A noise reduction method and noise reduction system for magnetic resonance imaging

Abstract

The present invention provides a noise reduction method and system for nuclear magnetic resonance imaging, belonging to the field of nuclear magnetic resonance technology. The method sets an environmental noise acquisition module around a nuclear magnetic resonance imaging coil or instrument without coupling with the imaging coil, collects environmental noise and inputs it into a data acquisition module; uses the received environmental noise to eliminate the environmental noise in the collected raw data at the same time, and obtains intermediate image data; after building and training a neural network DEEP-IFT, the intermediate image data is input into the trained neural network DEEP-IFT to eliminate the random noise of the intermediate image data; then builds an MRI-DN network based on the Noise2noise principle, further eliminates random noise and enhances the signal-to-noise ratio, completes the denoising of the nuclear magnetic resonance image, and outputs the final nuclear magnetic resonance image. The present invention removes environmental noise and random noise, improves the quality of nuclear magnetic resonance images, and improves the portability of the instrument.

Description

Technical Field

[0001] The present invention belongs to the field of nuclear magnetic resonance technology, and in particular relates to a noise reduction method and a noise reduction system for nuclear magnetic resonance imaging. Background Art

[0002] Since its discovery in 1946, nuclear magnetic resonance technology has been widely used in physics, chemistry, materials science, life science and medicine. As one of the most effective non-destructive testing methods, nuclear magnetic resonance technology has been used from using relaxation time spectra to reveal the composition of oil and water thousands of meters below the ground, to analyzing chemical molecular structures with spectra, to identifying human organ lesions with medical images. Among them, magnetic resonance imaging (MRI) is a clinical medical diagnostic method that uses biomagnetic spin imaging. The human body generates signals under the action of an external static magnetic field and radio frequency pulses, and the gradient field realizes the spatial encoding of the signals. Finally, the image is displayed on the screen after detection, processing and conversion with a probe.

[0003] In the prior art, MRIs generally rely on electromagnetic shielding devices such as shielded rooms to ensure imaging quality and stability. However, shielded rooms have certain limitations in eliminating electromagnetic effects, while increasing instrument costs and hindering the promotion and application of MRIs, especially lightweight low-field MRIs. Summary of the invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, the present invention aims to provide a noise reduction method and a noise reduction system for magnetic resonance imaging, which collects noise based on raw magnetic resonance data and environmental noise data, and then eliminates the collected noise based on a deep learning network algorithm to remove noise and improve the quality of magnetic resonance images.

[0005] In order to achieve the above purpose, the embodiment of the present invention adopts the following technical solution:

[0006] In a first aspect, an embodiment of the present invention provides a noise reduction method for magnetic resonance imaging, comprising the following steps:

[0007] Step S1, setting an environmental noise collection module around a nuclear magnetic resonance imaging coil or instrument without coupling with the imaging coil; collecting environmental noise through the environmental noise collection module, and inputting the collected environmental noise into an environmental noise elimination module;

[0008] Step S2: the data acquisition module acquires raw nuclear magnetic resonance data, and the environmental noise elimination module uses the received environmental noise to eliminate the environmental noise in the raw nuclear magnetic resonance data acquired by the data acquisition module.

[0009] As a preferred embodiment of the present invention, the method further includes:

[0010] The step S2 eliminates the environmental noise in the collected original NMR data to obtain intermediate data;

[0011] Step S3, constructing a neural network DEEP-IFT to simultaneously generate a real part image and an imaginary part image of a nuclear magnetic resonance;

[0012] Step S4, training the constructed neural network DEEP-IFT based on historical data to obtain a mature neural network DEEP-IFT;

[0013] Step S5, inputting the intermediate data into the mature neural network DEEP-IFT, eliminating the random noise of the intermediate data, and converting and outputting the first nuclear magnetic resonance image.

[0014] As a preferred embodiment of the present invention, the method further includes:

[0015] Step S6, constructing an MRI-DN network based on the Noise2noise principle;

[0016] Step S7, inputting the first nuclear magnetic resonance image obtained in step S5 into the MRI-DN network, further eliminating random noise and enhancing the signal-to-noise ratio based on the principle that noise is randomly generated relative to the signal, completing denoising of the nuclear magnetic resonance image, and outputting the final nuclear magnetic resonance image.

