WO2020134826A1 - Parallel magnetic resonance imaging method and related equipment - Google Patents

Abstract

A parallel magnetic resonance imaging method, comprising: obtaining undersampled image data acquired by magnetic resonance coils of multiple parallel channels in K-space (S101); obtaining a pre-built comprehensive neural network model, wherein the comprehensive neural network model comprises interconnected neural networks and a K-space-consistent layer (S102); inputting the undersampled image data of each channel into the comprehensive neural network model, so that the neural networks may initially reconstruct the undersampled image data and the space-consistent layer performs back substitution processing on an initial reconstruction result, so as to obtain target reconstruction image data (S103); and merging the target reconstruction image data of all channels to obtain image data of a single channel (S104). Hence, the present method adds the K-space-consistent layer to the neural network models, and said layer uses the undersampled image data before reconstruction to perform back substitution processing on the reconstructed image data, thereby retaining more image details and improving an image reconstruction effect. Also provided is related equipment for parallel magnetic resonance imaging.

Description

Parallel magnetic resonance imaging method and related equipment Technical field

The invention relates to the technical field of magnetic resonance image processing, and more specifically, to a magnetic resonance parallel imaging method and related equipment.

Background technique

The magnetic resonance imaging system uses static magnetic fields and radio frequency magnetic fields to image human tissues. It not only provides rich tissue contrast and does not cause side effects to the human body, so it has become a common tool for medical clinical diagnosis.

The magnetic resonance imaging system mainly includes two devices: an image acquisition device and a parallel imaging device. The image acquisition device includes multiple parallel channels. Different channels correspond to different acquisition coils. The different acquisition coils are located at different acquisition positions and are used to acquire image data of human tissue from different directions. The parallel imaging device uses the parallel imaging (PI) algorithm, that is, the image data collected by each parallel channel is reconstructed to obtain image data with better display effect.

Traditional parallel imaging algorithms are mainly divided into two categories: one based on K-space reconstruction, such as space harmonic synchronization acquisition (SMASH), generalized automatic calibration partial parallel acquisition (GRAAPA), self-consistent parallel imaging (SPIRiT), etc.; An image-based reconstruction, such as sensitivity encoding (SENSitivity Encoding, SENSE), etc. In addition, the emergence of compressed sensing has also greatly improved and formed image reconstruction, among which typical reconstruction algorithms are L1-SPIRiT and so on. In recent years, deep learning has been well used in fast magnetic resonance reconstruction. At present, there are two main methods of deep learning for parallel imaging, one is the use of multi-layer perceptron (MLP) for parallel imaging, and the other is A Variational Network (VN) network that networks traditional iterative algorithms.

However, for images reconstructed by existing parallel imaging algorithms, the image detail processing effect needs to be further improved.

Summary of the invention

In view of this, the present invention provides a parallel magnetic resonance imaging method to further improve the detail processing effect of the reconstructed image. In addition, the present invention also provides magnetic resonance parallel imaging related equipment to ensure the practical application and implementation of the method.

To achieve the stated objective, the technical solutions provided by the present invention are as follows:

In the first aspect, the present application provides a parallel magnetic resonance imaging method, including:

Obtain under-sampled image data acquired in parallel by K-space for multiple channels of magnetic resonance coils;

Obtain a pre-built neural network integrated model; wherein the neural network integrated model includes interconnected neural networks and a K-space consistent layer;

Input the undersampled image data of each channel into the neural network integrated model, so that the neural network integrated model performs the following steps:

The neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data to obtain preliminary reconstructed image data corresponding to each channel;

The K-space consistent layer performs the generation processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel, to obtain the target reconstructed image data corresponding to each channel;

The target reconstructed image data of all channels is merged to obtain single-channel image data.

In a second aspect, the present application provides a parallel magnetic resonance imaging device, including:

An image data acquisition module for acquiring undersampled image data acquired in parallel by K-space magnetic resonance coils of multiple channels;

A comprehensive model obtaining module, used to obtain a pre-built neural network comprehensive model; wherein the neural network comprehensive model includes interconnected neural networks and a K-space consistent layer;

The image data reconstruction module is used to input the undersampled image data of each channel into the neural network integrated model, so that the neural network integrated model performs the following steps:

The neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data to obtain preliminary reconstructed image data corresponding to each channel;

The K-space consistent layer performs the generation processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel, to obtain the target reconstructed image data corresponding to each channel;

The image data merging module is used to merge the target reconstruction image data of all channels to obtain single channel image data.

