Clinical validity of counterfactual whole slide images

Establish whether gigapixel-scale counterfactual whole slide images (WSIs) generated for computational pathology are realistic and clinically valid, demonstrating that the produced images are biologically plausible and suitable for clinical use.

Background

The paper motivates the need for realistic counterfactual WSIs to paper "what if" scenarios, such as modifying tumor size or immune cell distributions to understand model behavior. Creating such counterfactuals at the gigapixel scale requires precise segmentation and robust inpainting, and while generative AI shows promise, its application to WSIs has not yet been clinically validated.

HIPPO is proposed as a patch-level intervention framework that sidesteps direct image synthesis by adding or removing patches, but the broader question of whether fully generated counterfactual WSIs can meet clinical standards remains explicitly unproven.