Cross-Cancer Knowledge Transfer in WSI-based Prognosis Prediction
Pei Liu, Luping Ji, Jiaxiang Gou, Xiangxiang Zeng
TL;DR
This paper introduces a new paradigm for cancer prognosis prediction using whole-slide images, focusing on cross-cancer knowledge transfer to improve performance, especially for rare tumors, through a systematic study and a baseline approach.
Contribution
It pioneers the systematic study of cross-cancer knowledge transfer in WSI-based prognosis prediction and proposes a routing-based baseline method, CROPKT, for efficient knowledge utilization.
Findings
CROPKT effectively measures transferability across 26 cancers.
Deep insights into transfer mechanisms are gained through experiments.
Cross-cancer transfer improves prognosis prediction, especially for rare tumors.
Abstract
Whole-Slide Image (WSI) is an important tool for estimating cancer prognosis. Current studies generally follow a conventional cancer-specific paradigm in which each cancer corresponds to a single model. However, this paradigm naturally struggles to scale to rare tumors and cannot leverage knowledge from other cancers. While multi-task learning frameworks have been explored recently, they often place high demands on computational resources and require extensive training on ultra-large, multi-cancer WSI datasets. To this end, this paper shifts the paradigm to knowledge transfer and presents the first preliminary yet systematic study on cross-cancer prognosis knowledge transfer in WSIs, called CROPKT. It comprises three major parts. (1) We curate a large dataset (UNI2-h-DSS) with 26 cancers and use it to measure the transferability of WSI-based prognostic knowledge across different cancers…
Peer Reviews
Decision·Submitted to ICLR 2026
A key strength of the paper is its clear, hypothesis-driven approach. The authors propose that prognostic knowledge can be transferred across different cancer types and that this process is governed by identifiable factors. They then systematically test these hypotheses through a series of well-designed experiments. This structured investigation, which progresses from demonstrating feasibility to exploring underlying mechanisms, provides solid support for the paper's conclusions. Another signif
1. **Limited Generalizability Due to Reliance on a Single Foundation Model:** The study's conclusions are entirely contingent on the feature space of a single foundation model, UNI2-h. The observed transferability patterns (Figure 2) and the predictive power of inter-task factors might be specific artifacts of UNI2-h's architecture and training data. This limits the generalizability of the core findings. * **Actionable Suggestion:** To strengthen the claims, the authors should validate th
1. The paper is quite well-written and easy to follow. 2. The paper proposes using knowledge transfer methods to address the challenge of rare tumors, which is a natural and direct solution. 3. The experimental results demonstrate the main points of this paper (transfer learning does helps when samples are limited).
1. Using only the ABMIL model is questionable. Survival prediction is quite hard; even with gene information, multimodal survival prediction performance is still not fully satisfactory, unlike WSI classification (C-index around 0.5-0.7, few 0.8). The results in Figure 2 are obtained with only the simplest WSI model (ABMIL), which makes it less convincing. To prove the transferability across organs, it would be better if the authors could demonstrate multiple models all converge to similar patter
- Explainability with attention heatmaps and correlation analysis are interesting and insightful. - Evaluation is extensive, performed across 26 cancer types. - The routing method for fusing models is novel and interesting. Benchmarking against other ensemble approaches is also extensive. - The authors will release their UNI-2-h pan-cancer feature dataset, reducing the barrier for additional works on this topic.
- The writing quality is quite poor, in terms of both grammar and flow. I strongly recommend the authors work closely with an LLM to improve the writing quality while maintaining the desired meaning. - Line 203: Positive transfer should be compared against random initialization and/or mean pooling. With good feature encoders, even these simple baselines can achieve competitive performance. - Line 210: The existence of positive transfer for a single task is not necessarily indicative of transfera
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsRadiomics and Machine Learning in Medical Imaging · AI in cancer detection