Using GANs to Visualise Class-Specific Features in Digital Histopathology Images

Catherine Little, Richard Gault, Stephanie G. Craig, Jacqueline A. James

Abstract: Oropharyngeal squamous cell carcinoma (OPSCC) patients have an increased likelihood of testing positive for human papillomavirus (HPV). Additionally, people with OPSCC and HPV have a better prognosis than those without HPV. Improving our understanding of the relationship between HPV and OPSCC may improve diagnosis and treatment for OPSCC patients. There are known molecular and clinical entities that differentiate the two cohorts. Whilst it is beneficial to be able to computationally classify the HPV status of patients from histological images, it is important to understand which features a computational model is utilising to make such a prediction. Generative Adversarial Networks (GANs) can be trained to generate hematoxylin and eosin (H&E) stained biopsy samples of cancerous tissues with a specified HPV status, it may be possible to use its understanding of the image domain to gain insight into the connection between OPSCC and HPV. This paper proposes a novel model, PathologyAC-GAN, which combines the class labeling aspect of Auxiliary Classier GANs with the high image fidelity achieved by PathologyGAN to generate images of a specified HPV status. Through qualitative and quantitative assessment, this paper finds that PathologyAC-GAN can learn distinguishing morphological characteristics that distinguish the HPV status of OPSCC patients creating biological insights.

External IDs:dblp:conf/aiih/LittleGCJ24