Site Editor
Soo Park, MD
Soo Park, MD
Posted: Monday, May 5, 2025
Three pretrained machine-learning models significantly outperformed a standard baseline model in identifying nonmelanoma skin cancers from whole slides derived from biopsy samples, suggesting that such AI models could be used to diagnose cancer in resource-limited regions.
Song et al presented the findings at the 2025 American Association for Cancer Research (AACR) Annual Meeting (Abstract 1141). They noted that in resource-limited settings, diagnoses of nonmelanoma skin cancers are limited by the lack of expert pathologists to review biopsies. Based on the results of their study, the investigators believe that foundational models—AI models that are trained in vast amounts of data concerning nonmelanoma skin cancers—could be used as a ready-made tool for guiding diagnosis in more resource-limited regions.
Researchers tested three pathology foundation models (UNI, PRISM, and Prov-GigaPath) against the ResNet18 image recognition model as a baseline model. They used de-identified nonmelanoma skin cancers data from 2,130 whole-slide images derived from 553 biopsy samples that were part of the Bangladesh Vitamin E and Selenium Trial (BEST) conducted by the Institute for Population and Precision Health at the University of Chicago.
Accuracy in detecting nonmelanoma skin cancers from normal tissue was 92.5% with PRISM, 91.3% with UNI, and 90.8% with Prov-GigaPath compared with 80.5% for ResNet18 (P < .001). Even simplified versions of three pathology foundation models outperformed ResNet18, with accuracies of 88.2% with PRISM, 86.5% with UNI, and 85.5% with Prov-GigaPath.
“Overall, our results demonstrate that pretrained machine-learning models have the potential to aid diagnosis of nonmelanoma skin cancers, which might be particularly beneficial in resource-limited settings,” stated Steven Song, an MD/PhD candidate in the Medical Scientist Training Program at Pritzker School of Medicine and the Department of Computer Science at the University of Chicago, who presented the findings at AACR 2025. “Our study also provides insights that may advance the development and adaptation of foundation models for various clinical applications.”
Disclosure: For full disclosures of the study authors, visit aacr.org.
JNCCN Spotlights
