Evaluating Foundation Models with Pathological Concept Learning for Kidney Cancer

TL;DR

The paper addresses how foundation models can be evaluated for translational clinical use in kidney cancer via pathological concept learning grounded in TNM guidelines and pathology reports. It introduces a framework that uses patch-level features from four foundation models to build whole-slide graphs, which are analyzed by a graph neural network with ABMIL attention to identify interpretable pathological concepts. Through survival analysis with CoxPH models trained on these concepts, the work demonstrates improved prognostic performance and reveals actionable high-risk factors, while also assessing fairness across gender and race. The results suggest that concept-based explanations improve interpretability and may support subtype analyses and relapse prediction, advancing clinically meaningful deployment of foundation-model–driven pathology tools.

Abstract

To evaluate the translational capabilities of foundation models, we develop a pathological concept learning approach focused on kidney cancer. By leveraging TNM staging guidelines and pathology reports, we build comprehensive pathological concepts for kidney cancer. Then, we extract deep features from whole slide images using foundation models, construct pathological graphs to capture spatial correlations, and trained graph neural networks to identify these concepts. Finally, we demonstrate the effectiveness of this approach in kidney cancer survival analysis, highlighting its explainability and fairness in identifying low- and high-risk patients. The source code has been released by https://github.com/shangqigao/RadioPath.

Figures (5)

• Figure 1: Overview of pathological concept learning. (a) The framework of pathological concept learning; (b) benchmarking foundation models in identifying pathological concepts; (c) Explainable survival analysis based on pathological concepts; and (d) Identification of spatial phenotypes by concept-orientated attention map.
• Figure 2: Benchmarking foundation models by pathological concept learning on kidney cancer.
• Figure 3: Kidney cancer survival analysis. The left shows the coefficients of top 10 high risk factors leading to mortality. The middle shows the AUC at different time points. The right shows the survival curves of high- and low-risk groups.
• Figure 4: Evaluation of fairness in terms of gender. The left(right) shows the survival curves of female and male low-risk(high-risk) groups.
• Figure 5: Evaluation of fairness in terms of race. The left(right) shows the survival curves of white and black low-risk(high-risk) groups.