Diffusion-Guided Approaches to Medical Imaging: Enhancing MRI Reconstruction and Cortical Surface Registration

Abstract

Medical imaging is crucial in healthcare, enabling non-invasive visualization of internal structures and aiding in disease diagnosis and treatment. Techniques such as MRI, CT, PET, and ultrasound are integral for both clinical practice and research. However, image reconstruction and registration challenges persist, particularly in MRI, where noise, artifacts, and prolonged acquisition times can degrade image quality. Traditional deep learning methods, including VAEs, UNets, and GANs, have limitations such as high computational costs and training instability. Diffusion-guided approaches have emerged as promising solutions, leveraging diffusion models to address high-dimensional and complex medical imaging data. These methods offer improved image quality and registration accuracy by capturing intricate details and reducing computational burdens. In neuroimaging, cortical surface registration faces difficulties due to anatomical variability and computational inefficiency. Existing deep learning methods have not fully utilized relational information between data distributions, leading to suboptimal results. This research proposes a novel methodology inspired by DiffuseMorph, integrating spherical diffusion models and spectral attention mechanisms to enhance cortical surface registration. Additionally, a new approach, Consistent Direct Diffusion Bridge with Injection (CDDBI), is introduced for MRI reconstruction. CDDBI utilizes k-space information and consistency-imposing gradient steps, improving image quality and reducing reconstruction time. Overall, diffusion-guided approaches demonstrate significant potential in advancing medical imaging technologies. By addressing the limitations of current methods, these approaches enhance diagnostic accuracy, treatment planning, and disease understanding, contributing to improved patient care and medical research outcomes.