What question did this study set out to answer?

This review analyzes AI approaches for evaluating blood and bone marrow samples to identify blasts effectively.

June 21, 2026

Next Generation Digital Morphology: Blast Preclassification in Bone Marrow Aspirates

Key Points

This review analyzes AI approaches for evaluating blood and bone marrow samples to identify blasts effectively.
Performed a narrative review of literature on AI-based morphologic evaluation in hematology.
Analyzed classical machine learning and deep learning methodologies concerning their use in PB smears and BMA samples.
Reviewed commercial digital morphology platforms for performance in blast detection.
Classical machine learning showed moderate blast recognition, relying heavily on chosen features.
Deep learning models, especially convolutional neural networks, reached sensitivities and specificities over 90% for blast detection in PB smears.
Commercial platforms had high concordance with manual methods for mature leukocytes, but less so for immature cells including blasts.

Abstract

ABSTRACT Introduction Morphologic evaluation of peripheral blood (PB) smears and bone marrow aspirates (BMA) remains central to the diagnosis of acute leukemias, particularly for the identification and quantification of blasts. However, this process is time‐consuming, operator‐dependent, and subject to inter‐observer variability. Recent advances in artificial intelligence (AI), particularly deep learning, have enabled the development of automated systems for leukocyte classification and blast detection. Methods We performed a narrative review of the literature on AI‐based approaches for morphologic evaluation in hematology, focusing on blast recognition and leukemia screening. Both classical machine learning and deep learning methodologies were analyzed, along with their application to PB smears and BMA samples. In addition, currently available commercial digital morphology platforms were reviewed with respect to their performance in blast detection. Results Classical machine learning approaches demonstrated good performance on maturing cells of most lineages, and moderate performance in blast recognition, limited by reliance on manually selected features. Deep learning models, particularly convolutional neural networks, achieved improved accuracy and near‐human performance in PB smear analysis, with reported sensitivities and specificities often exceeding 90% for blast detection. However, performance in BMA analysis remains more variable due to increased cellular complexity. Commercial platforms show high concordance with manual microscopy for mature leukocyte classification, but more modest accuracy for immature and neoplastic cells, including blasts. Conclusion AI‐based digital morphology systems are promising tools for supporting morphologic evaluation in hematology laboratories. While current platforms improve efficiency and standardization, limitations in blast detection accuracy and generalizability prevent their use as standalone diagnostic tools. Further development, including large‐scale validation and improved model interpretability, will be essential for their integration into routine clinical practice. The International Council for Standardization in Haematology (ICSH) presently suggests prudence in the widespread clinical adoption of AI‐driven bone marrow analysis systems, except within environments that are properly validated, approved by regulatory authorities, and supported by rigorous quality assurance measures. Their integration into routine practice will require broader validation, regulatory approval, and quality assurance frameworks.

Mark Helpful

Bookmark

Relay