Systematic review and meta-analysis of AI methods for immature WBC subtype classification

Le résumé fourni par la source

Background: Automated morphology using computer vision and machine learning may support classification of immature leukocyte subtypes in acute myeloid leukemia (AML), but reported performance varies across studies and cell types. We systematically synthesized evidence on image-level classification of erythroblasts (EBO), monoblasts (MOB), myeloblast (MYO), and promyelocytes (PMO). Methods: We performed a systematic review and random-effects meta-analysis of diagnostic accuracy studies evaluating AI models for immature WBC subtype classification. For each subtype, one-vs-all 2 × 2 counts (TP, FP, FN, TN) were reconstructed and pooled sensitivity and specificity were estimated using random-effects models on the logit scale. Heterogeneity was quantified with I² and τ². Risk of bias and applicability were assessed using the QUADAS-AI framework. Results: Six studies were included. Pooled specificity was consistently high for all subtypes. Pooled sensitivity was high for MYO (98.7%), EBO (94.3%), MOB (89.1%), but substantially lower and more variable for PMO (65%), consistent with morphological overlap between adjacent maturation stages and conservative model behavior on borderline cells. Subgroup analysis by feature-extraction paradigm revealed a preliminary heterogeneity pattern: hand-crafted pipelines showed minimal heterogeneity (I² = 0% for EBO, MOB, and MYO), hybrid pipelines showed intermediate levels, and end-to-end deep learning pipelines exhibited the greatest heterogeneity (I² up to 99%). QUADAS-AI appraisal identified high risk of bias in dataset selection and index-test domains across all studies, limiting certainty and clinical generalizability. Conclusions: Current AI systems demonstrate consistently high specificity but more variable sensitivity for immature WBC classification in AML, with PMO remaining the most challenging subtype. All six studies are derived from a single benchmark dataset, limiting generalizability; independent prospective validation is required before clinical deployment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.