Title: Bayesian scalar-on-image regression for automatically detected of regions of interest
Authors: Sara Wade - University of Edinburgh (United Kingdom) [presenting]
Abstract: In biomedical studies, vast amounts of imaging, biological and clinical data are increasingly collected to improve understanding of diseases or conditions. In this setting, we develop scalable Bayesian scalar-on-image regression models that allow for the integration of such data. Scalar-on-image regression models utilise the entire imaging data, making it is possible to capture the complex pattern of changes associated with the disease and improve accuracy; however, the massive dimension of the images, which is often in the millions, combined with the relatively small sample size, that at best is usually in the hundreds, pose serious challenges. We propose a novel class of Bayesian nonparametric scalar-on-image regression models based on the Potts-product partition model that groups together voxels into spatially coherent clusters used as features in the regression model. This greatly eases the computational issues associated with the high-dimensional and highly-correlated inputs and allows for interpretable and reliable features that are automatically defined as the most discriminative. The posterior inference is based on a generalized Swendsen-Wang sampler, allowing efficient split-merge moves that take advantage of the spatial structure. Applications focus on early diagnosis and prognosis of Alzheimer's disease, irreversible brain disease and major international public health concerns.