US-12625214-B2 - Ultra-low field magnetic resonance imaging system and method thereof

US12625214B2US 12625214 B2US12625214 B2US 12625214B2US-12625214-B2

Abstract

An ultra-low field magnetic resonance imaging system and a method thereof are provided. The ultra-low field magnetic resonance imaging system includes means for collecting a down-sampled three-dimensional ultra-low field magnetic resonance data; means for processing a three-dimensional ultra-low field magnetic resonance data to obtain a three-dimensional isotropic ultra-low field magnetic resonance image; and means for processing data, for inputting the three-dimensional isotropic ultra-low field magnetic resonance image into a pre-trained image enhancement model to obtain a full-sampled three-dimensional magnetic resonance enhanced image.

Inventors

Linfang Xiao
Ruixing ZHU

Assignees

Hangzhou Weiying Medical Technology Co., Ltd.

Dates

Publication Date: 20260512
Application Date: 20240426
Priority Date: 20230829

Claims (10)

1 . An ultra-low field magnetic resonance imaging system, comprising: means for collecting a down-sampled three-dimensional ultra-low field magnetic resonance data; means for processing a three-dimensional ultra-low field magnetic resonance data to obtain a three-dimensional isotropic ultra-low field magnetic resonance image; and means for processing data, for inputting the three-dimensional isotropic ultra-low field magnetic resonance image into a pre-trained image enhancement model to obtain a full-sampled three-dimensional magnetic resonance enhanced image.
2 . The ultra-low field magnetic resonance imaging system of claim 1 , wherein the down-sampled three-dimensional ultra-low field magnetic resonance data comprises K-space data with Fourier down sampling along two dimensions, wherein T1-weighted magnetic resonance data is collected by inversion recovery fast spin echo sequence, and T2-weighted magnetic resonance data is collected by fast spin echo sequence.
3 . The ultra-low field magnetic resonance imaging system of claim 1 , wherein processing the three-dimensional ultra-low field magnetic resonance data comprises performing zero-padding and Fourier transform on the three-dimensional ultra-low field magnetic resonance data to obtain the three-dimensional isotropic ultra-low field magnetic resonance image.
4 . The ultra-low field magnetic resonance imaging system of claim 1 , further comprising: means for acquiring a plurality of full-sampled three-dimensional magnetic resonance images, and performing K-space clipping on each of the full-sampled three-dimensional magnetic resonance enhanced image to obtain a three-dimensional isotropic magnetic resonance image with first image quality; means for adding noise to the three-dimensional isotropic magnetic resonance image, and then sequentially performing K-space clipping, Fourier down sampling, zero-padding and Fourier transform on the three-dimensional isotropic magnetic resonance image added with the noise to obtain a three-dimensional isotropic magnetic resonance image with second image quality, wherein the first image quality is greater than the second image quality; and means for taking the three-dimensional isotropic magnetic resonance image with the first image quality and the corresponding three-dimensional isotropic magnetic resonance image with the second image quality as a training image pair to train and obtain the pre-trained image enhancement model.
5 . The ultra-low field magnetic resonance imaging system of claim 1 , wherein a network structure of the pre-trained image enhancement model comprises a first convolution component, a multi-scale feature extraction component, a spatial attention component, a second convolution component, a first channel attention convolution component, a sub-pixel convolution component, and a first convolution layer, an input of the first convolution component is an input of the image enhancement model, and an output of the first convolution component is connected to an input of a multi-scale feature extraction component; an input of the spatial attention component is connected to an output of the multi-scale feature extraction component, and an output of the spatial attention component is connected to an input of the second convolution component; an input of the first channel attention convolution component is connected to an output of the second convolution component, and an output of the first channel attention convolution component is connected to an input of the sub-pixel convolution component; and an input of the first convolution layer is connected to an output of the sub-pixel convolution component, and voxel intensity of the input of the image enhancement model after up-sampling and voxel intensity of an output of the first convolution layer are added to serve as an output of the image enhancement model.
