ComBat Harmonization for MRI Radiomics: Impact on Nonbinary Tissue Classification by Machine Learning

Abstract

Objectives: The aims of this study were to determine whether ComBat harmonization improves multiclass radiomics-based tissue classification in technically heterogeneous MRI data sets and to compare the performances of 2 ComBat variants.

Materials and methods: One hundred patients who had undergone T1-weighted 3D gradient echo Dixon MRI (2 scanners/vendors; 50 patients each) were retrospectively included. Volumes of interest (2.5 cm 3 ) were placed in 3 disease-free tissues with visually similar appearance on T1 Dixon water images: liver, spleen, and paraspinal muscle. Gray-level histogram (GLH), gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level size-zone matrix (GLSZM) radiomic features were extracted. Tissue classification was performed on pooled data from the 2 centers (1) without harmonization, (2) after ComBat harmonization with empirical Bayes estimation (ComBat-B), and (3) after ComBat harmonization without empirical Bayes estimation (ComBat-NB). Linear discriminant analysis with leave-one-out cross-validation was used to distinguish among the 3 tissue types, using all available radiomic features as input. In addition, a multilayer perceptron neural network with a random 70%:30% split into training and test data sets was used for the same task, but separately for each radiomic feature category.

Results: Linear discriminant analysis-based mean tissue classification accuracies were 52.3% for unharmonized, 66.3% for ComBat-B harmonized, and 92.7% for ComBat-NB harmonized data. For multilayer perceptron neural network, mean classification accuracies for unharmonized, ComBat-B-harmonized, and ComBat-NB-harmonized test data were as follows: 46.8%, 55.1%, and 57.5% for GLH; 42.0%, 65.3%, and 71.0% for GLCM; 45.3%, 78.3%, and 78.0% for GLRLM; and 48.1%, 81.1%, and 89.4% for GLSZM. Accuracies were significantly higher for both ComBat-B- and ComBat-NB-harmonized data than for unharmonized data for all feature categories (at P = 0.005, respectively). For GLCM ( P = 0.001) and GLSZM ( P = 0.005), ComBat-NB harmonization provided slightly higher accuracies than ComBat-B harmonization.

Conclusions: ComBat harmonization may be useful for multicenter MRI radiomics studies with nonbinary classification tasks. The degree of improvement by ComBat may vary among radiomic feature categories, among classifiers, and among ComBat variants.