[0017] As a preferred embodiment of the present invention, the constructed neural network DEEP-IFT includes at least one of the following configurations:

[0018] The first configuration adopts a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer and two branch structures in sequence, each branch structure includes two two-dimensional convolutional layers and two sub-branch structures, one of the sub-branch structures is a two-dimensional convolutional transposition layer and a reconstruction layer in sequence, and the other sub-branch structure is a two-dimensional convolutional layer; the two fully connected layers use the TanH function; the bifurcated network data processing process of the first configuration includes: the input data is processed by the TanH function in the one-dimensional fully connected network, and then the output of the one-dimensional fully connected network enters the two-dimensional zero-filling layer after the reconstruction layer and is re-transposed into two-dimensional data, the two-dimensional data is respectively introduced into two two-layer convolutional networks of the two branch structures, the real data and the imaginary data of the image are respectively generated, the values ​​of the real part and the imaginary part are respectively calculated, and then the attenuation values ​​of the real part and the imaginary part are respectively calculated by the two-dimensional convolutional transposition layer and the reconstruction layer in the branch structure; the attenuation function calculates the values ​​of the real part and the imaginary part respectively and superimposes them as the selection of the fitting direction;

[0019] The second configuration adopts a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer, two branch structures, a connection layer and a reconstruction layer in sequence, each branch structure includes two two-dimensional convolutional layers and one two-dimensional convolutional transposition layer; the two fully connected layers use the TanH function; the bifurcated network data processing process of the second configuration includes: the input data is processed by the TanH function in the one-dimensional fully connected network, and then the output of the one-dimensional fully connected network enters the two-dimensional zero-filling layer after the reconstruction layer and is re-transposed into two-dimensional data, the two-dimensional data is respectively introduced into the two bifurcated two-layer convolutional networks, the real data and the imaginary data of the image are respectively generated, and the values ​​of the real part and the imaginary part are respectively calculated; the real data and the imaginary data enter the connection layer and the reconstruction layer at the same time, and are integrated as one data to calculate the attenuation value;

[0020] The third configuration includes, in sequence, a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional stack filling layer, two two-dimensional convolutional layers and two branch structures, one of which includes, in sequence, a two-dimensional convolutional transposition layer and a reconstruction layer, and the other branch structure is a two-dimensional convolutional layer; the two fully connected layers use the TanH function; the data processing process of the third configuration includes: the input data is processed using the TanH function in a one-dimensional fully connected network, the output of the one-dimensional fully connected network is re-transposed into two-dimensional data, the two-dimensional data is simultaneously generated through a double-layer convolutional network. Real data and imaginary data of the image are integrated, and the attenuation values ​​of the real data and the imaginary data are calculated.

[0021] As a preferred embodiment of the present invention, the environmental noise acquisition module is multi-channel, and each channel includes a noise measurement probe and a preamplifier.

[0022] As a preferred embodiment of the present invention, the noise measurement probe is a resonant circuit composed of a coil and a capacitor, the resonant frequency is the Larmor frequency of the nuclear magnetic resonance imager, and has a bandwidth similar to that of nuclear magnetic resonance measurement.

[0023] As a preferred embodiment of the present invention, the environmental noise in the collected original nuclear magnetic resonance data is eliminated by using the received environmental noise, and the elimination is performed based on the similarity principle of the environmental noise.

[0024] As a preferred embodiment of the present invention, the MRI-DN network is constructed based on the Noise2noise principle, including:

[0025] Using the Noise2noise network, the target slice image and 2n adjacent slice images, n = 1 or 2, are passed through the U-NET network and then the mean square error value is calculated to eliminate the noise.

[0026] In a second aspect, an embodiment of the present invention further provides a noise reduction system for magnetic resonance imaging, the system comprising: an environmental noise acquisition module, an environmental noise elimination module, a DEEP-IFT module and an MRI-DN network module; wherein:

[0027] The environmental noise acquisition module is arranged around the imaging coil or the instrument, and is used to collect environmental noise without coupling with the imaging coil, and input the collected environmental noise into the data environmental noise elimination module;

[0028] The environmental noise elimination module is used to eliminate the environmental noise in the original nuclear magnetic resonance data collected by the data acquisition module by using the received environmental noise, so that the data acquisition module in the nuclear magnetic resonance imaging system obtains intermediate data and sends it to the DEEP-IFT module;

[0029] The DEEP-IFT module is used to construct and train the neural network DEEP-IFT, and input the intermediate data into the trained neural network DEEP-IFT, eliminate the random noise of the intermediate data, transform and output the first nuclear magnetic resonance image, and transmit it to the MRI-DN network module;

[0030] The MRI-DN network module is used to construct an MRI-DN network based on the Noise2noise principle, and according to the principle that noise is randomly generated relative to the signal, further eliminate the random noise in the first nuclear magnetic resonance image, while enhancing the signal-to-noise ratio, completing the denoising of the nuclear magnetic resonance image, and outputting the final nuclear magnetic resonance image.

[0031] As a preferred embodiment of the present invention, the environmental noise acquisition module is a multi-channel mode, each channel includes a noise measurement probe and a preamplifier, and the preamplifier is connected to the environmental noise elimination module. The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0032] The noise reduction method and system for nuclear magnetic resonance imaging eliminate environmental noise based on the processing of nuclear magnetic resonance raw data and environmental noise data, and then remove random noise based on a deep learning network algorithm to improve the quality of nuclear magnetic resonance images; and are also applicable to low-field and high-field nuclear magnetic resonance imagers in shielded rooms and unshielded states, thereby reducing the cost of the instrument, improving the mobility of the instrument, and expanding the scope of use of low-field and high-field nuclear magnetic resonance imagers.

