Unsupervised Denoising Method for Multi-sequence Low-field MRI in Veterinary Imaging

Motivation: The low signal-to-noise ratio (SNR) of low-field magnetic resonance images affects the diagnosis of relevant diseases and limits the widespread adoption of low-field MRI systems. Goal(s): We seek to improve the SNR of low field images using unsupervised learning without reference images. Approach: Using an improved ResNet model as the framework, we designed a trainable Sobel convolutional block for shallow edge feature extraction and introduced a convolutional attention module to extract deep edge features from the shallow edge feature map. Results: This network was tested for denoising on 0.4T veterinary images to demonstrate effective denoising while preserving the underlying image structure. Impact: In the absence of reference images, our method preserves the structural features of images while effectively reducing noise. It achieve noise reduction for low-field images in a shorter time and has great potential for improving low-field image quality improvement.