Predicting Cutaneous Squamous Cell Carcinoma Progression Risk from Whole Slide Images with Federated Learning

Jakub Zacharczuk, Juan Ignacio Pisula, Doris Helbig, Lucas Sancéré, Oana-Diana Persa, Corinna Bürger, Anne Fröhlich, Carina Lorenz, Sandra Bingmann, Dennis Niebel, Konstantin Drexler, Jennifer Landsberg, Roman Thomas, Johannes Brägelmann, Katarzyna Bozek

Keywords: Federated Learning, Cutaneous Squamous Cell Carcinoma

Abstract: Cutaneous squamous cell carcinoma (cSCC) is the second most common cancer globally. While surgical excision is typically successful, a significant proportion of patients experience disease progression leading to poor prognosis. Based on the fact that histopathological tumor features have been associated with increased risk of cSCC progression, we propose to predict this condition solely from Whole Slide Image (WSIs) scans of excised tumors. A major challenge in developing such predictive models is the fact that numerous clinical centers maintain patient cohorts that are often too small individually for robust deep learning (DL) applications. Here we use four small to medium-sized datasets from different clinical centers across Germany and demonstrate the feasibility of training federated DL models to predict cSCC progression. We compare various Federated Learning (FL) approaches, leveraging distributed datasets and developing center-specific models.

Submission Number: 66