Deep learning in bone marrow cytomorphology: advances in segmentation, classification, and clinical translation

Bone marrow cytomorphology analysis has long been a prerequisite for diagnosing hematologic disorders. It is often a tedious and subjective activity through conventional manual methods. The recent advancement in deep learning (DL) has a bright future with potential automation in cell classification, segmentation, and diagnosis workflows. This review outlines the current state-of-the-art application of DL for bone marrow analysis, focusing mainly on challenges such as data scarcity, class imbalance, and inter-center variability. It includes an evaluation of publicly available datasets together with strategies for overcoming limitations, including synthetic data generation and federated learning. This review analyzes segmentation techniques, ranging from the classical watershed algorithm to novel U-Net (a specialized neural network for image segmentation) and Vision Transformer hybrids, for their efficacy in isolating cells, tissue subsystems, and subcellular structures. DL classification models, including convolutional neural networks (CNNs), attention-based architectures, and ensembles, demonstrate expert-level accuracy in detecting malignancies like acute myeloid leukemia (AML) and myelodysplastic syndromes (MDS). Clinical applications involve AI-driven platforms that integrate into digital pathology workflows, with early reports suggesting reductions in diagnostic turnaround time and inter-observer variability. For crowded images of the bone marrow, context-aware analysis is enhanced by hybrid models combining CNN and Transformer architectures. However, there are challenges with generalizability and interpretability, as well as difficulties in integrating multimodal data. Therefore, future directions emphasize validation, explainable AI, and integrating cytomorphology with genetic and flow cytometry data. This interdisciplinary work aims to bridge the fields of AI and hematopathology towards standardization and precise diagnostics, which would improve clinical decision-making and patient outcomes.