A preprint posted on arXiv tackles a persistent problem in pathology: pattern recognition often depends on a clinician's training and experience, which may fall short when encountering rare cancer subtypes. The authors propose using pathology foundation models for case retrieval, drawing on archives of previously diagnosed tumors to provide a frame of reference.

The key innovation is that the method is described as "training-free" — it does not require fine-tuning on a specific task. Instead, the foundation model is used directly to retrieve similar cases from an archive, potentially helping pathologists when faced with unusual salivary gland tumors. The paper focuses on this tumor group, which includes many rare subtypes.

Because the source abstract is truncated, the full methodology and results are not available. The paper is a preprint and has not yet been peer-reviewed. Its significance lies in the possibility of supporting diagnosis of rare tumors without the need for extensive labeled datasets or task-specific model training.