Ultra-high-field 7T MRI provides superior detail for diagnosing Alzheimer's disease (AD) and mild cognitive impairment (MCI), but limited accessibility restricts clinical use. To bridge this gap, we developed the self-supervised masked attention-based refinement transformer network (SMaRT-Net), a feedback-driven superresolution framework that synthesizes 7T-equivalent MRI (7T*) from standard 3T MRI. The model combines a Barlow Twins-guided latent diffusion model for initial superresolution with a masked attention transformer that refines images through classification feedback, enabling generation of high-resolution brain imaging optimized for diagnostic use. Training was performed in two stages: a paired 3T-7T dataset from healthy adults was used for initial superresolution learning, and the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, including AD, MCI, and cognitively normal (CN) subjects, was used for refinement and classification. Prognostic evaluation was conducted in distinguishing stable from converting MCI (sMCI vs. cMCI). Image quality was assessed with BRISQUE and NIQE, while diagnostic performance was evaluated using accuracy and AUROC. The generated 7T* MRIs demonstrated superior quality compared to original 3T scans (BRISQUE: 41.6 vs. 43.6; NIQE: 6.2 vs. 6.4). Diagnostic performance improved consistently: AD vs. CN accuracy increased from 0.900 to 0.946, MCI vs. CN from 0.881 to 0.926, and AD vs. MCI from 0.900 to 0.938. Prognostic accuracy for predicting MCI conversion also rose from 0.680 to 0.800. These results highlight the synergistic benefit of combining superresolution with classification-guided refinement, demonstrating the potential of SMaRT-Net as a scalable tool for improving early diagnosis and prognosis of neurodegenerative disease using routine 3T MRI.
Yoon et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: