Optimal Mass Transport for Robust Texture Analysis

Abstract

The emerging field of radiomics, which consists of transforming standard-of-care images to quantifiable scalar statistics, endeavors to reveal the information hidden in these macroscopic images. This field of research has found different applications ranging from phenotyping and tumor classification to outcome prediction and treatment planning. Texture analysis, which often consists of reducing spatial texture matrices to summary scalar features, has been shown to be important in many of the latter applications. However, as pointed out in many studies, some of the derived texture statistics are strongly correlated and tend to contribute redundant information; and are also sensitive to the parameters used in their computation, e.g., the number of gray intensity levels. In the present study, we propose first to consider texture matrices, with an emphasis on gray-level co-occurrence matrix (GLCM), as a non-parametric multivariate objects. The proposed modeling approach avoids evaluating redundant and strongly correlated features and also prevents the feature processing steps. Then, via the Wasserstein distance from optimal mass transport theory, we propose to compare these spatial objects to identify computerized tomography slices with dental artifacts in head and neck cancer. We demonstrate the robustness of the proposed classification approach with respect to the GLCM extraction parameters and the size of the training set. Comparisons with the random forest classifier, which is constructed on scalar texture features, demonstrates the efficiency and robustness of the proposed algorithm. ¶

Supervised Image Classification Algorithm Using Representative Spatial Texture Features: Application to COVID-19 Diagnosis Using CT Images

“…In radiomics research field, texture analysis consists usually of generating features based on summary statistics from high-dimensional feature matrices such as GLCM and RLM among others. This reduction step may lead to a loss of the spatial information that is inherent in these texture matrices, which has been discussed in previous studies [13], [14], [15].…”

“…This reduction step may lead to a loss of the spatial information that is inherent in these texture matrices, which has been discussed in previous studies [13], [14], [15]. The present paper is a continuation of a previous work highlighting the importance of spatial (multidimensional) texture features for robust medical image classification [15]. It is based on the assumption that there exist representative samples, which we refer to as references as well, i.e.,”good or bad” samples that represent each class/data cohort in a given particular classification/regression problem.…”