BMENet: 3D Brain MRI Enhancement Against Various Degradations with Coarse to Fine Network

Magnetic resonance imaging (MRI) is crucial for medical diagnosis but often suffers from various degradations, such as noise, motion artifacts, and intensity inhomogeneity, leading to misdiagnoses or suboptimal treatment options. Existing methods typically focus on two-dimensional (2D) slices or individual types of degradation, which limits flexibility and applicability. This paper proposes BMENet: a two-stage three-dimensional (3D) Brain MRI Enhancement Network, which is designed to simultaneously handle denoising, artifact removal, and intensity inhomogeneity correction. In the first stage, we use a 2D slice coarse enhancement model to remove the bulk of the degradation; in the second stage, a 3D fine control latent diffusion generation model is used to restore missing details. Additionally, in the second stage, an edge-sensitive strategy is used to priority fine structural details at tissue boundaries, and a progressive constraint mechanism is applied to guide recovery. We train and test the proposed method on T1-weighted brain MRI images. Experimental results demonstrate that BMENet outperforms several state-of-the-art (SOTA) techniques in both quantitative and visual evaluations. Additionally, we conducted segmentation tests, which showed that the segmentation of brain regions improved significantly after the enhancement of degraded images. The relevant code can be found at: https://github.com/LSYLAB/BMENet.git.