[0033] Of course, it is not necessary to achieve all of the advantages described above at the same time to implement any product or method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 is a flow chart of a noise reduction method for magnetic resonance imaging provided by an embodiment of the present invention;

[0036] Figure 2 is a schematic diagram of an environmental noise collection module in an embodiment of the present invention;

[0037] Figure 3 is a schematic diagram of a noise measurement probe coil in an embodiment of the present invention;

[0038] Figure 4 is a graph of original data before removing environmental noise from nuclear magnetic resonance data in an embodiment of the present invention;

[0039] Figure 5 is a data diagram of nuclear magnetic resonance data after removing environmental noise in an embodiment of the present invention;

[0040] Figure 6 It is a first configuration diagram of the neural network DEEP-IFT in an embodiment of the present invention;

[0041] Figure 7 It is a second configuration diagram of the neural network DEEP-IFT in an embodiment of the present invention;

[0042] Figure 8 It is a third configuration diagram of the neural network DEEP-IFT in an embodiment of the present invention;

[0043] Fig. 9 1 is an example diagram of image reconstruction using Fourier transform and neural network DEEP-IFT in an embodiment of the present invention;

[0044] Fig.10 is an example diagram of reducing noise using the neural network DEEP-IFT in an embodiment of the present invention;

[0045] Fig.11 Schematic diagram of the algorithm architecture of the MRI-DN network in an embodiment of the present invention;

[0046] Fig.12 It is an example diagram of noise reduction by the MRI-DN network in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] After discovering the above problem, the inventor of the present application conducted a detailed study on the noise problem of existing nuclear magnetic resonance imaging. Noise is the most important problem encountered in nuclear magnetic resonance imaging. In order to expand the application scope of low-field nuclear magnetic resonance imaging devices and enhance the stability and mobility of low-field nuclear magnetic resonance imaging devices, it is necessary to develop a denoising technology that does not rely on electromagnetic shielding devices.

[0048] With the development and rapid popularization of deep learning technology, it has become feasible to use deep learning to improve the signal-to-noise ratio of MRI. Deep learning technology can be used to replace the classic Fourier transform technology to achieve image reconstruction, and at the same time, high-resolution data with random noise eliminated can be obtained. The latest development of Noise2noise technology can also achieve signal enhancement and noise elimination for various images. However, MRI images are different from general images, and there are still certain problems in how to apply denoising technology based on deep learning algorithms to MRI images.

[0049] It should be noted that the defects existing in the solutions in the above-mentioned prior art are the results obtained by the inventor after practice and careful research. Therefore, the discovery process of the above-mentioned problems and the solutions proposed in the embodiments of the present invention for the above-mentioned problems below should all be the contributions made by the inventor to the present invention in the process of the present invention.

[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. It should be noted that the embodiments of the present invention and the features in the embodiments can also be combined with each other without conflict.

[0051] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In the description of the present invention, the terms "first", "second", "third", "fourth", etc. are only used to distinguish the description and cannot be understood as just or implying relative importance.

[0052] After the above in-depth analysis, the present invention proposes a noise reduction method and noise reduction system for nuclear magnetic resonance imaging. In various noise environments without electromagnetic shielding, environmental noise and random noise will have a greater impact on the imaging quality of nuclear magnetic resonance imaging. Therefore, the present invention runs a software program based on a deep learning algorithm on a hardware device. After collecting the original nuclear magnetic resonance data and environmental noise data, the original data is firstly subjected to noise reduction processing by using the correlation of environmental noise to remove the influence of environmental noise; then, based on the deep learning network DEEP-IFT of a special configuration model, model-based image reconstruction and further noise reduction are realized; finally, the deep learning network MRI-DN based on the Noise2noise principle is used to eliminate random noise and enhance the image. The present invention makes it no longer necessary to perform electromagnetic shielding when using a nuclear magnetic resonance imager, which has great advantages over the traditional Fourier transform image reconstruction algorithm, and can realize the effective restoration and image reconstruction of high-noise data; at the same time, it can be used for the original signal environment noise reduction, image reconstruction and random noise elimination of low-field nuclear magnetic resonance imagers, and can also be extended to the field of high-field and ultra-high-field nuclear magnetic resonance imaging.

[0053] like Figure 1 As shown, the noise reduction method for magnetic resonance imaging provided by the embodiment of the present invention comprises the following steps:

[0054] Step S1, setting an environmental noise collection module around a nuclear magnetic resonance imaging coil or instrument without coupling with the imaging coil; collecting environmental noise through the environmental noise collection module, and inputting the collected environmental noise into an environmental noise elimination module.

