Childhood Leukemia Classification via Information Bottleneck Enhanced Hierarchical Multi-Instance Learning.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning in hematology

Background:

• Leukemia classification requires expert cytomorphological analysis of bone marrow smears, a time-consuming manual process.
• Current deep learning methods for bone marrow analysis face limitations in generalization and require extensive cell-level annotations.
• Existing methods fail to leverage hierarchical relationships among leukemia subtypes.

Purpose of the Study:

• To develop a data-efficient deep learning framework for leukemia classification using patient-level labels.
• To overcome limitations of existing methods by incorporating hierarchical relationships and improving generalization.
• To reduce the reliance on manual, cell-level annotations in bone marrow smear analysis.

Main Methods:

• Proposed a hierarchical Multi-Instance Learning (MIL) framework.
• Integrated attention-based learning to identify diagnostically valuable cells across different hierarchies.
• Employed a hierarchical Information Bottleneck (IB) to refine feature representations for enhanced accuracy and generalization.

Main Results:

• The framework successfully identified diagnostic cells without requiring cell-level annotations.
• Achieved superior performance compared to existing methods on a large-scale childhood acute leukemia dataset.
• Demonstrated high generalizability on an independent test cohort.

Conclusions:

• The proposed hierarchical MIL with IB framework offers a robust and data-efficient approach for leukemia classification.
• This method significantly improves accuracy and generalization in medical image analysis for hematological malignancies.
• It provides a viable alternative to manual bone marrow smear examination, reducing workload and potential for error.