Foundation models match but sometimes lag task-specific models in prostate pathology
The study evaluated more than 100,000 prostate biopsy images from 6,268 patients across 15 sites, comparing self‑supervised foundation models (FMs) with task‑specific (TS) end‑to‑end models. Both model families reached high agreement with pathologists, achieving quadratic Cohen’s kappa of 0.91 (95 % CI 0.90‑0.92) on internal data. On external cohorts, TS models scored 0.87 (0.86‑0.88) while the…
Key points
- Using over 100,000 prostate biopsies from 6,268 patients across 15 sites, the study compared foundation models and task‑specific models.
- Task‑specific models achieved Cohen’s kappa up to 0.91 internally and 0.87 externally, matching or exceeding foundation models.
- Foundation models required up to 35× more energy than task‑specific models while offering no clear accuracy advantage.
When sufficient labeled data were available, TS models converged with or surpassed FMs, reducing misgrading, misdiagnosis of challenging morphologies, and technical variability. However, FMs consumed up to 35 times more energy than TS models, raising concerns about resource efficiency. The authors suggest that both approaches be considered for robust, clinically applicable AI pathology solutions.
The work was supported by multiple European research grants and acknowledges shareholders in Clinsight AB and A3P Biomedical AB among the authors. The article is licensed under a Creative Commons Attribution 4.0 International License.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- iMerit launches Scholars platform to scale expert AI training and evaluation · 1 src
- Machine learning explores quantum superpositions but cannot solve Schrödinger’s cat paradox · 1 src
- KDnuggets outlines four-stage roadmap for computer Vision engineers · 1 src
- Resolution autoformalizes math to speed AI alignment research · 1 src
- Graphite finds ChatGPT still rarely mentions many brands even when using search · 1 src
Comments
via GitHub Discussions