[0055] This step first collects environmental noise by setting corresponding components, and then uses the collected environmental noise in subsequent steps to eliminate environmental noise in the original data of nuclear magnetic resonance according to the principle of similarity, thereby expanding the scope of use of nuclear magnetic resonance imaging. Environmental noise has a particularly serious impact on imaging quality in low-field nuclear magnetic resonance imaging. When low-field nuclear magnetic resonance imaging devices are used in a mobile manner and in different environments without the use of shielded rooms, it becomes unrealistic to use physical electromagnetic shielding to shield environmental noise. Therefore, this step collects environmental noise by setting an environmental noise probe.

[0056] like Figure 2 As shown, in this step, the environmental noise acquisition module is multi-channel, each channel includes a noise measurement probe 201 and a preamplifier 202, and the preamplifier 202 is connected to the data acquisition module 203. Specifically, 1 to N noise measurement probes 201 are placed at different positions of the nuclear magnetic resonance imager to detect environmental noise, and the noise signal is amplified by the preamplifier 202 and input into the environmental noise elimination module of the data acquisition module 203.

[0057] In a specific embodiment, the noise measurement probe 201 is composed of a coil and a capacitor to form a resonant circuit, the resonant frequency is the Larmor frequency of the nuclear magnetic resonance imager, and has a bandwidth similar to that of nuclear magnetic resonance measurement; Figure 3 As shown, the coil includes but is not limited to being implemented by a surface coil 204 or a spiral coil 205. In another specific embodiment, the noise measurement probe 201 is implemented by an electrode probe.

[0058] Step S2, the data acquisition module acquires original nuclear magnetic resonance data, and the environmental noise elimination module uses the received environmental noise to eliminate the environmental noise in the original nuclear magnetic resonance data acquired by the data acquisition module; and intermediate data is obtained after eliminating the environmental noise in the acquired original nuclear magnetic resonance data.

[0059] In this step, the environmental noise elimination module is set separately or in the data acquisition module. The environmental noise in the collected nuclear magnetic resonance raw data is eliminated by using the received environmental noise data, and the environmental noise is eliminated according to the similarity principle of the environmental noise. After receiving the environmental noise signal, the environmental noise elimination module of the data acquisition module 203 first converts the analog signal of the received environmental noise into a digital signal, and then eliminates the environmental noise in the collected raw nuclear magnetic resonance data according to the similarity principle of the environmental noise.

[0060] Figure 4 This is a specific example of removing environmental noise from NMR raw data. Figure 5 This is the comparison of the original NMR data in this specific example after removing the environmental noise. Figure 4 and Figure 5 As shown, the collected nuclear magnetic resonance image data is a multi-echo data, including two data of real part and imaginary part; after the step of eliminating the environmental noise, the nuclear magnetic resonance raw data becomes smooth and uniform.

[0061] Step S3, constructing a neural network DEEP-IFT to simultaneously generate a real part image and an imaginary part image of a nuclear magnetic resonance.

[0062] In this step, the constructed neural network DEEP-IFT includes at least one of the following configurations:

[0063] The first configuration uses a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer, and two branch structures, each of which includes two two-dimensional convolutional layers and two sub-branch structures, one of which is a two-dimensional convolutional transposition layer and a reconstruction layer, and the other is a two-dimensional convolutional layer. Figure 6As shown, 402a is a fully connected network, which consists of two layers of fully connected networks, and both layers of fully connected networks use the TanH function; the bifurcated network data processing process of the first configuration includes: the input data is processed using the TanH function in a one-dimensional fully connected network, and then the output of the one-dimensional fully connected network is re-transposed into two-dimensional data, and the two-dimensional data is respectively imported into the two bifurcated two-layer convolutional networks to generate the real data 403a and imaginary data 404a of the image, and the two-dimensional convolutional network uses the ReLU function. Among them, the part 402a as the input end can also be implemented using a fully connected network of 3 layers or more, and the 403a and 404a as the convolutional network part can also be implemented using a convolutional network of 3 layers or more, and the attenuation function uses the mean square error (MSE). The attenuation function calculates the attenuation values ​​of the real part and the imaginary part respectively and then superimposes them as the selection of the fitting direction. In addition, the kernel and bias parameters in the network architecture can be adjusted according to the acquisition channel and the size of the generated image; a two-dimensional convolutional transpose layer, a reset layer, and a two-dimensional zero-filling layer are used to connect the fully connected layer and the convolutional layer.

