- The paper systematically benchmarks 13 pathology foundation models for breast cancer survival prediction, revealing incremental performance gains in various cohorts.
- It employs a unified pipeline with frozen patch encoders and the PANTHER framework to ensure rigorous cross-cohort validation and risk stratification.
- Distilled compact models like H0-mini achieve near state-of-the-art accuracy with over 92% parameter reduction, emphasizing efficiency for clinical deployment.
Benchmarking Pathology Foundation Models for Breast Cancer Survival Prediction
Introduction
The systematic benchmarking of pathology foundation models (PFMs) for survival prediction is a critical step for their clinical adoption in computational oncology. Despite rapid progress in large-scale representation learning for digital pathology, robust cross-cohort comparisons for time-to-event prediction tasks remain limited. This work provides a comprehensive head-to-head evaluation of thirteen PFMs for survival analysis in breast cancer, utilizing datasets covering more than 5,400 patients with long-term follow-up. The evaluation framework emphasizes both generalization to independent external cohorts and practical aspects relevant for deployment, such as model size and inference cost.
Experimental Framework and Model Selection
The benchmark employs a unified pipeline leveraging patch-level feature extraction from WSIs and risk score aggregation with the PANTHER survival modeling framework. Each PFM is used as a frozen patch encoder. Survival models are trained on the SöS-BC-4 cohort (2,315 patients) and evaluated on two external datasets (KS-Solna and SCAN-B-Lund; 3,119 patients), strictly excluding any pretrain/evaluation overlap.
The 13 evaluated models span multiple generations and paradigms:
- Natural image baseline: Resnet-IN (ImageNet pretrained).
- Early pathology encoders: CTransPath, RetCCL.
- State-of-the-art PFMs: UNI, UNI2-h, H-optimus-0, H-optimus-1, Prov-GigaPath, Virchow, Virchow2.
- Compact PFM: H0-mini (distilled from H-optimus-0).
- Vision-language PFMs: CONCH, CONCHv1.5.
Survival prediction is assessed using both recurrence-free survival (RFS) and progression-free survival (PFS) endpoints, on all patients and the clinically distinct ER+ & HER2- subgroup. Performance is quantified via the concordance index (C-index), complemented by qualitative risk stratification analyses using Kaplan-Meier estimation.
Main Findings
- Best overall results are consistently produced by H-optimus-1, which achieves top C-index results for three of four evaluation settings, with values reaching 0.702 for PFS (All Patients).
- Distilled compact model H0-mini achieves the second-best ranking, slightly surpassing its teacher, H-optimus-0, despite a parameter reduction by over 92%. This demonstrates efficient knowledge transfer through distillation and significant reductions in computational footprint.
- Other leading models include Virchow2, UNI2-h, and CONCHv1.5, with aggregated rankings indicating generational improvements.
- Older architectures and non-domain-specific models (Resnet-IN, CTransPath, RetCCL) underperform considerably, establishing the necessity of domain-adapted pretraining on large pathology datasets.
- Performance differences among the top PFMs are modest, with substantial overlap in C-index confidence intervals, especially outside the two lowest-ranked models.
Generalization and Robustness
Model rankings are largely stable across endpoints (RFS/PFS) and subpopulations, but some models (e.g., CONCHv1.5, Virchow) change their relative ranking depending on the endpoint, reflecting the impact of event rates and possibly endpoint-specific feature learning.
The data-efficiency analysis indicates that stronger PFMs maintain their performance advantage across all training data regimes. Notably, H-optimus-1 with 25% training data matches or exceeds the full-data performance of Resnet-IN, indicating the robustness of PFM representations even in limited data scenarios.
Qualitative Risk Stratification
Kaplan-Meier analysis shows all evaluated models can stratify patients by risk group, but the separation is most pronounced for H-optimus-1, moderate for UNI, and weakest for Resnet-IN. The superiority of PFMs is clearest in more challenging multi-group stratification tasks.
Representation Analysis
UMAP projections reveal clear clustering by dataset for recent PFMs, reflecting sensitivity to cohort-specific factors. However, no large-separation of patients by survival outcome is observed, suggesting that prognostic features are highly subtle and non-linearly embedded in the representation space, and that institution- or acquisition-based batch effects are prominent.
Theoretical and Practical Implications
A salient outcome is that scaling PFMs by pretraining data and model size exhibits diminishing returns for survival prediction. The improvements, while statistically significant, are small. For instance, H-optimus-1 (2× pretraining data versus H-optimus-0) yields only incremental C-index gains. Similarly, top-10 model rankings show no strong correspondence with absolute parameter count (cf. H0-mini vs. Prov-GigaPath), emphasizing that training strategy and data curation outpace brute-force model enlargement.
Efficiency and Deployment
Efficient PFMs such as H0-mini present a practical pathway to clinical integration by reducing computation and storage burdens while matching the prognostic signal extraction of massive teacher models. Distillation using relatively small, publicly available WSI sets is sufficient for high-fidelity compression, providing evidence that distillation and compact model design should be prioritized for deployment in resource-constrained and real-time settings.
Vision-Language Pretraining
Vision-language PFMs (CONCH, CONCHv1.5) performed competitively, suggesting that paired multimodal supervision can enhance pathology representation learning. However, due to confounded factors (dataset, protocol, architecture), definitive attribution of gains to the multimodal approach remains open. Systematic, controlled comparisons are warranted.
Data and Event Sampling
Survival modeling benefits from increased amounts of clinical endpoint data (i.e., more events, longer follow-up), not just overall sample size. The limited number of PFS events relative to RFS underscores the necessity of aggregating larger cohorts and maximizing longitudinal coverage.
Study Limitations
- All datasets originate from Swedish regional hospitals, so results may not generalize to broader international or cross-modality settings.
- Only breast cancer is considered; extension to other cancer types is not directly validated.
- The evaluation restricts PFMs to frozen feature extractors in a standardized pipeline (PANTHER + MLP). Alternative fine-tuning strategies or survival models could yield different rankings.
Conclusion
This work establishes a robust, externally validated benchmark for pathology foundation models in breast cancer survival prediction. The principal findings are:
- Recent PFMs consistently outperform early baselines, but advances are incremental with overlapping absolute performance.
- Distilled compact models provide key efficiency benefits without sacrificing accuracy, motivating their use in scalable and real-world applications.
- Further scaling of models and data shows diminishing returns; quality and diversity of training data, and efficient model compression, are now critical bottlenecks.
- Prognostic feature learning remains sensitive to batch effects and cohort-specific variance, highlighting the importance of robust validation and potential domain adaptation strategies.
- Vision-language pretraining is a promising, yet underexplored, strategy for enhancing generalizability in morphological representation learning.
Future work should aim to extend these benchmarks to multi-center and multi-disease settings, dissect the factors influencing representation quality, and optimize for the best balance of efficiency and performance using distillation and multimodal learning.
