Reproduce “Momentum Memory for Knowledge Distillation in Computational Pathology” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yongxin Guo, Hao Lu, Onur C. Koyun, Zhengjie Zhu, Muhammet Fatih Demir, Metin Nafi Gurcan

Researchers at Wake Forest University developed Momentum Memory Knowledge Distillation (MoMKD), a framework that uses a momentum-updated memory bank to transfer genomic supervision to histopathology models, improving molecular biomarker prediction accuracy. The method achieved superior performance, including a +7.0% AUC increase for HER2 prediction over WSI-only MIL models and robust cross-domain generalization with a +3.8% AUC gain for Oncotype-DX on an independent dataset compared to existing multimodal KD approaches.

Key insights

  • A stable, momentum-updated memory bank can effectively mediate knowledge transfer between disparate modalities, providing a global supervisory context that significantly improves distillation stability and cross-domain generalization.
  • Decoupling gradients between genomic and histopathology branches is critical to prevent modality dominance and ensure the histology model learns robust, genuinely genomics-informed visual representations without relying on genomic input during inference.
  • The learned memory components are interpretable, demonstrating a direct correlation with specific histopathological patterns, such as tumor-rich or benign structures, linked to molecular biomarkers.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.