Explainable artificial intelligence for label-free red blood cell morphometrics using brightfield microscopy towards clinical applications

Abstract Background For diagnosing certain hematological disorders, variations in the red blood cell (RBC) morphology are considered as a marker. Advances in artificial intelligence (AI) offer the potential for automated, reproducible, label-free analysis of RBC morphology; however, most AI classifiers lack explainable, clinician-friendly outputs. Methods Single-cell RBC images of 17,781 across 10 morphological classes were acquired from three brightfield microscopes as a multicentric dataset. Mask R-CNN-based segmentation and 74 morphometric feature extractions were performed using the RedTell tool. Seven different machine learning classifiers were compared, and the best classifier was evaluated using SHapley Additive exPlanations (SHAP) analysis for interpretability and confidence calibration analysis for decision reliability. This led to the development of a comprehensive framework for evaluating the AI workflow through explainable AI in label-free RBC morphometric studies. Results The random forest classifier achieved an F1 score of 0.89. SHAP analysis indicated that intensity features strongly influenced the model predictions, followed by shape and texture features. These intensity-specific features were tested for PCA and were preserved across all three microscopes (0.49–0.88) with no meaningful device-driven confounding (silhouette ~ − 0.09). Expected calibration error was less (0.050), and the prediction confidence was good for clinically meaningful morphologies, supporting a practical reject-option strategy for flagging low-confidence predictions. Conclusion This fully explainable workflow, with feature-level interpretability and decision reliability, connects the AI and label-free RBC imaging domains, enabling trust and cost-effective adoption of label-free RBC morphology analysis, even in resource-limited clinical settings.