AI-assisted Diagnosis of Nonmelanoma Skin Cancer in Resource-Limited Settings

Spencer Ellis, Steven Song, Derek Reiman, Xuan Hui, Renyu Zhang, Hasan Shahriar, Maria Argos, Mohammed Kamal, Christopher R. Shea, Robert L. Grossman, Aly A. Khan, Habibul AhsanCancer 2025 · DOI

TL;DR

This paper explores the use of AI-powered "foundation models" (FMs) for diagnosing nonmelanoma skin cancer (NMSC) from histology images, particularly in resource-limited settings where expert pathologists are scarce. They found that a vision-language model called PRISM, when combined with specific aggregation and classification techniques, significantly outperformed other FMs and a traditional baseline, offering a promising avenue for improving global cancer diagnostics.

What Problem Does This Paper Address?

Accurate and timely diagnosis of skin cancer is crucial for effective treatment and patient outcomes. However, in many parts of the world, access to skilled pathologists is limited, leading to delayed diagnoses and poorer prognoses. Traditional computational pathology approaches often require vast amounts of labeled data, which is expensive and time-consuming to acquire. This paper addresses this challenge by investigating the potential of general-purpose "foundation models" (FMs) – large, pre-trained AI models that can generalize across various tasks and datasets – to diagnose NMSC from whole-slide histology images without extensive task-specific fine-tuning. The goal is to develop AI tools that can assist in diagnosis, especially in settings where expert human resources are constrained.

Methods

The researchers evaluated three pathology FMs: Prov-GigaPath, UNI, and PRISM. Their approach involved:

  • Zero-Shot Whole-Slide Embeddings: Using the FMs' tile encoders to extract feature representations (embeddings) from individual image tiles of NMSC histology slides.
  • Tile Aggregation: Developing methods to combine these tile embeddings into a single "slide embedding" that represents the entire slide. They explored standard methods like Global Average Pooling (GAP) and attention-based Deep Multiple Instance Learning (ABMIL), as well as model-specific aggregators provided by the FMs.
  • Downstream Classification: Training lightweight classifiers (logistic regression, XGBoost, and a shallow multi-layer perceptron) on these slide embeddings to predict NMSC subtypes.
  • Few-Shot Annotation: Demonstrating a novel approach for annotating regions of interest (ROIs) using a small number of manually annotated "representative tiles" to guide the AI in identifying specific tissue types, even with limited expert input.

What's novel here is the application of these large, pre-trained FMs to the specific problem of NMSC diagnosis in a resource-limited context, and the exploration of both zero-shot and few-shot learning paradigms for this task.

Datasets

The study utilized a dataset of 2,130 hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) from 455 patients with NMSC. These samples were collected as part of the Bangladesh Vitamin E and Selenium Trial (BEST). The dataset primarily consisted of benign cases, followed by Bowen disease, basal-cell carcinoma (BCC), and invasive squamous-cell carcinoma (SCC). The quality of these slides was noted as generally lower than that typically found in accredited diagnostic pathology labs in the US.

Key Insights and Ideas

  • Foundation Models Show Promise: The study confirms that FMs can achieve high diagnostic performance for NMSC, significantly outperforming traditional deep learning baselines like ResNet18.
  • PRISM's Superiority: The PRISM vision-language model, with its distinct perceiver-based architecture and multimodal training, consistently achieved the highest diagnostic accuracy across various NMSC subtypes.
  • Attention Mechanisms Matter for Explainability: Visualizing attention maps from the aggregators revealed that PRISM's broader focus on tissue regions, including the dermis, likely aids in distinguishing between subtle NMSC subtypes like Bowen disease and invasive SCC. This aligns with dermatopathologist expertise.
  • Few-Shot Learning for Annotation: The few-shot annotation framework proved effective in identifying relevant tissue types with minimal expert guidance, demonstrating its utility in settings with limited annotation resources. It also showed an ability to implicitly handle image quality issues by distinguishing artifacts.
  • Resource-Limited Settings Benefit: The findings strongly suggest that FMs can be a valuable tool for improving NMSC diagnosis in regions with limited access to expert pathologists, potentially democratizing access to advanced diagnostic capabilities.

Key Results

  • The best-performing FM (PRISM) achieved a mean AUROC of 0.925 (P < 0.001) for NMSC subtype prediction, significantly outperforming a ResNet18 baseline (mean AUROC = 0.805, P < 0.001).
  • For UNI and Prov-GigaPath, ABMIL aggregation with an MLP classifier yielded the highest weighted mean AUROC (0.913 and 0.908, respectively).
  • PRISM's optimal configuration (its own aggregator with an MLP classifier) achieved the highest per-class performance, particularly in distinguishing benign and SCC cases.
  • Confusion matrices showed that while all models struggled somewhat with distinguishing benign from Bowen disease, PRISM demonstrated the least confusion for actual benign and SCC samples.
  • The few-shot annotation framework, using PRISM embeddings, correctly identified tiles of different cancer types in their respective slides, aligning well with expert annotations.

Limitations and Future Work

  • Dataset Generalizability: The primary dataset was derived from a single cohort (BEST) in Bangladesh, which may limit the generalizability of the findings to other populations.
  • Clinical Integration: Further research is needed to evaluate the practical integration of these AI tools into existing clinical workflows, especially considering the infrastructure requirements (e.g., slide scanners) in resource-limited settings.
  • Ethical Considerations: The authors acknowledge the need to address potential algorithmic bias, ensure transparency and explainability, and guarantee equitable access to AI technologies in healthcare.
  • Nuanced Classification Challenges: The study noted confusion between Bowen disease and benign cases, suggesting that finer-grained, tile-level supervision or models designed for enhanced interpretability might be needed for such nuanced distinctions.
  • Future Directions: The authors propose further validation in diverse populations, integration into clinical workflows, and continued exploration of FM architectures that effectively integrate multimodal data. They also emphasize the potential for FMs in prospective clinical trials for NMSC management.

Summary updated

Abstract

Pathology FMs offer a promising pathway to improve early and precise NMSC diagnosis, especially in resource-limited environments. These tools could also facilitate patient stratification and recruitment for prospective clinical trials aimed at improving NMSC management.
BibTeX
@inproceedings{ellis2025ai,
  author = {Spencer Ellis and Steven Song and Derek Reiman and Xuan Hui and Renyu Zhang and Hasan Shahriar and Maria Argos and Mohammed Kamal and Christopher R. Shea and Robert L. Grossman and Aly A. Khan and Habibul Ahsan},
  title = {{AI-assisted Diagnosis of Nonmelanoma Skin Cancer in Resource-Limited Settings}},
  booktitle = {Cancer},
  year = {2025},
  doi = {10.1158/1055-9965.epi-25-0132},
}
Figure