Abstract B014: Are histopathology foundation models clinically ready for survival prediction?

Abstract Purpose: Recently, large foundation models (FMs) have advanced survival prediction from whole slide images (WSIs) in histopathology. However, their generalizability to new datasets without retraining remains largely untested. In this study, we assess their deployment performance by training on one dataset and testing on another, effectively testing their utility in real-world clinical settings. Methods: For our study, we used 1,830 slides and 1,754 patients (TCGA BRCA: 1,060 slides, 994 patients; Ontario Tumor Bank (OTB): 760 slides, 760 patients). First, we confirmed significant distribution differences using log-rank tests, t-SNE plots, and logistic regression - an issue commonly encountered during model deployment. We then selected a diverse set of state-of-the-art models to represent the current AI landscape. Due to the large size of WSIs, slide-level predictions are typically performed using the Multiple Instance Learning (MIL) framework, where information from small patches is aggregated to generate WSI-level predictions. We used the patch-level FM CTransPath as the backbone for MIL-based slide-level survival models, evaluating various popular approaches, including attention-based, transformer-based, and graph-based models. Additionally, we assessed the slide-level FM TITAN, a pre-trained model that directly generates slide-level embeddings from WSIs. To predict progression-free interval, we applied a linear Cox proportional hazards (CPH) model to the generated slide embeddings. We also implemented a simple clinical model using the linear CPH model based on five widely available variables: progesterone receptor status, tumor stage, lymph node stage, metastasis status, and age. The models were rigorously evaluated in-domain (ID) using five-fold cross-validation and out-of-domain (OOD) for deployability by testing the trained fold models on the external dataset. Results: Our findings highlight two key points. First, the clinical model outperformed the best image-only survival models in both ID [BRCA: +9.2%, OTB: +14.9%] and OOD settings [BRCA: +8.6%, OTB: +25.8%], achieving state-of-the-art C-index scores [ID: BRCA 0.69, OTB 0.74; OOD: BRCA 0.67, OTB 0.76]. Notably, the deployed clinical model remained robust, even exceeding its ID performance on OTB. Second, while MIL-based image models performed on par with or better than slide-level FMs, they showed substantial drops from ID to OOD [BRCA: -6.8%, OTB: -8.0%]. In contrast, slide-level FMs showed better deployability, with only minor declines across domains [BRCA: -2.0%, OTB: -2.4%]. However, they still lagged significantly behind clinical models in both ID and OOD settings. Conclusions: Despite being trained on large datasets and requiring substantial computation, FMs face significant challenges. Although WSI-based survival models have made considerable progress, our findings show they still fall short of simple clinical models. Further research is needed to improve algorithm performance and generalizability, ensuring their readiness for real-world clinical applications. Citation Format: Vishwesh Ramanathan, Dianne Chadwick, Lincoln Stein, Anne L. Martel. Are histopathology foundation models clinically ready for survival prediction? [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B014.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.