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

Area of Science:

• Hematology
• Computational Biology
• Artificial Intelligence

Background:

• Leukemia diagnosis relies on flow cytometry, but manual analysis is slow and error-prone.
• Automated diagnostic tools are needed to improve accuracy and efficiency in leukemia detection.

Purpose of the Study:

• To develop and evaluate a deep learning-based automated system for analyzing bone marrow flow cytometry samples in leukemia diagnosis.
• To compare the performance of recurrent, graph-based, and attention-enhanced convolutional neural networks.

Main Methods:

• Developed three deep learning models: recurrent, graph-based, and attention-enhanced convolutional networks (VGG19 architecture).
• Trained and evaluated models on over 2,000 bone marrow flow cytometry samples from leukemia patients and healthy individuals.
• Utilized Explainable AI techniques for model transparency and evaluated generalization on unseen patients.

Main Results:

• The attention-enhanced VGG19 convolutional model achieved 96% accuracy, with near-perfect discrimination between disease states.
• Explainable AI confirmed the model's focus on biologically relevant cell clusters.
• Generalization challenges were addressed via hyperparameter optimization; standard augmentation degraded performance.

Conclusions:

• The developed deep learning system offers a highly accurate and interpretable tool for accelerating leukemia diagnosis.
• This AI approach can significantly reduce human error and support pathologists in clinical decision-making.
• Further research is needed for application in low-level disease states and minimal residual disease (MRD) detection.