Scanner Change in MRI: Domain Transfer for Harmonization of MR Image Contrasts

In magnetic resonance imaging (MRI), various factors such as scanner model and acquisition protocol affect the resulting image contrast. This limits the application of MRI as imaging modality in long-term or multicenter studies, where image consistency is required. Furthermore, the performance of deep-learning, which showed great success in various applications in the medical field, suffers in presence of a domain shift between training and test data introduced by different MR devices or protocols. As large datasets are costly to obtain, training of a new network for every domain is impossible. Therefore, image harmonization, aiming to reduce the domain gap between a source and a target image, would greatly improve the applicability of deep-learning networks in MRI. In this thesis, a supervised harmonization network is proposed to harmonize the contrast in MR images in a scanner change scenario, using a cohort of healthy subjects scanned in both MR scanners for network training. The proposed network architecture was able to generate images, that showed increased image similarity after harmonization of the image contrasts between source and target domain. Furthermore, the volumetric difference in segmented structures introduced by the scanner change was successfully reduced in the brainstem segments mesencephalon (-0.462% ± 0.224) and pons (-0.809% ± 0.283). Additionally, volumetric bias in longitudinal data of Multiple Sclerosis patients could be reduced in mesencephalon (-0.286 % ± 0.198) and pons (-0.361 % ± 0.160) by harmonizing image contrasts across scanners.