Comprehensive performance assessment of the BMIA-12 a system for bone marrow cell quantification in normal and hematological malignancy samples.

Manual bone marrow (BM) differential counting is labor-intensive, time-consuming, and prone to inter-observer variability. Artificial intelligence (AI)-based systems can standardize BM cytological assessments. This study evaluated the BMIA-12 A system (UIMD, Seoul, Korea) for automated BM cell recognition and differential counting. A total of 298 BM aspirate smears from 149 patients were analyzed, including normal controls (n = 50), multiple myeloma (n = 33), monoclonal gammopathy of undetermined significance (n = 6), acute myeloid leukemia (AML; n = 40), acute promyelocytic leukemia (n = 4), and acute lymphoblastic leukemia (ALL; n = 16). Three classification methods were compared: AI-automated, expert-reviewed AI, and manual microscopic counting. Both wedge and squash preparations were assessed. System performance was evaluated using recall, precision, F1-score, and accuracy. BMIA-12 A achieved accuracies of 94.6% (wedge) and 94.0% (squash), with recall > 90% for 14/16 cell types. Wedge preparations showed superior precision for key diagnostic cells, including plasma cells, blasts, and basophils. Strong correlations (r ≥ 0.9) were observed between AI-automated and expert-reviewed classifications for nine cell types. However, disease-specific quantification varied significantly by method, particularly for plasma cell and blast percentages. Inter-method discrepancies were pronounced in AML with NPM1 mutation and B-ALL with BCR::ABL1 fusion. Overall, BMIA-12 A provides robust classification for normal BM samples. Persistent inter-method differences highlight the need for further validation.