A Bayesian Nonparametric model for textural pattern heterogeneity

TL;DR

This paper introduces a Bayesian nonparametric framework for analyzing GLCM-based texture patterns in medical images, improving reproducibility and capturing heterogeneity more effectively than traditional methods.

Contribution

It develops a novel Bayesian probabilistic model that accounts for skewness, zero-inflation, and spatial autocorrelation in GLCM data for unsupervised clustering of radiological images.

Findings

Successful clustering of adrenal lesion images into meaningful subtypes.

Clusters showed correspondence with pathological diagnoses.

Model outperformed existing machine-learning approaches in simulations.

Abstract

Cancer radiomics is an emerging discipline promising to elucidate lesion phenotypes and tumor heterogeneity through patterns of enhancement, texture, morphology, and shape. The prevailing technique for image texture analysis relies on the construction and synthesis of Gray-Level Co-occurrence Matrices (GLCM). Practice currently reduces the structured count data of a GLCM to reductive and redundant summary statistics for which analysis requires variable selection and multiple comparisons for each application, thus limiting reproducibility. In this article, we develop a Bayesian multivariate probabilistic framework for the analysis and unsupervised clustering of a sample of GLCM objects. By appropriately accounting for skewness and zero-inflation of the observed counts and simultaneously adjusting for existing spatial autocorrelation at nearby cells, the methodology facilitates estimation of texture pattern distributions within the GLCM lattice itself. The techniques are applied to cluster images of adrenal lesions obtained from CT scans with and without administration of contrast. We further assess whether the resultant subtypes are clinically oriented by investigating their correspondence with pathological diagnoses. Additionally, we compare performance to a class of machine-learning approaches currently used in cancer radiomics with simulation studies.