Deep learning approaches for the diagnosis and classification of leukemia using bone marrow aspirate smear images

Background: Leukemia is a group of hematologic malignancies characterized by uncontrolled proliferation of abnormal white blood cells, classified into acute and chronic subtypes with distinct clinical and pathological features. Early diagnosis is challenging due to nonspecific symptoms, and while peripheral blood smears are frequently used for initial evaluation, bone marrow aspirate smears remain the gold standard for definitive diagnosis. They provide essential morphological insights into leukemic cell populations. However, manual examination of bone marrow aspirate smears is labor-intensive, subject to inter-observer variability, and prone to fatigue, which may compromise diagnostic accuracy. To address these limitations, recent studies have explored the integration of artificial intelligence, particularly deep learning algorithms, to automate the detection and classification of leukemic cells. This review highlights the potential of deep learning approaches to improve diagnostic accuracy and efficiency using bone marrow aspirate smears, and also to identify the existing challenges in order to address them in future research.

Methods: This review was conducted through a comprehensive search of databases including PubMed, Science Direct and Google Scholar to identify studies published between 2014 and 2024. Keywords included “bone marrow aspirate smear,” “leukemia,” “classification,” “diagnosis,” and “deep learning.”. filters were applied to retrieve only free full-text articles, and searches were limited to titles and abstracts. Retrieved articles were screened for relevance, and 12 studies were selected. These articles specifically examined deep learning methods applied to bone marrow aspirate smear image analysis for leukemia diagnosis and classification,

Results: Deep learning models demonstrated high accuracy in classifying leukemia from bone marrow aspirate smears. CNN-based frameworks achieved AUROCs up to 0.97 and accuracies above 91% for distinguishing AML from healthy samples. ResNet-50 and ensemble approaches reached accuracies exceeding 92% and 99% for multi-class and early ALL detection. Novel architectures such as AMLnet and MILLIE matched or outperformed expert hematologists, achieving AUROCs of 0.885–0.96 for subtype classification and robust recognition of lymphoblasts and promyelocytes. Faster R-CNN variants with rank-sort optimization improved detection efficiency, reducing inference time to 0.3s per image. These findings highlight the robustness, scalability, and clinical potential of deep learning approaches in augmenting conventional bone marrow aspirate smear analysis for leukemia diagnosis and classification.

Conclusion: Deep learning has emerged as a powerful tool for the cytomorphological evaluation of bone marrow aspirate smear images, achieving performance comparable to, and in some cases exceeding of expert hematologists. Models such as CNNs, ResNet variants, and novel architectures like AMLnet, BMSNet, and MILLIE demonstrated robust capabilities in detecting leukemic subtypes, predicting mutations, and assisting in rare entity recognition such as APL. These approaches reduce diagnostic delays and offer promise in both high-resource and low-resource clinical settings. Despite these advances, challenges remain, including limited sample sizes, variability in image quality, and the need for multicentre datasets to improve generalizability. Integrating cytomorphology with complementary modalities such as flow cytometry, cytogenetics, and molecular profiling is a crucial next step. Overall, deep learning applied to bone marrow aspirate smear images holds significant potential to enhance diagnostic accuracy, accelerate clinical workflows, and strengthen decision support in hematologic malignancies.

Dr. Rezvan Rahimi is an Assistant Professor of Medical Informatics at Tarbiat Modares University, Iran. She specializes in applying Artificial Intelligence (AI) and machine learning to enhance cancer care. Her PhD research developed an AI-powered clinical decision support system to reduce chemotherapy prescription errors in pediatric leukemia. Dr. Rahimi currently leads projects on AI-driven diagnostic tools, including CNN models for leukemia detection and classification from medical images. She is also working on AI applications for preeclampsia diagnosis and urine sediment analysis. Her work focuses on leveraging intelligent systems to improve clinical outcomes in oncology and healthcare.