Diagnosis of Leukemia from Bone Marrow Flow Cytometry Data Using Deep Learning and Explainable Artificial...

Leukemia is a life-threatening blood cancer requiring rapid and accurate diagnosis to improve patient survival. Although flow cytometry offers powerful multiparametric analysis, manual interpretation remains slow, subjective, and, thus, prone to error. This study presents an automated diagnostic system using deep learning to analyze bone marrow flow cytometry samples from this technology. Three advanced neural network models were developed and evaluated: recurrent, graph-based, and attention-enhanced convolutional networks. The data set comprised >2000 samples from patients with leukemia and healthy individuals. The attention-enhanced Visual Geometry Group 19‑layer (VGG19) architecture convolutional model delivered the highest performance, achieving 96% accuracy and near-perfect discrimination between disease states. Additionally, explainable artificial intelligence techniques enhanced transparency, confirming the model focused on biologically relevant cell clusters. To ensure real-world applicability, the model was rigorously evaluated on completely unseen patients, revealing generalization challenges addressed through careful hyperparameter optimization. Furthermore, standard image augmentation techniques degraded performance significantly, highlighting the need for domain-specific data approaches. Notably, this study focused on newly diagnosed bone marrow specimens with high tumor burden. Consequently, overall reported performance may not directly generalize to low-level disease states or minimal residual disease scenarios. Therefore, this work demonstrates a highly accurate, interpretable, and clinically relevant tool that can ultimately accelerate leukemia diagnosis, significantly reduce human error, and support expert pathologists in their critical clinical work.