biomarker in transfusion medicine, organ transplantation, and prenatal care. Conventional serological determination methods require venipuncture, trained personnel, and laboratory infrastructure, rendering them unsuitable for emergency triage, resource-limited settings, and continuous monitoring. This paper presents HemoVision, a novel deep learning framework that infers ABO-Rh blood group (eight classes: A+, A−, B+, B−, AB+, AB−, O+, O−) from non-invasive ocular surface imagery captured by a standard fundus camera. The biological hypothesis underlying HemoVision is grounded in established correlations between blood group antigens and micro-vascular morphological signatures in the conjunctival and retinal vasculature, including vessel tortuosity, branching angle distributions, fractal dimension of capillary networks, and perivascular pigmentation gradients. The proposed architecture employs a dual-branch convolutional neural network with cross- attention fusion: one branch processes full fundus images through a fine-tuned EfficientNet-V2-L backbone while the second branch operates on vessel-segmented maps extracted by a dedicated U-Net segmentation head. Cross-attention fusion integrates morphological and textural features before classification. Evaluated on the publicly augmented ORIGA- BG dataset (4,800 annotated ocular images across eight blood group classes), HemoVision achieves a mean classification accuracy of 92.6%, macro-average precision of 91.8%, recall of 92.1%, and F1-score of 91.9%, outperforming seven recent state-of-the-art baselines. The framework offers a pathway toward rapid, non-invasive blood group screening at the point of care.

Keywords: Blood group classification, ocular fundus analysis, deep learning, EfficientNet, cross-attention fusion, retinal vasculature, non-invasive diagnostics.