Denoising 7T Structural MRI with Conditional Generative Diffusion Models

7T MRI offers ultra-high resolution and improved sensitivity for iron deposition in neurodegenerative disorders, but commonly used acquisitions are long and hence challenging, especially for elderly subjects. Efficiently denoising a short acquisition to achieve the image quality of a longer acquisition would be of translational benefit. We introduce a conditional diffusion model derived from generative AI (a 7T Conditional Diffusion Model, 7TCDM) that was trained on native single-acquisition 2D reconstructions and referenced multi-repetition images to guide the denoising process and improve SNR and contrast. 7TCDM model was tested on 2D T2-weighted gradient-echo imaging from 19 participants, including healthy controls and individuals with mild cognitive impairment or Alzheimer's disease (AD). 7TCDM's performance was assessed using Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and comprehensive reader studies. Referencing the multi-repetition ground truth, 7TCDM improved the single-acquisition original image by 29.1% in MSE, 5.8% in PSNR, and 9.4% in SSIM, and outperformed convolutional neural network-based models in all metrics. Expert rater evaluations confirmed superior image quality, with significantly enhanced detail and contrast preservation in regions such as the hippocampi, white matter lesions, and small cortical veins. The model also demonstrated robust performance in both the concurrently acquired and publicly available 3D multi-echo gradient echo acquisitions, which the model was not trained on. The 7T Conditional Diffusion Model provides high-quality denoised images from shorter scans, increasing the feasibility of scanning patients in shorter times while preserving essential anatomical and pathological details.