6 . The ultra-low field magnetic resonance imaging system of claim 5 , wherein the multi-scale feature extraction component comprises a first stride convolution component, a second channel attention convolution component, a third channel attention convolution component, a fourth channel attention convolution component, a first up-sampling layer, a first residual channel attention component, a second up-sampling layer, a second residual channel attention component, a second stride convolution component, an input of the first stride convolution component is connected to the output of the first convolution component, and an output of the first stride convolution component is connected to an input of the second channel attention convolution component; an input of the third channel attention convolution component is connected to an output of the second channel attention convolution component, and an output of the third channel attention convolution component is connected to an input of the fourth channel attention convolution component; an input of the first up-sampling layer is connected to an output of the fourth channel attention convolution component, and an output of the first up-sampling layer and the output of the third channel attention convolution component are connected to an input of the first residual channel attention component after channel splicing; an input of the second up-sampling layer is connected to an output of the first residual channel attention component, and an output of the second up-sampling layer and the output of the second channel attention convolution component are connected to an input of the second residual channel attention component after channel splicing; and an input of the second stride convolution component is connected to an output of the second residual channel attention component, and an output of the second stride convolution component is connected to the input of the spatial attention component.
7 . The ultra-low field magnetic resonance imaging system of claim 6 , wherein the first channel attention convolution component, the second channel attention convolution component, the third channel attention convolution component, and the fourth channel attention convolution component all comprise a third residual channel attention component, a strided convolution layer, and a linear rectification function with leakage which are connected in sequence.
8 . The ultra-low field magnetic resonance imaging system of claim 7 , wherein the first residual channel attention component, the second residual channel attention component, and the third residual channel attention component all comprise a third convolution component, a second convolution layer, and a channel attention component connected in sequence, and voxel intensity of an input of the third convolution component and voxel intensity of an output of the channel attention component are added to serve as an output of the first residual channel attention component, an output of the second residual channel attention component, or an output of the third residual channel attention component.
9 . An ultra-low field magnetic resonance imaging method, which is applied to the ultra-low field magnetic resonance imaging system of claim 1 , comprising: step 1 , the ultra-low field magnetic resonance imaging system collecting a down-sampled three-dimensional ultra-low field magnetic resonance data; step 2 , the ultra-low field magnetic resonance imaging system processing a three-dimensional ultra-low field magnetic resonance data to obtain a three-dimensional isotropic ultra-low field magnetic resonance image; and step 3 , the ultra-low field magnetic resonance imaging system inputting the three-dimensional isotropic ultra-low field magnetic resonance image into a pre-trained image enhancement model to obtain a full-sampled three-dimensional magnetic resonance enhanced image.
10 . The ultra-low field magnetic resonance imaging method of claim 9 , wherein before step 1 is executed, the ultra-low field magnetic resonance imaging method further comprises a model training process, and the model training process comprises following steps: step A 1 , the ultra-low field magnetic resonance imaging system acquiring a plurality of full-sampled three-dimensional magnetic resonance images, and performs K-space clipping on each full-sampled three-dimensional magnetic resonance image to obtain a three-dimensional isotropic magnetic resonance image with first image quality; step A 2 , the ultra-low field magnetic resonance imaging system adding noise to the three-dimensional isotropic magnetic resonance image, and then sequentially performing K-space clipping, Fourier down sampling, zero padding and Fourier transform on the three-dimensional isotropic magnetic resonance image added with noise to obtain a three-dimensional isotropic magnetic resonance image with a second image quality, wherein the first image quality is greater than the second image quality; and step A 3 , the ultra-low field magnetic resonance imaging system taking the three-dimensional isotropic magnetic resonance image with the first image quality and the corresponding three-dimensional isotropic magnetic resonance image with the second image quality as training image pairs to train and obtain the pre-trained image enhancement model.