Fig. 15
From: Slideflow: deep learning for digital histopathology with real-time whole-slide visualization
Example of model explanations using saliency maps and generative adversarial networks. a Various gradient-based pixel attribution methods with APIs available in Slideflow were used to generate saliency maps for an example HPV-negative image. b A class-conditional GAN based on the StyleGAN2 architecture was trained using Slideflow to generate synthetic histology images belonging to HPV-negative and HPV-positive negative classes, visually highlighting morphologic differences between classes