[0064] The second configuration uses a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer, two branch structures, a connection layer, and a reconstruction layer. Each branch structure includes two two-dimensional convolutional layers and one two-dimensional convolutional transposition layer. Figure 7 As shown, 402b is a fully connected network, which consists of two layers of fully connected networks, and both layers of fully connected networks use the TanH function; the second type of bifurcated network data processing process includes: the input data is processed using the TanH function in a one-dimensional fully connected network, and then the output of the one-dimensional fully connected network is re-transposed into two-dimensional data, and the two-dimensional data is respectively imported into the two bifurcated two-layer convolutional networks to generate the real data 403b and imaginary data 404b of the image, and the two-dimensional convolutional network uses the ReLU function. Among them, the 402b part as the input end can also be implemented using a fully connected network of 3 layers or more, and the 403b and 404b as the convolutional network part can also be implemented using a convolutional network of 3 layers or more, and the attenuation function uses MSE. The real data and the imaginary data enter the connection layer and the reconstruction layer at the same time, and after integration, the attenuation value is calculated as a data as the choice of fitting direction. In addition, the core parameters (kernel) and bias parameters (bias) in the network architecture can be adjusted according to the acquisition channel and the size of the generated image. We also use a 2D convolution transpose layer, a reset layer, and a 2D zero padding layer to connect the fully connected layer and the convolution layer. The second configuration is similar to the first configuration, except that the real and imaginary parts are integrated into the same data to calculate the attenuation value and fit it.

[0065] The third configuration includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional stacking filling layer, two two-dimensional convolutional layers, and two branch structures, one of which includes a two-dimensional convolutional transposition layer and a reconstruction layer, and the other branch structure is a two-dimensional convolutional layer. Figure 8 As shown, 402c is a fully connected network, which is composed of two layers of fully connected networks, and both layers of fully connected networks use TanH function; the data processing process of the third configuration includes: the input data is processed by TanH function in a one-dimensional fully connected network, and then the output of the one-dimensional fully connected network is re-transposed into two-dimensional data, and the two-dimensional data is respectively used by a two-layer convolution network to simultaneously generate the real data and imaginary data 403c of the image, and we use ReLU function for the two-dimensional convolution network. Among them, the 402c part as the input end can also be implemented using a fully connected network of 3 layers or more, and the 403c as the convolution network part can also be implemented using a convolution network of 3 layers or more, and the attenuation function uses MSE. The two-dimensional data is simultaneously generated by the two-layer convolution network for the real data and the imaginary data of the image, and the integrated value of the real data and the imaginary data is calculated, and then the attenuation value of the real data and the imaginary data is simultaneously calculated as the selection of the fitting direction. In addition, the core parameters (kernel) and bias parameters (bias) in the network architecture can be adjusted according to the acquisition channel and the size of the generated image. We also use a 2D convolution transpose layer, a reset layer, and a 2D zero padding layer to connect the fully connected layer and the convolution layer. The third configuration 401c is significantly different from the previous two configurations in that the real data and the imaginary data are output as a (64×64×2 or 128×128×2) data.

[0066] The above three configurations can generate real images and imaginary images at the same time without training two networks separately. In other words, there is no need to train two networks at the same time, but to realize the reconstruction of real and imaginary images on the same network.

[0067] Step S4, training the constructed neural network DEEP-IFT based on historical data to obtain a mature neural network DEEP-IFT.

[0068] In this step, in a specific embodiment, preferably, 50,000 nuclear magnetic resonance imaging data are used to train DEEP-IFT. The training is iterated 200 times in total, and the loss function reaches 10 -5 Through training, DEEP-IFT can successfully replace the inverse Fourier transform and convert time domain data into images. Fig. 9The image conversion results shown show that compared with the data obtained by DEEP-IFT and inverse Fourier transform (501, 502 and 503), the DEEP-IFT output results (504 and 505) almost achieve 100% high-fidelity level restoration, meeting the design requirements.

[0069] Another important function of DEEP-IFT is to achieve a certain degree of noise elimination, which is random noise. First, 50,000 nuclear magnetic resonance images are converted into time domain data (or spatial data) and Gaussian noise is introduced as input data. Then, the 50,000 nuclear magnetic resonance images are used as output data to retrain the neural network. The training was iterated for about 200 times, and the loss function reached 10 -5 about. Fig.10 The result of the test data comparison is shown in Figure 6. Compared with the data 602 with noise, the data 603 after noise removal is highly similar to the original data 601, indicating that DEEP-IFT can eliminate random noise to a certain extent. In the example, the network architecture of DEEP-IFT improves the signal-to-noise ratio by 3 times.

[0070] Step S5, inputting the intermediate image data into the mature neural network DEEP-IFT, eliminating the random noise of the intermediate data, and converting and outputting the first nuclear magnetic resonance image.

[0071] In this step, the trained neural network DEEP-IFT is used to simultaneously reconstruct the real and imaginary parts of the MRI raw data. DEEP-IFT's network architecture is a new network architecture that can simultaneously reconstruct the real and imaginary parts of the MRI raw data. The main features of DEEP-IFT's network architecture include the use of a fully connected network for signal processing Figure 6 and bifurcation processing based on two-dimensional convolutional networks Figure 7 and Figure 8 At the same time, the neural network DEEP-IFT architecture can reconstruct images well, replacing the traditional Fourier transform algorithm.

[0072] Step S6, constructing an MRI-DN network based on the Noise2noise principle.

