Consensus and Complementary Feature Guided Multi-modal Knowledge Distillation Network for Breast Cancer Diagnosis.

Area of Science:

• Digital pathology
• Computational oncology
• Biomedical imaging analysis

Background:

• Accurate HER2 status assessment is crucial for breast cancer treatment decisions.
• Hematoxylin and Eosin (H&E) and Immunohistochemistry (IHC) staining provide complementary data for HER2 evaluation.
• IHC's cost and accessibility limitations hinder routine clinical use of multi-modal approaches.

Purpose of the Study:

• To develop an effective HER2 grading framework utilizing solely H&E images.
• To enable precise HER2 status assessment without relying on expensive IHC.

Main Methods:

• A two-stage diagnostic framework involving Consensus and Complementary Feature Co-Embedding Network (CoCoFNet) and Hierarchical Multi-modal Knowledge Distillation (HM-KD).
• CoCoFNet extracts modality-specific and cross-modal features for comprehensive representation.
• HM-KD transfers knowledge from a multi-modal teacher to a unimodal student network.

Main Results:

• The proposed method achieves performance comparable to state-of-the-art distillation techniques for HER2 grading using only H&E images.
• CoCoFNet enhances feature fusion, leading to improved supervision and generalization in the unimodal student network.
• Experimental validation on two public datasets confirms the efficacy of the H&E-only approach.

Conclusions:

• The developed framework offers a cost-effective and accessible alternative for HER2 grading in breast cancer.
• This approach democratizes precise HER2 status assessment, particularly in resource-limited settings.
• The study highlights the potential of advanced deep learning techniques to overcome limitations in traditional diagnostic workflows.