Comprehensive Multicenter Evaluation of an AI-Based Deep Learning System for Automated Blood Cell Classification and Morphologic Analysis.

Abstract

Background and objectives: Deep learning-based systems have shown promise in automating white blood cell (WBC) classification; however, most validations are limited to single institutions. This study aimed to evaluate the multicenter performance and robustness of a CNN-based deep learning system (DLS) for peripheral blood cell classification and morphologic abnormality detection across three hospitals.

Methods: A total of 766 peripheral blood smears (613 routine and 153 abnormal) from Juntendo University Hospital, Kyoto University Hospital, and the National Cancer Center Hospital were analyzed using the Sysmex DI-60. The DLS, trained on 1.49 million annotated cell images, classified 16 cell types and 42 morphological features. Cell classification performance was assessed by standard metrics, including accuracy and F1 score, while agreement with manual WBC counts was evaluated through correlation and regression analyses.

Results: The DLS achieved high specificity, negative predictive value, and accuracy (> 95%) across institutions. Mean F1 scores exceeded 0.80 for major WBC types in both routine and abnormal smears. Agreement with manual microscopy was strong for neutrophils, lymphocytes, and eosinophils (r² > 0.8; Cb > 0.9), while monocytes and basophils showed moderate variability. The DLS identified smears containing blasts or abnormal cells that required expert review and detected morphological features associated with specific diseases.

Conclusions: This multicenter validation demonstrates robust and reproducible performance of the DLS across laboratories, supporting its clinical applicability for routine WBC differentials and detection of disease-specific morphological abnormalities.