Shixart1985 / Wikimedia Commons · CC BY 2.0 · Image source
Research: models may learn the laboratory as well as the tissue
Published on 11 June 2026, PathoROB evaluated twenty foundation models for sensitivity to non-biological features such as laboratory procedures and scanner differences. All showed robustness weaknesses, with substantial variation between models. Mitigations reduced but did not eliminate the risk. This is a research benchmark, not an incident report establishing harm to patients. Some authors disclosed relationships with companies developing models included in the evaluation.
Compare the task, not just a league table
Campanella and colleagues’ study, published on 17 April 2025, compared public models using slides generated during routine work at three medical centres. It covered endpoints including cancer diagnoses and biomarkers and offered an evaluation pipeline for external users. This helps frame model selection around a specific task. Evaluation within those datasets does not replace validation in a new laboratory.
Clinical translation extends beyond the algorithm
NCI’s 15 April 2026 seminar addresses both large annotated datasets and the framework needed to translate robust models into clinical practice. Validation and evaluation of AI-enabled medical devices are distinct concerns. The seminar is educational material, not a regulatory authorisation document for the models discussed here. Research findings, local acceptance and a product’s authorised scope should be documented separately.
Editorial interpretation: define a local pilot with four questions
Which tissue and task will be evaluated? What preparation and scanning conditions apply? How is test data separated from data used to develop the model? Who sets acceptance criteria and reviews disagreements? Our recommendation is to answer these questions explicitly in the pilot record and include images that cannot be assessed. The distribution of failures deserves attention alongside average performance.
Keep changes visible
When a model version, scanner or preparation process changes, it is useful to revisit which conditions the previous acceptance covered. This is our practical interpretation of the sources, not a mandatory validation protocol for a particular product. The article offers no individual diagnosis or treatment advice and should not be read as evidence of clinical effectiveness for RAI products.
Sources
- Kömen et al. — Towards robust foundation models for digital pathology (Nature Communications, 11 June 2026) ↗
Accessed: 2026-10-14 - Campanella et al. — A clinical benchmark of public self-supervised pathology foundation models (Nature Communications, 17 April 2025) ↗
Accessed: 2026-10-14 - NCI — Validating AI in Digital Pathology for Clinical Use Beyond the Algorithm (15 April 2026) ↗
Accessed: 2026-10-14
This article is for education and information. Referencing an organization does not imply its endorsement of RAI PACS Pro. This is not a claim of clinical effectiveness for a product or model.