[0073] like Fig.11As shown, the MRI-DN network is a magnetic resonance imaging data noise elimination network designed using the Noise2noise network. Unlike the Noise2noise network, MRI-DN uses its own data for noise reduction. The target slice image and the adjacent 2n slice images (n=1 or 2) 701 are passed through the U-NET network 702 and then the mean square error MSE value is calculated to eliminate the noise. Through testing, it is found that MRI-DN makes full use of the three-dimensional MRI data, the randomness of the noise of adjacent slice images, and the similarity of the signals to extract the signal.

[0074] Step S7, inputting the first nuclear magnetic resonance image obtained in step S5 into the MRI-DN network, further eliminating random noise and enhancing the signal-to-noise ratio based on the principle that noise is randomly generated relative to the signal, completing denoising of the nuclear magnetic resonance image, and outputting the final nuclear magnetic resonance image.

[0075] like Fig.12 As shown, a is the nuclear magnetic resonance imaging picture before processing, and b is the nuclear magnetic resonance imaging picture after processing. Fig.12 It can be seen that random noise is eliminated well through this step.

[0076] The MRI-DN network architecture is mainly used for image enhancement. The MRI-DN network adopts a self-learning algorithm and uses the randomness of noise and the correlation of signals to distinguish and eliminate noise. The MRI-DN network can obtain high signal-to-noise ratio data, thereby achieving image enhancement. Compared with traditional image enhancement algorithms such as GNN, the MRI-DN network does not produce unexplained abnormal signals, making image enhancement closer to the real situation. In the embodiment, the MRI-DN network architecture improves the signal-to-noise ratio by 1.5 times.

[0077] It should be noted that the noise reduction method for nuclear magnetic resonance imaging provided in this embodiment is also applicable to the noise reduction and image reconstruction of low-field and high-field nuclear magnetic resonance imagers in shielded rooms. The shielded room can isolate external environmental noise, but cannot shield internal environmental noise (such as noise generated by the human body). The multi-channel environmental noise acquisition module can be used to simultaneously obtain various noises near the nuclear magnetic resonance imaging system, so the multi-channel environmental noise acquisition module can be simultaneously applied to low-field and high-field nuclear magnetic resonance imagers based on shielded rooms, and low-field and high-field nuclear magnetic resonance imagers that can be used movably without shielded rooms.

[0078] It can be seen from the above technical solutions that the noise reduction method for nuclear magnetic resonance imaging provided in this embodiment collects noise based on nuclear magnetic resonance raw data and environmental noise data, eliminates the collected noise, and then removes random noise based on a deep learning network algorithm, thereby improving the quality of nuclear magnetic resonance images; it is also suitable for low-field and high-field nuclear magnetic resonance imagers in shielded rooms and unshielded states, thereby improving the mobility of the instrument and expanding the scope of use of low-field and high-field nuclear magnetic resonance imagers.

[0079] Based on the same idea, an embodiment of the present invention further provides a noise reduction system for magnetic resonance imaging, the system comprising: an environmental noise acquisition module, an environmental noise elimination module, a DEEP-IFT module and an MRI-DN network module.

[0080] The environmental noise acquisition module is arranged around the imaging coil or the instrument, and is used to collect environmental noise without coupling with the imaging coil, and input the collected environmental noise into the data environmental noise elimination module; the environmental noise acquisition module is a multi-channel mode, each channel includes a noise measurement probe and a preamplifier, and the preamplifier is connected to the environmental noise elimination module;

[0081] The environmental noise elimination module is used to eliminate the environmental noise in the original nuclear magnetic resonance data collected by the data acquisition module by using the received environmental noise, so that the data acquisition module in the nuclear magnetic resonance imaging system obtains intermediate data and sends it to the DEEP-IFT module;

[0082] The DEEP-IFT module is used to construct and train the neural network DEEP-IFT, and input the intermediate data into the trained neural network DEEP-IFT, eliminate the random noise of the intermediate data, transform and output the first nuclear magnetic resonance image, and transmit it to the MRI-DN network module;

[0083] The MRI-DN network module is used to construct an MRI-DN network based on the Noise2noise principle, and according to the principle that noise is randomly generated relative to the signal, further eliminate the random noise in the first nuclear magnetic resonance image, while enhancing the signal-to-noise ratio, completing the denoising of the nuclear magnetic resonance image, and outputting the final nuclear magnetic resonance image.

[0084] In this embodiment, each module is implemented by a processor, and a memory is appropriately added when storage is required. Among them, the processor can be but is not limited to a microprocessor MPU, a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), other programmable logic devices, discrete gates, transistor logic devices, discrete hardware components, etc. The memory can include a random access memory (RAM) and can also include a non-volatile memory (NVM), such as at least one disk storage. Optionally, the memory can also be at least one storage device located away from the aforementioned processor.

[0085] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center.

