Foundation models in computational pathology: evidence and limits

A foundation model is generally pretrained on a broad dataset and then adapted to downstream tasks. In computational pathology, the approach is attractive because whole-slide images contain many tissue patterns and because task-specific labels can be expensive. The central question is not whether a large encoder can produce useful features, but where those features transfer and where task-specific evidence is still required.

What the published studies show

The UNI study evaluated a self-supervised pathology model across representative computational pathology tasks and reported transfer across tissue types and task settings. The Virchow study examined large-scale pretraining and downstream applications including biomarker prediction and cancer detection. These are results from particular datasets, methods, baselines, and evaluation designs. They are evidence about those studies, not a universal guarantee for every pathology task.

Transfer can fail for ordinary reasons

• The target tissue, stain, scanner, or preparation process may differ from the pretraining data.
• A representation can support one task while missing features needed for another.
• Slide-level prediction often needs an aggregation method whose behavior is separate from the tile encoder.
• Labels may reflect local practice, incomplete records, or noisy reference standards.
• Calibration and uncertainty can change after adaptation, especially on out-of-distribution data.

Large pretraining datasets can improve the diversity of learned representations, but dataset size alone does not establish generalization. The provenance, patient and specimen separation, tissue coverage, label quality, preprocessing, model access, and evaluation population all affect what a reported result means.

Evaluate the transfer claim, not just the encoder

A useful evaluation compares the foundation model with appropriate baselines on the intended task, includes external or held-out data, and reports uncertainty and failure patterns. It should distinguish tile-level, region-level, and slide-level behavior. It should also examine whether the downstream head, aggregation strategy, threshold, and reference standard are driving the result.

A foundation model may be a research component, a feature extractor, or part of a system intended for clinical decision support. Those uses require different descriptions of intended use and different evidence. Benchmark performance does not by itself establish safety, clinical utility, regulatory status, or suitability for a particular laboratory.

The most durable conclusion is cautious: foundation models can reduce repeated representation-learning work and support broader experiments, while their limits remain task-specific. Reproducible data documentation, external evaluation, transparent reporting, and explicit uncertainty are necessary for deciding what a model can reasonably support.