Background: Diagnosing leukemia from bone marrow biopsies or aspiration smears is expert-dependent and time-consuming. The rise of machine learning(ML) and deep learning(DL) offers a promising avenue for automating image analysis, though approaches across the literature remain varied. Objective: This systematic review examines the use of ML models for leukemia classification using bone marrow images, summarizing recent methods, model types, validation strategies, and performance while highlighting methodological gaps. Methods: PubMed, Scopus, and Web of Science databases were systematically searched from Jan 1, 2015 up to August 8, 2024 for current peer-reviewed studies applying AI to bone marrow biopsy/aspiration images for leukemia diagnosis. the studies were selected in two steps, title/abstract and full-text screening. Extracted data included dataset characteristics, model architecture, preprocessing, validation, explainability tools, and performance metrics. Results: Thirty-three studies met the inclusion criteria. A diverse range of ML techniques was identified, from classical algorithms using handcrafted features to modern DL frameworks including CNNs(Convolutional Neural Networks) and ensemble models. DL-based approaches consistently outperformed traditional ML models, achieving high classification accuracy in leukemia subtype differentiation. However, most studies lacked external validation and explainable AI integration, limiting clinical applicability. Conclusions: ML—particularly DL—demonstrates significant promise in automating leukemia diagnosis using bone marrow images. Despite encouraging performance, challenges persist around standardization, generalizability, and transparency. Future research should prioritize robust external validation, explainability, and standardized datasets to support clinical adoption and regulatory approval.
