Improving segmentation reliability of multi-scanner brain images using a generative adversarial network

Background: Magnetic resonance (MR) images generated by different scanners generally have inconsistent contrast properties, making it difficult to perform a combined quantitative analysis of images from a range of scanners. In this study, we aimed to develop an automatic brain image segmentation model to provide a more reliable analysis of MR images taken with different scanners. Methods: The spatially localized atlas network tiles-27 (SLANT-27) deep learning model was used to train the automatic segmentation module, based on a multi-center dataset of 1,917 three-dimensional (3D) T1-weighted MR images. Subsequently, a framework called Qbrain, consisting of a new generative adversarial network (GAN) image transfer module and the SLANT-27 segmentation module, was developed. Another 3D T1-weighted MRI interscan dataset of 48 participants who were scanned in 3 MRI scanners (1.5T Siemens Avanto, 3T Siemens Trio Tim, and 3T Philips Ingenia) on the same day was used to train and test the Qbrain model. Volumetric T1-weighted images were processed with Qbrain, SLANT-27, and FreeSurfer (FS). The automatic segmentation reliability across the scanners was assessed using test-retest variability (TRV). Results: . 3T Philips scanners was reduced by 1.57%, 2.01%, and 0.56%, respectively. In addition, the segmentation model improved intra-scanner variability (0.9-1.67%) compared with that of FS (2.47-4.32%). Conclusions: The newly developed QBrain method combined with GAN image transfer module and a SLANT-27 segmentation module was shown to improve the reliability of whole-brain automatic structural segmentation results across multiple scanners, thus representing a suitable alternative quantitative method of comparative brain tissue analysis for individual patients.