Adversarial time-to-event modeling for pathological whole-slide images

Develop a conditional generative adversarial network–based time-to-event modeling approach that operates directly on histological whole-slide images for survival analysis, extending adversarial time-to-event modeling beyond tabular and CT imaging data to the whole-slide image domain.

Background

Adversarial time-to-event modeling using conditional GANs (e.g., DATE) has shown promise on clinical tabular data and has been extended to CT imaging, demonstrating advantages in predictive accuracy and robustness. However, these advances have not yet been translated to histological whole-slide images (WSIs), which present unique challenges such as extreme resolution and the need for multiple-instance learning paradigms.

Addressing this gap requires adapting the adversarial time-to-event framework to WSI data characteristics, enabling survival distribution estimation directly from WSIs in computational pathology.