Acute Myeloid Leukemia subtype classification using single blood cell images.

Area of Science:

• Hematology
• Computational Biology
• Medical Imaging

Background:

• Acute Myeloid Leukemia (AML) is a complex blood cancer requiring precise subtype identification for treatment.
• Current diagnostic methods can be labor-intensive and may benefit from advanced computational approaches.

Purpose of the Study:

• To develop and validate a novel AI-driven framework for the accurate detection and subtype classification of Acute Myeloid Leukemia (AML).
• To enhance the robustness and generalizability of image-based AML diagnostics across different datasets.

Main Methods:

• A two-step classification strategy employing custom Convolutional Neural Networks (CNNs) and machine learning algorithms.
• An "ensemble of ensembles" approach with soft voting for improved classification accuracy.
• Implementation of a preprocessing pipeline including stain normalization and brightness correction to handle dataset variability.

Main Results:

• The single-cell classifier achieved 94.7% accuracy, 95.2% precision, and 94.8% F1-score.
• The proposed ensemble model demonstrated 94% accuracy in distinguishing between healthy and AML patients.
• Exploration of Multiple Instance Learning and pseudo-labeling provided comparative insights.

Conclusions:

• The developed framework shows significant promise for accurate, image-based AML detection and subtype classification.
• The methodology offers a practical foundation for advancing automated hematological diagnostics.
• Addressing domain shift through preprocessing and cross-dataset validation is crucial for real-world applicability.