Artificial Intelligence Classification Algorithms for Leukaemia Diagnosis: A Structured Review of Model Performance and Clinical Readiness

Abstract

Leukaemia diagnosis and classification increasingly combine morphology with immunophenotypic, cytogenetic and molecular information, while artificial intelligence (AI) offers a potential means of standardising image-based assessment and reducing repetitive workload. This structured review critically examines contemporary AI approaches for leukaemia cell and case classification, with emphasis on dataset scale, unit of analysis, validation design, reported performance and clinical readiness. Peer-reviewed literature was identified through a targeted PubMed-focused search updated on 30 July 2026, with publisher and DOI records used to verify study design, numerical results and bibliographic identity. Representative primary human studies were selected only when the reported task, unit of analysis, testing design and principal performance metrics could be verified from the primary report. The evidence shows substantial technical progress, but the interpretation of apparently high performance changes markedly with the level at which generalisation is tested. Image-wise or closely related internal partitions provide weaker evidence of transportability than patient-level internal testing, while genuinely independent institutional or imaging-platform evaluation provides the most clinically informative test of robustness. Across acute lymphoblastic leukaemia, acute myeloid leukaemia and acute promyelocytic leukaemia applications, patient-level separation, external testing, class-balance reporting, transparent reference standards and multimodal diagnostic integration remain decisive. AI should therefore be considered an adjunct to expert haematopathology rather than a stand-alone substitute for contemporary integrated diagnosis. Multicentre prospective evaluation, rigorous reporting and clinically meaningful human-AI workflow studies are priorities for translation.