A Knowledge-Reserved Distillation with Complementary Transfer for Automated FC-based Classification Across...

Area of Science:

• Hematology
• Computational Biology
• Machine Learning

Background:

• Acute leukemia presents a significant clinical challenge with life-threatening outcomes.
• Measurable residual disease (MRD) detection via flow cytometry (FC) is crucial but interpretation is subjective and time-consuming.
• Existing machine learning models for MRD classification lack generalization due to small datasets or single-leukemia type focus.

Purpose of the Study:

• To develop a generalized machine learning framework for automatic MRD classification in acute leukemia.
• To address the limitations of physician-dependent MRD interpretation and improve diagnostic efficiency.
• To enhance the classification performance on smaller acute lymphoblastic leukemia (ALL) cohorts by leveraging larger acute myeloid leukemia (AML) datasets.

Main Methods:

• A transfer learning approach using a knowledge-reserved distilled AML pre-trained network.
• Complementary learning was applied to adapt the pre-trained model for ALL MRD classification.
• The framework was validated for its transferability across different acute leukemia types.

Main Results:

• The proposed framework achieved an average AUC of 84.5%, demonstrating effective transferability.
• The model showed improved performance in classifying MRD in ALL samples.
• Analysis indicated that younger and elder ALL patients particularly benefited from the AML pre-trained model.

Conclusions:

• Transfer learning offers a viable solution for generalized MRD classification in acute leukemia.
• Leveraging large AML datasets can significantly improve ALL MRD classification performance.
• The developed model shows promise for more objective and efficient MRD assessment in clinical practice, with specific benefits for distinct age groups.