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to Chinese patent application No. 202311098681.5, filed on Aug. 29, 2023, and titled “RAPID ULTRA-LOW FIELD MAGNETIC RESONANCE IMAGING SYSTEM AND METHOD THEREOF”. The content of the above identified application is hereby incorporated by reference in its entirety. TECHNICAL FIELD The present disclosure relates to the field of magnetic resonance imaging technology, and in particular, to an ultra-low field magnetic resonance imaging system and a method thereof. BACKGROUND Ultra-low field magnetic resonance imaging technology has many advantages over magnetic resonance imaging approaches, including excellent openness, ultra-quiet operation, no requirement for electromagnetic shielding, miniaturization, lightweight portability, and the ability to be positioned bedside for patient convenience. It provides a low-cost, low-power, patient-centered solution for clinical magnetic resonance examination, and a tool for specialist diagnosis and general diagnosis non-imaging departments. However, ultra-low field magnetic resonance signals are weak and susceptible to noise interference, resulting in diminished imaging quality and low image resolution. At present, methods to enhance an image quality of the ultra-low field magnetic resonance mainly include a multiple averaging method, an interpolation method, an image reconstruction algorithm, and so on. Although the multiple averaging method does enhance the image quality to a certain extent, it takes a lot of time, consumes significant resources and is costly. In addition, an image obtained by the multiple averaging method lacks details, clarity, and accuracy. Although time cost of the interpolation method is low, its capability to enhance image quality is limited and effect is inconsistent. Relying solely on the conventional image reconstruction algorithms, the image quality of ultra-low field magnetic resonance imaging still exhibits shortcomings in spatial resolution, signal-to-noise ratio, and artifact level. These limitations stem from the fundamental principle of magnetic resonance signal generation, that is, an intensity of magnetic resonance signal is proportional to the square of the main magnetic field strength. Moreover, a typical MRI (magnetic resonance imaging) resolution can be anisotropic, and clinical MRI scans often require imaging in multiple directions, leading to extended scanning time. Prolonged scanning time leads to motion artifacts, which degrades the image quality. Subpar image quality will result in unsatisfactory diagnostic results, significantly impeding a widespread application of ultra-low field magnetic resonance in clinical practice. SUMMARY According to various embodiments, the present disclosure provides an ultra-low field magnetic resonance imaging system, including means for collecting a down-sampled three-dimensional ultra-low field magnetic resonance data; means for processing the three-dimensional ultra-low field magnetic resonance data to obtain a three-dimensional isotropic ultra-low field magnetic resonance image; and means for inputting the three-dimensional isotropic ultra-low field magnetic resonance image into a pre-trained image enhancement model to obtain a full-sampled three-dimensional magnetic resonance enhanced image. In some embodiments, the down-sampled three-dimensional ultra-low field magnetic resonance data includes K-space data with Fourier down sampling along two dimensions. T1-weighted magnetic resonance data is collected by inversion recovery fast spin echo sequence, and T2-weighted magnetic resonance data is collected by fast spin echo sequence. In some embodiments, processing the three-dimensional ultra-low field magnetic resonance data includes performing zero-padding and Fourier transform on the three-dimensional ultra-low field magnetic resonance data to obtain the three-dimensional isotropic ultra-low field magnetic resonance image. In some embodiments, the ultra-low field magnetic resonance imaging system further includes means for acquiring a plurality of full-sampled three-dimensional magnetic resonance images, and performing K-space clipping on each of the full-sampled three-dimensional magnetic resonance enhanced image to obtain a three-dimensional isotropic magnetic resonance image with first image quality; means for adding noise to the three-dimensional isotropic magnetic resonance image, and then sequentially performing K-space clipping, Fourier down sampling, zero-padding, and Fourier transform on the three-dimensional isotropic magnetic resonance image added with the noise to obtain a three-dimensional isotropic magnetic resonance image with second image quality; the first image quality is greater than the second image quality; andmeans for taking the three-dimensional isotropic magnetic resonance image with the first image quality and the corresponding three-dimensional isotropic magnetic resonance image with the second imag