JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

Explainable, federated deep learning model predicts disease progression risk of cutaneous squamous cell carcinoma

  • 0Institute for Biomedical Informatics, Faculty of Medicine and University Hospital Cologne, University of Cologne, Köln, Germany.
Npj Precision Oncology +

|

June 28, 2025

|

June 28, 2025

View abstract on PubMed

Summary

This summary is machine-generated.

Predicting cutaneous squamous cell carcinoma (cSCC) progression is improved using a transformer-based deep learning model. This AI approach analyzes histopathology slides, enhancing risk stratification for personalized cancer care and secondary prevention.

Area Of Science

  • Computational pathology
  • Artificial intelligence in oncology
  • Digital histopathology

Background

  • Accurate prediction of cancer patient disease progression is crucial for personalized medicine and secondary prevention.
  • Current risk stratification systems for cutaneous squamous cell carcinoma (cSCC) have limitations in predictive accuracy.
  • Deep learning models offer potential for improved patient risk prediction, but their interpretability is often limited.

Purpose Of The Study

  • To develop and validate a transformer-based deep learning model for predicting cSCC progression using histopathology slides.
  • To enhance model generalizability and address privacy concerns through federated learning.
  • To investigate the interpretability of the deep learning model to identify predictive histopathological features.

Main Methods

  • Development of a transformer-based deep learning model utilizing diagnostic histopathology slides of cSCC patients.
  • Initial model training and validation on a held-out test set.
  • Federated learning approach implemented across three clinical centers to improve generalizability and privacy.
  • Interpretability analysis to identify key spatial and morphological features associated with disease progression.

Main Results

  • The initial model achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.92 on the test set and an average AUROC of 0.65 on external validation cohorts.
  • The federated learning approach resulted in an AUROC of 0.82 across all cohorts.
  • Image-based risk scores demonstrated significant association with progression, with hazard ratios up to 7.42 (p < 0.01) in multivariable analyses.
  • Interpretability analysis highlighted tumor boundary features and tissue heterogeneity as predictive of cSCC progression.

Conclusions

  • A transformer-based deep learning model trained on routine histopathology slides can effectively predict cSCC progression.
  • Federated learning enhances model performance and generalizability while preserving patient privacy.
  • The model provides biological insights into cSCC progression, identifying key histopathological features.
  • This approach facilitates cross-clinical center deployment for improved secondary prevention and understanding of cSCC.