[0086] It should also be noted that the noise reduction system for magnetic resonance imaging described in this embodiment can be an independent system, or it can be integrated into or modify an existing magnetic resonance imaging system; the system corresponds to the noise reduction method for magnetic resonance imaging, and the description and limitation of the method are also applicable to the system, which will not be repeated here.

[0087] The above description is only a preferred embodiment of the present invention and an explanation of the technical principles used. It is not intended to limit the scope of the invention claimed for protection, but only represents the preferred embodiment of the present invention. Those skilled in the art should understand that the scope of the invention involved in the present invention is not limited to the technical solution formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the inventive concept. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

Claims

1. A noise reduction method for magnetic resonance imaging, characterized in that: The steps include: Step S1, setting an environmental noise collection module around a nuclear magnetic resonance imaging coil or instrument without coupling with the imaging coil; collecting environmental noise through the environmental noise collection module, and inputting the collected environmental noise into an environmental noise elimination module; Step S2, the data acquisition module acquires raw nuclear magnetic resonance data, and the environmental noise elimination module uses the received environmental noise to eliminate the environmental noise in the raw nuclear magnetic resonance data acquired by the data acquisition module, and obtains intermediate data; Step S3, constructing a neural network DEEP-IFT to simultaneously generate a real part image and an imaginary part image of a nuclear magnetic resonance; Step S4, training the constructed neural network DEEP-IFT based on historical data to obtain a mature neural network DEEP-IFT; Step S5, inputting the intermediate data into the mature neural network DEEP-IFT, eliminating random noise of the intermediate data, and converting and outputting the first nuclear magnetic resonance image; Step S6, constructing an MRI-DN network based on the Noise2noise principle; Step S7, inputting the first nuclear magnetic resonance image obtained in step S5 into the MRI-DN network, further eliminating random noise and enhancing the signal-to-noise ratio based on the principle that noise is randomly generated relative to the signal, completing denoising of the nuclear magnetic resonance image, and outputting the final nuclear magnetic resonance image; The constructed neural network DEEP-IFT includes at least one of the following configurations: The first configuration adopts a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer and two branch structures in sequence, each branch structure includes two two-dimensional convolutional layers and two sub-branch structures, one of the sub-branch structures is a two-dimensional convolutional transposition layer and a reconstruction layer in sequence, and the other sub-branch structure is a two-dimensional convolutional layer; the two fully connected layers use the TanH function; the bifurcated network data processing process of the first configuration includes: the input data is processed by the TanH function in the one-dimensional fully connected network, and then the output of the one-dimensional fully connected network enters the two-dimensional zero-filling layer after the reconstruction layer and is re-transposed into two-dimensional data, the two-dimensional data is respectively introduced into two two-layer convolutional networks of the two branch structures, the real data and the imaginary data of the image are respectively generated, the values ​​of the real part and the imaginary part are respectively calculated, and then the attenuation values ​​of the real part and the imaginary part are respectively calculated by the two-dimensional convolutional transposition layer and the reconstruction layer in the branch structure; the attenuation function calculates the values ​​of the real part and the imaginary part respectively and superimposes them as the selection of the fitting direction; The second configuration adopts a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer, two branch structures, a connection layer and a reconstruction layer in sequence, each branch structure includes two two-dimensional convolutional layers and one two-dimensional convolutional transposition layer; the two fully connected layers use the TanH function; the bifurcated network data processing process of the second configuration includes: the input data is processed by the TanH function in the one-dimensional fully connected network, and then the output of the one-dimensional fully connected network enters the two-dimensional zero-filling layer after the reconstruction layer and is re-transposed into two-dimensional data, the two-dimensional data is respectively introduced into the two bifurcated two-layer convolutional networks, the real data and the imaginary data of the image are respectively generated, and the values ​​of the real part and the imaginary part are respectively calculated; the real data and the imaginary data enter the connection layer and the reconstruction layer at the same time, and are integrated as one data to calculate the attenuation value; The third configuration includes, in sequence, a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional stack filling layer, two two-dimensional convolutional layers and two branch structures, one of which includes, in sequence, a two-dimensional convolutional transposition layer and a reconstruction layer, and the other branch structure is a two-dimensional convolutional layer; the two fully connected layers use the TanH function; the data processing process of the third configuration includes: the input data is processed using the TanH function in a one-dimensional fully connected network, the output of the one-dimensional fully connected network is re-transposed into two-dimensional data, the two-dimensional data is simultaneously generated through a double-layer convolutional network. Real data and imaginary data of the image are integrated, and the attenuation values ​​of the real data and the imaginary data are calculated.

2. The noise reduction method for magnetic resonance imaging according to claim 1, characterized in that: The environmental noise acquisition module is multi-channel, and each channel includes a noise measurement probe and a preamplifier.

