Optimising diffusion models for histopathology image synthesis

Oropharyngeal Squamous Cell Carcinoma (OPSCC) is a sub-type of head and neck cancer linked to human papillomavirus infection (HPV). HPV-positive OPSCC patients have an improved prognosis compared to HPV-negative OPSCC patients however, the reasoning for this is unknown. Visualising the clinical and molecular differences in HPV status would be highly interpretable and could aid our understanding of the impact these distinguishing features have on patient prognosis. A generative model trained to de-lineate features of HPV status provides both a synthetic visualisation of HPV-related OPSCC and a classification of HPV status. Conditional diffusion models (CDMs) have been shown to produce state-of-the-art (SOTA) quality and fidelity in the image synthesis domain. Furthermore, they can generate representative Haematoxylin and Eosin(H&E) stained histopathology images of cancerous tissue. This paper proposes two novel weighting schemes, one of which is designed to prioritise spatial features during training which enables the model to learn important pathological markers associated with HPV-related OPSCC tissue. Through experimental analysis of histological data, we demonstrate that our proposed approach improves the performance of CDMs and provides insightful, interpretable features that aid our understanding of HPV-related OPSCC.