An Explainable AI and Optimized Multi-Branch Convolutional Neural Network Model for Eye Anemia Diagnosis

Abstract

Anemia stands as one of the foremost global health challenges, with an estimated 1.62 billion individuals worldwide carrying insufficient haemoglobin levels. Standard diagnostic workflows depend on venous blood sampling and laboratory analysis, creating substantial barriers for communities lacking healthcare infrastructure. This study presents a vision-based, fully non-invasive detection framework that examines conjunctival pallor captured in ordinary eye photographs to identify anemia and generate haemoglobin estimates. A pre-trained MobileNetV2 network is systematically adapted for binary classification of conjunctival pallor through a structured two-phase fine-tuning process. The system further incorporates a severity-stratification module that converts raw classifier output into five clinically meaningful tiers, each paired with an estimated haemoglobin range based on WHO guidelines. Results are delivered through a Gradio web interface supporting both uploaded photos and live webcam streams. Benchmarked on 400 test images drawn from a public Kaggle conjunctival dataset, the framework attains a classification accuracy of 95.4%, AUC of 0.981, sensitivity of 95.0%, specificity of 95.5%, and a haemoglobin MAE of 0.92 g/dL, surpassing all evaluated baseline methods. Grad-CAM explainability heatmaps validate that the network directs its attention to the conjunctival region, confirming clinical interpretability.