Denoising and Segmentation of Ultra-Low Field Magnetic Resonance Imaging Based on TriNet

Ultra-Low field magnetic resonance Imaging (ULF MRI) systems operate in static magnetic fields of B0 < 250 μT. By eliminating the need for superconducting magnets, ULF MRI offers advantages such as portability, reduced cost, and insensitivity to metal objects. However, ULF MRI systems face the challenge of low signal-to-noise ratio (SNR), primarily due to noise, which can be categorized into power line noise and system white noise. To more effectively mitigate noise and enhance image quality for diagnostic purposes, this paper proposes a convolutional neural network, TriNet. TriNet consists of three sub-networks designed to extract high-frequency, low-frequency, and edge features from images, while employing attention echanism to better integrate these features. The performance of TriNet outperforms existing denoising networks, with a 2 dB increase in peak signal-to-noise ratio (PSNR) and enhanced structural similarity index (SSIM). Additionally, segmentation evaluations were conducted, comparing clean images with noisy images processed by TriNet. The results show that images denoised by TriNet yield segmentation results with an intersection over union (IoU) score only 3% less and a Dice score 6% less compared to direct segmentation of noisy images. The TriNet network architecture significantly improves the imaging SNR in the image domain, laying the foundation for further advancement of ULF imaging in human applications.