High Resolution TSE Image Synthesis Using Denoising Diffusion Models Trained on 7T Image Pairs: Application for Hippocampal Subfield Analysis

Motivation: Severe motion artifacts on high resolution TSE images disrupt the image processing pipeline, leading to failed or erroneous segmentation outputs and forcing subject exclusion from studies. Goal(s): To address motion artifacts' impact on hippocampus subfield segmentation by developing an alternative solution. Approach: We implemented a denoising diffusion model for MR image translation, training it to synthesize TSE-like contrast from motion-resistant sequences (MPRAGE and MP2RAGE), and compared different image sampling strategies. Results: The synthesized images successfully enabled hippocampus subfield segmentation through the ASHS pipeline, demonstrating diffusion models' effectiveness in providing alternatives for motion-corrupted TSE images. Impact: This work introduces an alternative solution for salvaging motion-corrupted TSE images, potentially reducing patient exclusion rates and improving the statistical power of neuroimaging studies through diffusion model based image translation.