MRI Radiomic Features Harmonization: A Multi-Center Phantom Study

In cancer imaging, radiomic features are a vital tool for diagnosing, quantifying intratumor heterogeneity, and predicting response to therapy and prognosis. However, the lack of a robust evaluation makes applying them across multi-center studies and larger patient populations complex. The main goal of this study is to use an MRI phantom to examine the reproducibility of MRI radiomic features that represent feature variability with ComBat harmonization in a multi-center examination with various acquisition settings and image pre-processing steps. An MRI phantom was used to acquire images using four scanners. Manual segmentation was used for 19 phantom compartments. Wavelet, bin discretization, and Laplacian of Gaussian were employed to pre-process each image. Ninety-two radiomic features were extracted. ComBat harmonization was applied to all feature sets. The Kruskal-(KW) Wallis’s and the Intraclass Correlation Coefficient (ICC) tests were utilized to indicate radiomic features’ variability. The number of non-significant p-values ranged from 29-74 and 0-5 features after and before harmonization over different image pre-processing. The results showed that the radiomic features are influenced most by various scanners. The impact of ComBat harmonization on the reproducibility of MRI radiomic features is considerable.