Pretrained Machine Learning Models Enhance Diagnosis of Nonmelanoma Skin Cancer in Low-Resource Settings

Advancements in artificial intelligence are opening new horizons in the detection of nonmelanoma skin cancers (NMSC), especially in areas with limited medical resources. A recent study presented at the AACR 2025 Annual Meeting highlights how pretrained machine learning models, also known as foundation models, can significantly improve the accuracy and accessibility of skin cancer diagnosis.

The research team from the University of Chicago Medical Center evaluated three state-of-the-art foundation models—PRISM, UNI, and Prov-GigaPath—in analyzing digital pathology images of skin lesions. These models work by breaking down high-resolution tissue images into smaller tiles, extracting relevant features, and calculating the likelihood of the presence of NMSC. The study involved 2,130 tissue images from 553 patients in Bangladesh, a population with heightened risk for NMSC due to arsenic exposure from contaminated water.

Results demonstrated that all three foundation models outperformed traditional models like ResNet18, achieving accuracy rates of over 90%. Specifically, PRISM correctly identified NMSC in 92.5% of cases, UNI in 91.3%, and Prov-GigaPath in 90.8%, compared to 80.5% with ResNet18. Simplified versions of these models, requiring less computational power and analysis, still maintained high accuracy, proving their potential for deployment in resource-constrained environments.

Additionally, the researchers developed an innovative annotation framework that leverages small sets of biopsy images to highlight suspect regions on tissue slides. This approach aids clinicians by directing attention to areas most likely to contain cancerous tissue without needing extensive training data.

Despite these promising results, the study acknowledged limitations, such as its focus on a Bangladeshi cohort, which may affect the generalizability of the findings. Practical challenges, including digital pathology infrastructure and integration into clinical workflows, need further exploration before widespread implementation.

Overall, this research suggests that pretrained machine learning models can serve as effective, low-resource tools to assist in diagnosing NMSC globally, particularly where specialist expertise and equipment are scarce. Future work will aim to validate these models across diverse populations and address logistical aspects of real-world deployment, bringing us closer to more equitable skin cancer diagnosis.

Source: https://medicalxpress.com/news/2025-04-pretrained-machine-nonmelanoma-skin-cancer.html