Artificial Intelligence in Leukemia Diagnosis: Applications, Datasets, and Barriers to Clinical Translation

Abstract

Leukemia diagnosis requires the integration of morphology, hematological parameters, immunophenotyping, cytogenetic or molecular findings, and clinical information. Artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), is increasingly being investigated as an adjunct to these diagnostic processes. This review examines recent applications of AI in leukemia diagnosis, with emphasis on microscopic image analysis, classical machine learning, deep learning, routine laboratory parameters, and the datasets that support model development. The literature demonstrates that AI systems can achieve high performance in selected benchmark tasks; however, reported performance varies substantially with disease subtype, dataset composition, and validation strategy. Major barriers to clinical translation include limited and imbalanced datasets, underrepresentation of rare cell types and populations, single-center development, dataset shift across institutions, incomplete coverage of contemporary leukemia classifications, and limited integration of multimodal diagnostic information. Recent evidence also indicates that most published systems remain unimodal, whereas clinical leukemia diagnosis combines morphology, immunophenotyping, molecular findings, and laboratory information. Accordingly, future research should prioritize diverse and clinically representative datasets, transparent reporting, independent external validation, multimodal integration, explainability, and prospective assessment of clinical utility. AI should currently be regarded as a decision-support technology that complements, rather than replaces, expert hematopathology and clinical judgment.