Residual Implicit Diffusion Model for arbitrary-scale MRI super-resolution

Abstract

<p>Magnetic resonance imaging (MRI) is essential for characterizing neurological disorders, yet high-resolution scans are often time-consuming and technically challenging; furthermore, radiologists require the ability to zoom MRI images at arbitrary scales for comprehensive lesion visualization. Arbitrary-scale super-resolution (ASSR) is a vital technique addressing these needs, widely applied in real-world scenarios to achieve high-quality MRI images. Most relevant studies have combined Implicit Neural Representation (INR) with diffusion models to achieve continuous-resolution, diverse, and high-quality ASSR results on general images. However, MRI slice data, characterized by a large proportion of black background, exhibit highly skewed distributions, leading to instability in diffusion training and degraded generation quality. Additionally, the scarcity of high-quality MRI images limits the feasibility of high-magnification ASSR tasks. To tackle these challenges, we propose a novel paradigm called Residual Implicit Diffusion Model (RIDM), which performs the INR-based diffusion process in the residual space. In addition, RIDM integrates an Enhanced Low-resolution Conditioning Network (ELCN) and a training-free self-cascade strategy (SCS) for sampling. ELCN enhances in-distribution ASSR by adaptively extracting high-frequency components from a preliminary upsampling result and thus providing rich features, while SCS leverages in-distribution multi-scale cues to steer sampling on out-of-distribution data. Extensive experiments on two public MRI datasets demonstrate that our method outperforms existing state-of-the-art super-resolution approaches across various scaling factors, showing its potential for clinical applications.<br></p>