PathPKT: Towards Understanding and Harnessing the Transferability of Prognostic Knowledge in Computational Pathology
In this work, we use pan-cancer WSI datasets (26 cancer, 11,188 WSIs, 9,190 patients) to explore the knowledge transferability in cancer prognosis. Concretely,
- We curate a WSI-based survival dataset for this study, called UNI2-h-DSS, derived from UNI2-h. It contains 8,818 WSIs from 7,268 patients, covering 13 cancer diseases.
- Based on UNI2-h-DSS, we find positive transfers across a wide range of cancer diseases, in line with human understanding of tumor biology.
- To understand the transferability of prognostic knowledge, we investigate
- What Knowledge Can Transferred Prognostic Models Offer?
- What Factors May Affect Transfer Performance?
- To harness the transferability, we porpose an MoE-based approach that shows to be effective and promising in exploiting the prognostic knowledge from other cancers.
Paper & Codes will come soon. Stay tuned.