ComBat harmonization for multicenter MRI based radiomics features

Radiomics, the high-throughput feature extraction from medical images, has gained momentum in radiology and the noninvasive assessment of the tumor, treatment response and therapy monitoring. A significant effort has been put to incorporate radiomics in the clinical setting, towards developing radiomics-based machine learning models from a large amount of images across multiple centers. However, multicenter radiomics has been proven to be sensitive to different scanner models, acquisition protocols and reconstruction settings, deriving imaging features of increased variability that hamper robustness and generalizability of the models. Recent literature has stressed the importance of utilizing harmonization techniques to compensate for multicenter effects. These include traditional image-based techniques and harmonization at the feature level using ComBat. The latter has been shown to have increased performance in several radiomics studies. To this end, this study investigated the capability of ComBat and M-ComBat to eliminate the “center-effect” of multicenter images from an heterogeneous TCIA public dataset. Both methods were tested on 102 patients with WHO grade IV glioblastomas and 64 patients with WHO grade II and III glioma and performance was assessed using a machine learning pipeline comprising combinations of different feature selection, data resampling and machine learning models. A significant increase in the performance was obtained using M-ComBat, yielding a Matthew Correlation Coefficient of $0.91 \pm 0.05$ compared to $0.76 \pm 0.08$ that was obtained with the absence of any harmonization prior to model development.