ULTRA LOW-FIELD TO HIGH-FIELD MRI TRANSLATION USING ADVERSARIAL DIFFUSION

Ultra Low-field Magnetic Resonance Imaging (MRI) scanners can potentially make a substantial impact in the field of medical imaging and radiology due to their cost-effectiveness, potential for portability and utility in an environment where the resource is in shortage. However, low-field MRI encounters challenges such as a low signal-to-noise ratio which results in lower-quality images. In this study, we introduce a novel image translation technique that relies on an adversarial diffusion-based deep learning approach to generate high-field MRI images from ultra low-field MR images. We have integrated a non-diffusive attention-guided module to enhance areas recognized as critical high-level features using self-attention maps from the diffusion process. To evaluate our approach, we use paired datasets consisting of different MRI sequences from both 64mT ultra low-field and 3T high-field scanners. We compare the performance of our method against state-of-the-art GAN and diffusion-based models, demonstrating its superior performance both quantitatively and qualitatively.