Cell-level Image patches of Blood Smear

Iron deficiency anemia (IDA) and thalassemia (THL) are common hematological disorders that require efficient and accurate screening for early diagnosis. Traditional blood smear analysis is both labor-intensive and prone to subjectivity, highlighting the need for AI-driven methods to improve diagnostic sensitivity and specificity.This study introduced a novel hybrid AI-based patient-level classification framework that integrates soft voting with optimized probability-based thresholds. The model was trained and validated using a real-world dataset from Hatyai Hospital, Thailand, and evaluated at both the patch and patient levels. The proposed approach achieved 96% accuracy on the test set, with precision-recall values of 1.00 and 0.83 for IDA and 0.95 and 1.00 for THL, respectively. THL sensitivity reached 1.00 at the patient level, while IDA sensitivity was 0.83. Bayesian probability updates confirmed prediction reliability, yielding post-test probabilities exceeding 99.99% for IDA and 80\% for THL. The model explained 62.84% of the variance in patient classifications, demonstrating strong discriminatory power.Model interpretability was assessed using SHAP and Grad-CAM, which highlighted key red blood cell morphological features. The proposed framework serves as a cost-effective and clinically interpretable AI-assisted hematology screening tool, supporting decision-making in resource-limited settings.Limitations include the use of a single-center dataset and the need for adaptive threshold optimization. Future work will focus on multi-center validation and real-world clinical integration. This study establishes a structured baseline for AI-assisted hematology screening, supporting early detection and improved clinical decision-making.