In a third aspect, the present application provides a parallel magnetic resonance imaging device, including:

The receiver is used to receive the under-sampled image data collected by the parallel multiple channel magnetic resonance coils in the K space;

A processor, used to obtain a pre-built neural network integrated model; wherein the neural network integrated model includes interconnected neural networks and a K-space consistent layer; input the undersampled image data of each channel into the neural network integrated model , So that the comprehensive model of the neural network performs the following steps: the neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data, and obtains the preliminary corresponding to each channel Reconstructed image data; the K-space consistent layer performs back-processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel to obtain the target reconstructed image data corresponding to each channel; merges the target reconstruction of all channels Image data to get single-channel image data;

A display is used to display the single-channel image data.

According to a fourth aspect, the present application provides a readable storage medium on which a computer program is stored, which is characterized in that, when the computer program is executed by a processor, the above-mentioned parallel magnetic resonance imaging method is realized.

It can be seen from the above technical solutions that the present invention provides a parallel magnetic resonance imaging method, which obtains undersampled image data acquired in parallel by the coils of multiple channels in K-space, and inputs the undersampled image data of each channel to the A comprehensive model of the constructed neural network. In addition to the neural network, the model also includes a K-space consistent layer. The neural network can reconstruct the undersampled image data. The K-space consistent layer can use the undersampled image data to reconstruct the image data of the neural network. After performing the generation process, the target reconstructed image data is obtained, and finally, the target reconstructed image data of all channels are merged to obtain a single-channel image. Compared with the prior art, the present invention adds a K-space consistent layer to the neural network model, which can use the under-sampled image data before reconstruction to perform post-processing on the reconstructed image data, thereby retaining more images The details improve the image reconstruction effect.

BRIEF DESCRIPTION

In order to more clearly explain the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only This is an embodiment of the present invention. For those of ordinary skill in the art, without paying any creative labor, other drawings may be obtained according to the provided drawings.

FIG. 1 is a schematic flowchart of a parallel magnetic resonance imaging method provided by the present invention;

2 is a schematic structural diagram of a neural network integrated module provided by the present invention;

3 is a schematic flowchart of a training convolutional neural network provided by the present invention;

4 is a comparison diagram of the imaging effect of the magnetic resonance parallel imaging method provided by the present invention and the traditional parallel imaging algorithm for reconstruction;

5 is a schematic structural diagram of a magnetic resonance parallel imaging device provided by the present invention;

6 is another schematic structural diagram of a magnetic resonance parallel imaging device provided by the present invention;

7 is a schematic diagram of a computer architecture of a magnetic resonance parallel imaging device provided by the present invention.

detailed description

The technical solutions in the embodiments of the present invention will be described clearly and completely 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, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present invention.

The magnetic resonance imaging system is a system commonly used in the medical field for imaging human tissue. The system mainly includes two devices, an image acquisition device and a parallel imaging device.

The image acquisition device includes multiple parallel channels, and each channel uses different positions of magnetic resonance coils to acquire image data of human tissue from different directions. Slow imaging speed has always been a major bottleneck restricting the application of this system, so it is particularly important to increase the scanning speed and reduce the scanning time under the premise that the imaging quality is clinically acceptable.

To achieve this, magnetic resonance imaging systems generally use parallel imaging technology, in which the image collected by the image acquisition device is undersampled image data. Specifically, in an image acquisition device, multiple coils simultaneously acquire image data during one scan, and each coil only acquires image data of a certain local position of the object. These image data are undersampled image data. At the same time, the parallel imaging device uses a parallel imaging algorithm to reconstruct the undersampled image data collected in each channel. It can be seen that parallel imaging mainly uses the sensitivity information collected by the coils at different positions to reconstruct the image. It can achieve the purpose of accelerating scanning through a certain degree of undersampling, which is one of the important methods in fast magnetic resonance imaging.

The invention provides a parallel magnetic resonance imaging method, which is used in a parallel imaging device and used to improve the detail processing effect of a reconstructed image. See FIG. 1, which shows a schematic flow diagram of a parallel magnetic resonance imaging method, which specifically includes steps S101 to S104.

S101: Obtain under-sampled image data acquired in parallel by the magnetic resonance coils of multiple channels in K-space.

In practical applications, when it is necessary to detect the condition of human tissues such as the head, the image data of the human tissues is collected by the image acquisition device first. As mentioned above, the image acquisition device includes multiple parallel channels, and the magnetic resonance coil of each channel uses the under-sampling method to collect local information of the human tissue, so that each channel will obtain an under-sampled image data. The number of channels is related to the actual image acquisition device, such as 6 channels, 8 channels, 12 channels, 32 channels, etc.

It should be noted that K-space is a domain that represents images and is also the image data collection domain, that is, image data is collected in K-space. Undersampling image data is data collected by undersampling in the image K-space. The image data in K-space is obtained by Fourier transform. The image data in the K-space domain is not an image that can be recognized by the human eye. It needs to undergo inverse Fourier transform to obtain an image that can be recognized by the human eye.

S102: Obtain a pre-built neural network integrated model; where the neural network integrated model includes interconnected neural networks and a K-space consistent layer.

