Evaluation of conventional and deep learning based image harmonization methods in radiomics studies

Objective. To evaluate the impact of image harmonization on outcome prediction models using radiomics. Approach. 234 patients from the Brain Tumor Image Segmentation Benchmark (BRATS) dataset with T1 MRI were enrolled in this study. Images were harmonized to a reference image using histogram matching (H-HM) and a generative adversarial network (GAN)-based method (H-GAN). 88 radiomics features were extracted on H-HM, H-GAN and original (H-NONE) images. Wilcoxon paired test was used to identify features significantly impacted by the harmonization protocol used. Radiomic prediction models were built using feature selection with the Least Absolute Shrinkage and Selection Operator (LASSO) and Kaplan-Meier analysis. Main results. More than 50% of the features (49/88) were statistically modified by the harmonization with H-HM and 55 with H-GAN (adjusted p-value < 0.05). The contribution of histogram and texture features selected by the LASSO, in comparison to shape features that were not impacted by harmonization, was higher in harmonized datasets (47% for H-none, 62% for H-HM and 71% for H-GAN). Both image-based harmonization methods allowed to split patients into two groups with significantly different survival (p