Abstract 1141: Improved diagnosis of non-melanoma skin cancer in resource-limited settings

Abstract Background: In resource-limited settings, cancer diagnosis is often challenging due to a shortage of expert pathologists. Early and precise diagnosis is vital to enhancing treatment outcomes and reducing morbidity of patients. This issue is particularly prevalent in regions like Bangladesh, where high levels of arsenic exposure increase the risk of non-melanoma skin cancer (NMSC). Here, we assess the effectiveness of general-purpose pathology foundation models (FMs) for diagnosis of NMSC. In particular, we aim to determine if FMs have value in resource-constrained environments, such that encodings from scanned whole slide images can be used with simple classification models that can run efficiently in resource-limited settings. Method: Using 5-fold cross validation, we evaluated three pathology foundation models (UNI, PRISM, and Prov-GigaPath) as well as a ResNet18 baseline model using de-identified NMSC data from the Bangladesh Vitamin E and Selenium Trial (BEST) by the Institute for Population and Precision Health at the University of Chicago. This data contained 2,130 hematoxylin and eosin (H&E)-stained whole slide images from 553 suspected NMSC biopsy samples from 455 participants. The slides included normal tissue (n=706), Bowen’s disease (n=638), basal cell carcinoma (n=575), and invasive squamous cell carcinoma (n=211). In addition to comparing the three FMs for generating tile embeddings in a zero-shot framework, we also evaluated tile aggregation methods, including global average pooling (GAP), attention-based multi-instance learning (ABMIL), and methods specific to respective FMs. Lastly, we compared the performance of logistic regression, XGBoost, and shallow multilayer perceptron neural networks (MLP) for the prediction of cancer subtype from the slide embedding. Results: All three FMs significantly outperformed ResNet18 (mean AUROC=0.805; p<0.001). We found that the overall best model used the PRISM tile embeddings aggregated using PRISM’s intrinsic Perceiver network to train an MLP model to predict NMSC subtype (mean AUROC=0.925; p<0.001). Within the other FMs specifically, we found that using ABMIL to aggregate tile embeddings to train an MLP model was optimal for both UNI (mean AUROC=0.913; p<0.001) and Prov-GigaPath (mean AUROC=0.908, p<0.001). We also found the simplest method with logistic regression of GAP aggregated embeddings was able to attain reasonable results for PRISM (mean AUROC=0.882), UNI (mean AUROC=0.865), and Prov-GigaPath (mean AUROC=0.855). Conclusion: Our study highlights the importance of innovation in confronting public health challenges and exhibits a real-world potential for machine learning aided cancer diagnosis. We demonstrate how leveraging whole slide embeddings from pre-trained foundation models can provide considerable potential for the improvement of treatment outcomes and patient survival rates, especially in resource-limited settings. Citation Format: Spencer Ellis, Steven Song, Derek Reiman, Xuan Hui, Renyu Zhang, Mohammad H. Shahriar, Mohammed Kamal, Christopher R. Shea, Robert L. Grossman, Aly A. Khan, Habibul Ahsan. Improved diagnosis of non-melanoma skin cancer in resource-limited settings [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1141.