Reconstruction of Super-Resolution T1 Maps Using Efficient Residual Denoising Diffusion Probabilistic Models at 7T Field

Motivation: Achieving high-resolution T1 mapping requires extended scan time due to substantially prolonged tissue T1 times at ultra-field, leaving the data quality susceptible to patient motion and other interferences. Goal(s): Develop an efficient DDPM DL model using that produces high-resolution T1 maps from low-resolution inputs with minimal sampling steps, enhancing clinical feasibility. Approach: The proposed method combines residual learning with a novel DDPM architecture, reducing the required sampling steps from 1000 to four. This model was trained and tested on institutional 7T MRI data. Results: The model significantly reduced inference time by over 240 times, providing high-resolution T1 maps with improved structural detail. Impact: The proposed model can reduce the scan time required for generating high-resolution T1 maps within a clinically acceptable time. Its capacity to produce high-quality brain images with reduced artifacts may improve diagnosis and accelerate advancements in neuroimaging research.