A0738
Title: Spatial nested mixture models for MALDI-MSI image segmentation
Authors: Francesco Denti - University of Padua (Italy) [presenting]
Abstract: Mass spectrometry imaging is emerging as a valuable tool for measuring in-situ cancer biomarkers, as it allows the detection of the critical biological traits that would be overlooked with a simple visual morphological assessment of a sample. This technique measures the abundance of several specific molecules over multiple locations of a biological sample. The analysis of these complex data structures calls for developing tailored statistical methods. Over the last few years, the Bayesian community has dedicated increased attention to mixture priors inducing nested random partitions. Employing models for nested, separate exchangeable data is proposed to estimate a biclustering solution, i.e., cluster locations characterized by similar abundance profiles. This way, molecules can be simultaneously detected with similar expressions within clusters of pixels. Moreover, a hidden Markov random field prior is employed to perform appropriate image segmentations. To address the large dimensionality of these datasets and the need for timely results, an efficient coordinate ascent variational inference algorithm is applied that dramatically scales the model's applicability. The estimated biclustering structure is showcased, allowing the detection of meaningful image segmentation and patterns of activated molecules.
