everyone
since 12 Apr 2025">EveryoneRevisionsBibTeXCC BY 4.0
Recent foundation models for pathology—such as UNI, CONCH, RudolfV, Prov-GigaPath, Atlas, and Virchow2 —have demonstrated impressive performance, but are typically trained on images captured by costly whole-slide imaging scanners. In contrast, many hospitals in developing countries still rely on optical microscopes and low-cost cameras or smartphones for image acquisition. To bridge the gap, we demonstrate that pathology foundation model pre-trained on whole-slide images (WSIs) can be fine-tuned on smartphone-captured cytopathology images for applications in low-resource settings. We used over 3,000 smartphone-captured cytology images from Tanzania and Vietnam for Virchow2 foundation model fine-tuning and testing. Our approach not only resulted in high classification performance (98.17% AUC, 92.93% accuracy, 94.13% F1-score) but also enhanced interpretability through principle component visualization of the embedding space, thereby fostering clinical trust in resource-constrained settings.