Performance Evaluation of an Artificial Intelligence-Assisted Digital Morphology System for Bone Marrow Aspirate Examination.

INTRODUCTION: Morphological examination of bone marrow aspirates (BMA) remains essential for the diagnosis of hematologic diseases, but conventional microscopy is labor-intensive and subject to interobserver variability. METHODS: A total of 158 samples from patients were analyzed. Differential counts obtained by optical microscopy were compared with those of the CellaVision BMA Application used according to its intended workflow of automated preclassification followed by expert review. The two methods were compared by Pearson correlation, Passing-Bablok regression, and Bland-Altman analysis, both overall and within the blast and plasma cells intervals that bracket the diagnostic thresholds. RESULTS: Pearson correlation coefficients ranged from 0.85 to 0.99 for all populations except basophils (r = 0.68). Passing-Bablok regression showed no significant systematic or proportional bias for erythroblasts, granulocytic cells, eosinophils, and monocytes; a small systematic difference was observed for blasts, a small proportional difference for lymphocytes, and significant deviations from linearity for plasma cells and basophils. Bland-Altman mean differences ranged from -1.4% to 1.2% across all populations. At the clinically relevant decision thresholds for blasts and plasma cells, overall agreement between the two methods ranged from 96.8% to 99.4%, although sensitivity at the 20% blast threshold was 77.3%. CONCLUSION: Used according to its intended workflow, the CellaVision BMA Application produced differential counts comparable with conventional microscopy for most nucleated cells, but agreement was weaker for basophils and plasma cells. Expert review and validation remain necessary. These findings, obtained in a single-center series, provide a basis for further evaluation of this and similar systems.