Does Harmonization Impact Radiomics Features in Longitudinal MRI of Glioblastoma? A Comparison between ComBat and longComBat

Motivation: MRI-derived radiomics features are highly affected by acquisition and reconstruction protocols. This may impact the performance of Machine Learning models developed for managing disease progression, such as in Glioblastoma. Goal(s): To compare longitudinal ComBat (longComBat) with ComBat for harmonizing radiomics features in a longitudinal Glioblastoma MRI dataset. Approach: Both methods (ComBat and longComBat) were applied to harmonize radiomics features from multi-scanner longitudinal MRI of Glioblastoma patients. The resulting features were compared using PCA, and the performance of models trained to classify images as pre- or post-operative was assessed. Results: ComBat was best at removing batch effects, while longComBat yielded the best classification performance. Impact: Technical variations greatly influence MRI-derived radiomics features. We harmonized radiomics features from longitudinal MRI of Glioblastoma using both ComBat and longComBat. The latter improved Machine Learning classification into pre- and post-operative scans.