3. The noise reduction method for magnetic resonance imaging according to claim 2, characterized in that: The noise measurement probe is a resonant circuit composed of a coil and a capacitor, the resonant frequency is the Larmor frequency of a nuclear magnetic resonance imager, and has a bandwidth similar to that of nuclear magnetic resonance measurement.

4. The noise reduction method for magnetic resonance imaging according to claim 1, characterized in that: The received environmental noise is used to eliminate the environmental noise in the collected original nuclear magnetic resonance data, and the elimination is performed according to the similarity principle of the environmental noise.

5. The noise reduction method for magnetic resonance imaging according to claim 1, characterized in that: The MRI-DN network is constructed based on the Noise2noise principle, including: Using the Noise2noise network, the target slice image and 2n adjacent slice images, n = 1 or 2, are passed through the U-NET network and then the mean square error value is calculated to eliminate the noise.

6. A noise reduction system for magnetic resonance imaging, characterized in that: The system includes: an environmental noise acquisition module, an environmental noise elimination module, a DEEP-IFT module and an MRI-DN network module; wherein, The environmental noise acquisition module is arranged around the imaging coil or the instrument, and is used to collect environmental noise without coupling with the imaging coil, and input the collected environmental noise into the data environmental noise elimination module; The environmental noise elimination module is used to eliminate the environmental noise in the original nuclear magnetic resonance data collected by the data acquisition module by using the received environmental noise, so that the data acquisition module in the nuclear magnetic resonance imaging system obtains intermediate data and sends it to the DEEP-IFT module; The DEEP-IFT module is used to construct and train the neural network DEEP-IFT, and input the intermediate data into the trained neural network DEEP-IFT, eliminate the random noise of the intermediate data, transform and output the first nuclear magnetic resonance image, and transmit it to the MRI-DN network module; The MRI-DN network module is used to construct an MRI-DN network based on the Noise2noise principle, and further eliminate the random noise in the first nuclear magnetic resonance image according to the principle that noise is randomly generated relative to the signal, while enhancing the signal-to-noise ratio, completing the denoising of the nuclear magnetic resonance image, and outputting the final nuclear magnetic resonance image; The environmental noise acquisition module is a multi-channel mode, each channel includes a noise measurement probe and a preamplifier, and the preamplifier is connected to the environmental noise elimination module; The neural network DEEP-IFT comprises at least one of the following configurations: The first configuration adopts a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer and two branch structures in sequence, each branch structure includes two two-dimensional convolutional layers and two sub-branch structures, one of the sub-branch structures is a two-dimensional convolutional transposition layer and a reconstruction layer in sequence, and the other sub-branch structure is a two-dimensional convolutional layer; the two fully connected layers use the TanH function; the bifurcated network data processing process of the first configuration includes: the input data is processed by the TanH function in the one-dimensional fully connected network, and then the output of the one-dimensional fully connected network enters the two-dimensional zero-filling layer after the reconstruction layer and is re-transposed into two-dimensional data, the two-dimensional data is respectively introduced into two two-layer convolutional networks of the two branch structures, the real data and the imaginary data of the image are respectively generated, the values ​​of the real part and the imaginary part are respectively calculated, and then the attenuation values ​​of the real part and the imaginary part are respectively calculated by the two-dimensional convolutional transposition layer and the reconstruction layer in the branch structure; the attenuation function calculates the values ​​of the real part and the imaginary part respectively and superimposes them as the selection of the fitting direction; The second configuration adopts a bifurcated network, which includes a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional zero-filling layer, two branch structures, a connection layer and a reconstruction layer in sequence, each branch structure includes two two-dimensional convolutional layers and one two-dimensional convolutional transposition layer; the two fully connected layers use the TanH function; the bifurcated network data processing process of the second configuration includes: the input data is processed by the TanH function in the one-dimensional fully connected network, and then the output of the one-dimensional fully connected network enters the two-dimensional zero-filling layer after the reconstruction layer and is re-transposed into two-dimensional data, the two-dimensional data is respectively introduced into the two bifurcated two-layer convolutional networks, the real data and the imaginary data of the image are respectively generated, and the values ​​of the real part and the imaginary part are respectively calculated; the real data and the imaginary data enter the connection layer and the reconstruction layer at the same time, and are integrated as one data to calculate the attenuation value; The third configuration includes, in sequence, a one-dimensional fully connected network consisting of two fully connected layers, a reconstruction layer, a two-dimensional stack filling layer, two two-dimensional convolutional layers and two branch structures, one of which includes, in sequence, a two-dimensional convolutional transposition layer and a reconstruction layer, and the other branch structure is a two-dimensional convolutional layer; the two fully connected layers use the TanH function; the data processing process of the third configuration includes: the input data is processed using the TanH function in a one-dimensional fully connected network, the output of the one-dimensional fully connected network is re-transposed into two-dimensional data, the two-dimensional data is simultaneously generated through a double-layer convolutional network. Real data and imaginary data of the image are integrated, and the attenuation values ​​of the real data and the imaginary data are calculated.

Images