Clinical Brain MRI Super-Resolution with 2D Slice-Wise Diffusion Model

Runqi Wang, Zehong Cao, Yichu He, Jiameng Liu, Feng Shi, Dinggang Shen

Abstract: Magnetic resonance imaging (MRI) plays a vital role in brain imaging, offering exceptional soft tissue contrast without the use of ionizing radiation, ensuring safe and effective medical diagnosis. In clinic settings, 2D acquisitions are preferred by physicians due to fewer slices, large spacing, and high in-plane resolution, balancing spatial resolution, signal-to-noise ratio (SNR), 0 and acquisition time. However, these MR images may lack through-plane resolution, which may hinder lesion detection, tissue segmentation, accurate volumetric measurements, and cortical reconstruction. Most existing deep learning methods are built with purely synthetic data by collecting only high-resolution images, creating a gap between synthetic data and real-world paired data. To address these issues, we propose a slice-wise framework using a diffusion model for inter-slice super-resolution of brain MR images: 1) Employ a real-world coarse super-resolution model for initial prediction; 2) Use a score-based diffusion model for detailed iterative refinement; 3) Leverage total variation (TV) penalty with a plug-and-play (PnP) optimization module for enhanced consistency. We validate our method on over 450 real paired cases, demonstrating that our method could generate realistic images with satisfactory 3D consistency and significantly reduce over-smooth problems, thereby improving current data quality. This 2D slice-wise diffusion model also provides an effective solution for improving the quality of brain MRI images in real-world scenarios.