Patch-Level Tissue Context Improves Learning from Frozen Pathology Foundation Model Embeddings

Pathology foundation models are often reused as frozen feature extractors. This is efficient, but a downstream classifier may treat every cell or image region independently even though cells exist inside organized tissue environments.

This paper evaluates lightweight fusion of local patch-level context with frozen embeddings from Prov-GigaPath, Virchow2, and DINOv2 on BRCA-M2C. Patch context generally improves or stabilizes classification, with gains that vary by encoder, while keeping the foundation-model backbone frozen.

• Read the paper
• Code and reproducibility artifacts

Citation

@inproceedings{dip2026patch,
  author    = {Dip, Sajib Acharjee and Zhang, Liqing},
  title     = {Patch-Level Tissue Context Improves Learning from Frozen Pathology Foundation Model Embeddings},
  booktitle = {Proceedings of the 17th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics},
  year      = {2026},
  articleno = {72},
  numpages  = {6},
  doi       = {10.1145/3807503.3820870}
}

Recommended citation: Dip, Sajib Acharjee, and Liqing Zhang. (2026). "Patch-Level Tissue Context Improves Learning from Frozen Pathology Foundation Model Embeddings." Proceedings of ACM BCB 2026, Article 72, 6 pages.
Download Paper