Evaluation of a Semi‐Supervised AI Model ( ASUS Blade) for Peripheral Blood Film Leukocyte Classification
Bingwen Eugene Fan, David Tao Yi Chen, Chiew Yan Lee, Kian Guan Eric Lim, Yi Xiong Ong, Wei Yong Kevin Wong, Shu-Yu Hsu, Cathy Chang, Pik Wan Erica Chiang, Siti Thuraiya Binte Abdul Latiff, Shu Ping Lim, Christina Lai Lin Sum, Sanchalika Acharyya, Moh Sim Wong, Hemalatha Shanmugam, Ponnudurai Kuperan, STEFAN S. WINKLER · International Journal of Laboratory Hematology · 2025
BACKGROUND: The peripheral blood film (PBF) analysis traditionally relies on manual microscopy (MM), a labour-intensive method with inter-observer variability. This study evaluates Blade (a semi-supervised AI model) and CellaVision DM9600 (commercial benchmark) against MM in automated leukocyte classification. METHODS: PBFs from 168 patients were prepared using automated staining and scanned digitally. Blade, trained on 185 412 cells (75 435 labelled, 109 977 unlabelled) via ResNet34 and RetinaNet architectures, underwent pseudo-labelling and AdamW optimisation. Performance was evaluated on 1675 cells against MM using the concordance correlation coefficient (CCC), Bland-Altman analysis, Deming/Passing-Bablok regression and diagnostic accuracy measures across nine leukocyte subtypes. RESULTS: When evaluated individually against MM, both systems showed high agreement. Blade achieved excellent correlation for common cells (neutrophils: ccc = 0.988; lymphocytes: ccc = 0.985; eosinophil: ccc = 0.953) and comparable results to CellaVision for monocytes (ccc = 0.852 vs. 0.847) and basophils (ccc = 0.762 vs. 0.794). Blade performed better for metamyelocytes (ccc = 0.905 vs. 0.756) and showed higher sensitivity for monocytes (75% vs. 63%) and myelocytes (87% vs. 74%). Regression analysis showed slopes close to 1.0 for most cell types, with Blade displaying narrower Limits of Agreement in Bland-Altman analysis. Both systems achieved 100% sensitivity for blasts and reactive lymphocytes. Overall macro-averaged performance was comparable between Blade (sensitivity 89.2%, specificity 96.3%) and CellaVision (86.3% and 96.7%). CONCLUSION: Blade and CellaVision demonstrated strong concordance with MM, validating their clinical utility. Blade's semi-supervised learning confers marginal advantages in rare cell detection and stability, highlighting AI's potential to enhance diagnostic accuracy. While both systems reduce labour and variability, Blade's performance has potential for integration into haematology workflows. Future validation in diverse cohorts is recommended.