The purpose of this study is to recover 3D-T1 weighted brain MRI scans from Multiple Sclerosis (MS) patients to reproduce a homologous representation of their lesion-free brain.

In population-level studies, the comparison of MS brain MRIs with their synthetic lesion-free counterpart may highlight disease-specific variability which may provide insights into MS pathophysiology.

A generative adversarial network (GAN) based on StyleGAN2 was trained to generate synthetic healthy brain scans from a 512-dimensional latent variable. Then, each MS brain was projected to this latent-space to recover the latent variable that generates the synthetic ‘healthy’ brain which is the most similar to the MS brain.

For model training, 3500 3D T1-weighted brain MRIs from healthy adults were pooled across the HCP, OASIS2 and IXI datasets. For evaluation, scans from 2398 MS patients from ADVANCE (NCT00906399) and ASCEND (NCT01416181) trials were used. All scans were skull-stripped, resampled to 2-mm isotropic voxel spacing, and intensity-normalized via min-max mapping to the range [-1, 1].

Model evaluation included visual inspection and quantitative measures: Mean Squared Error (MSE), Structural Similarity Metric (SSIM) and Peak Signal to Noise Ratio (PSNR) between the original MS brain scans and their synthetic lesion-free counterpart, excluding lesion mask areas.

We applied our method to 3D T1-weighted MRIs of MS patients to recover lesion-free brains with an isotropic resolution of 2mm. The generated lesion-free brain scans showed no MS lesions and displayed great similarity to the diseased brain outside of the MS lesion areas, as highlighted by the following metrics: SSIM = 0.99, PSNR = 34.5, MSE = 1.49e-3.

Our method reconstructed realistic lesion-free representations of 3D T1-weighted brain MRI scans of MS patients. This method could be generalizable to other neurological diseases.