comBat versus cycleGAN for multi-center MR images harmonization

Stenzel Cackowski, Emmanuel Luc Barbier, Michel Dojat, Thomas Christen

Keywords: Brain, MRI, harmonization, deep-learning, radiomic features, classification

Abstract:

Pooling Magnetic Resonance Imaging (MRI) scans from different sites is difficult due to uncontrolled variations introduced by the use of different acquisition protocols or scanners. Image harmonization is a way to remove site-specific bias while preserving the intrinsic image properties. While multiple harmonization techniques exist, it is yet difficult to evaluate their efficiencies in specific applications.
In this paper, we propose to carry out five experiments, performed on synthetic and real data, in order to be able to benchmark two different existing, but never compared in the literature, harmonization approaches: ComBat and CycleGAN. We focus on T1-weighted MR images (one of the most widely used MR images) and propose to investigate the effects of each harmonization approach using radiomic features to extract image properties and Support Vector Machine (SVM) for classification. We show that both methods perform well for removing various types of noises while preserving manually added synthetic lesions, but also for removing site effects on data coming from 2 different sites while preserving biological information. Moreover, while each harmonization method improves autism classification,we show that CycleGAN outperforms ComBat in terms of accuracy.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: both

Primary Subject Area: Transfer Learning and Domain Adaptation

Secondary Subject Area: Validation Study