BTS-Net: Barlow twins-based superresolution for 7T human brain MRI

This study aimed to develop and validate a Barlow Twins-based superresolution diffusion network (BTS-Net) for 7T human brain magnetic resonance imaging (MRI) superresolution (SR). A paired 3T-7T human brain MRI database was constructed from 50 healthy adult participants (26 females; 52.0%; age range: 27-68 years; median age: 38 years), with anonymized scans. To develop BTS-Net, we employed Barlow Twins, a self-supervised learning (SSL) method, in a latent diffusion model (LDM) to enhance feature representation learning for SR from 3T MRI to 7T-like (BTS-7T) MRI. The image quality of the 3T, 7T, LDM-7T, and BTS-7T MRI was evaluated using the peak signal-to-noise ratio, structural similarity index measure, and normalized root mean squared error. The paired t-test was used to evaluate the mean difference between each imaging group. The three-dimensional structural fidelity was evaluated in 14 anatomical regions (the bilateral thalamus, caudate, putamen, globus pallidus, hippocampus, amygdala, and nucleus accumbens) using voxel-based morphometry. An external dataset consisting of 10 healthy participants was used to validate BTS-Net by performing identical SR, image quality, and volumetric analyses. In both the in-house and external validation datasets, BTS-7T MRI exhibited superior image quality across all three metrics compared to 3T MRI. Ground-truth-based error maps showed that BTS-7T images displayed qualitatively improved anatomical fidelity compared to 3T images. There were no statistically significant differences in volumetry between 10 (in-house) and 11 (validation) of the 14 anatomical regions. The right hippocampus, putamen, and amygdala volumes showed significantly higher agreement in BTS-7T images of the in-house dataset; whereas the bilateral putamen, right thalamus and amygdala volumes showed significantly higher agreement in the validation dataset. This study highlights the potential of BTS-Net to enhance both qualitative visualization and quantitative analysis of 3T brain MRI, with possible applicability to early neurodegenerative conditions characterized by subtle morphological changes. Therefore, additional research using larger patient datasets could aid the possible adoption of this technology in clinical settings.