Comprehensive learning and adaptive teaching: Distilling multi-modal knowledge for pathological glioma grading.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Multi-modal data fusion (pathology slides, genomic profiles) enhances glioma grading but genomic data is costly and difficult to obtain.
• This limits clinical applications of multi-modal diagnostic approaches.
• A realistic scenario involves paired data for training and pathology-only data for inference.

Purpose of the Study:

• To develop a framework for improving pathological glioma grading using knowledge transfer from multi-modal data.
• To address the challenge of limited genomic data availability during clinical inference.
• To enhance the performance of pathology-based grading models by leveraging privileged multi-modal information.

Main Methods:

• A comprehensive learning and adaptive teaching framework is proposed.
• A Saliency-Aware Masking (SA-Mask) strategy is introduced for richer feature extraction from multi-modal data.
• A Local Topology Preserving and Discrepancy Eliminating Contrastive Distillation (TDC-Distill) module aligns teacher-student feature distributions.
• A Gradient-guided Knowledge Refinement (GK-Refine) module adaptively refines knowledge transfer.

Main Results:

• The proposed distillation framework significantly improves pathological glioma grading performance.
• The method outperforms existing knowledge distillation techniques.
• Achieves comparable performance to existing multi-modal methods using only pathology slides.

Conclusions:

• The developed framework effectively transfers knowledge from multi-modal data to pathology-only models for glioma grading.
• The approach overcomes the limitations of genomic data accessibility in clinical settings.
• This method offers a viable solution for enhancing diagnostic accuracy in glioma grading.