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

To evaluate the impact of image harmonization on outcome prediction models using radiomics.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) and a generative adversarial network (GAN)-based method (H). 88 radiomics features were extracted on H, Hand original (H) 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.More than 50% of the features (49/88) were statistically modified by the harmonization with Hand 55 with H(adjusted-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, 62% for Hand 71% for H). Both image-based harmonization methods allowed to split patients into two groups with significantly different survival (<0.05). With the Himages, we were also able to build and validate a model using only features impacted by the harmonization (median survivals of 189 versus 437 days,= 0.006)Data harmonization in a multi-institutional cohort allows to recover the predictive value of some radiomics features that was lost due to differences in the image properties across centers. In terms of ability to build survival prediction models in the BRATS dataset, the loss of power from impacted histogram and heterogeneity features was compensated by the selection of additional shape features. The harmonization using a GAN-based approach outperformed the histogram matching technique, supporting the interest for the development of new advanced harmonization techniques for radiomic analysis purposes.