Among them, the neural network synthesis model is pre-built. Unlike the existing neural network model, the neural network synthesis model constructed by the present invention includes two parts, one is the neural network and the other is the K-space consistency layer (Data Consistency, DC). The neural network needs to be pre-trained by the neural network training algorithm. The role of the K-space consistent layer is to make full use of the collected K-space data information in the image data to correct the image reconstructed by the neural network, so that the reconstructed image is more accurate and retained. More image details.

The inventor found through research that the traditional parallel imaging algorithm is often greatly affected by the number and arrangement of coils, and the reconstruction result has an amplification effect on noise, and the acceleration factor is also limited. Compressed sensing-based parallel imaging algorithm (CS-PI) only has good reconstruction effect for non-coherent sampling sampling mode, poor reconstruction result for one-dimensional undersampling aliasing artifacts, and the weight parameters in the objective function It is more difficult to adjust, while the iterative solution process takes a little longer.

In the parallel imaging method based on deep learning, the multi-layer perceptron is a fully connected neural network, which requires a large amount of parameters, and this method needs to be reconstructed line by line, focusing on solving the aliasing artifacts caused by one-dimensional undersampling. The sampling method has certain restrictions. VN network is one of the methods to network traditional algorithms. It needs to calculate the sensitivity information of the coil array in advance, so the reconstruction result depends on the accuracy of the coil sensitivity estimation.

Therefore, optionally, the neural network selected in the present invention is specifically a convolutional neural network (Convolutional Neural Network, CNN). Compared with other neural networks such as multilayer perceptrons, convolutional neural networks have the characteristics of local connection and weight sharing, and have more unique advantages in the field of computer vision. More specifically, convolutional neural network CNN has advantages in image processing. It is suitable for both one-dimensional and two-dimensional undersampling, and is not affected by the number of traditional auto-calibration signal (ACS) lines. The invention uses the convolutional neural network CNN for parallel magnetic resonance imaging, so that the invention has no special limitation on the sampling mode and is not sensitive to the ACS line.

In the comprehensive model of neural network, the connection between the neural network and the K-space consistent layer can be a multi-layer cascade. For example, a specific structure of a comprehensive model of a neural network is shown in Figure 2. As shown in FIG. 2, the comprehensive model of the neural network includes N sub-modules Block. The N sub-modules are connected back and forth, and the image reconstruction result output by the previous sub-module is input to the latter sub-module to continue the reconstruction process. It should be noted that N is a known value obtained in advance based on training experience. In addition, the sub-module may be referred to as a residual module.

Each submodule includes a convolutional neural network CNN and a K-space consistent layer DC connected to each other. The K-space consistent layer is connected after the convolutional neural network CNN to ensure that the image reconstruction results of the convolutional neural network CNN are more For accuracy. In the training process, it can be seen that the structure of cascading neural networks connected to the K-space consistent layer can make the reconstructed image closer to the label image (that is, fully sampled image data) during the forward propagation process.

A specific structure of the convolutional neural network includes M superposed complex convolutional layers, M is a known value obtained in advance according to training experience.

S103: Input the under-sampled image data of each channel into the neural network integrated model, so that the neural network performs preliminary reconstruction on the under-sampled image data and the spatial reconstruction layer to perform preliminary processing on the preliminary reconstruction results to obtain target reconstructed image data.

Specifically, the undersampled image data of each channel is input into the neural network integrated model, so that the neural network integrated model executes steps A1 and A2.

A1: The neural network reconstructs each under-sampled image data separately according to the correlation between each under-sampled image data to obtain the preliminary reconstructed image data corresponding to each channel; A2: The K-space consistent layer is based on each channel’s Undersampling the image data, the preliminary reconstructed image data of each channel is back-processed to obtain the target reconstructed image data corresponding to each channel.

Specifically, the input data of the comprehensive model of the neural network is undersampled image data of multiple parallel channels obtained in step S101. The processing of the neural network integrated model is divided into two parts: in step A1, the neural network first performs preliminary reconstruction on the undersampled image data. During the reconstruction process, the correlation of the image data between multiple coils is fully utilized, and the resulting image is called the preliminary reconstructed image Data; in step A2, the K-space consistent layer uses the undersampled image data input to the neural network synthesis model to perform the retrograde processing on the preliminary reconstructed image data. data.

The generation process is: in the preliminary reconstructed image data after reconstruction, if the pixel value of a pixel at a certain position does not exist in the undersampled image data corresponding to the preliminary reconstructed image data, the pixel value is retained; if a certain The pixel values of the pixels at each position exist in the under-sampled image data corresponding to the preliminary reconstructed image data, and then the pixel values in the under-sampled image data are replaced with the pixel values in the preliminary reconstructed image data.

More specifically, one way to go back is to perform weighted processing on the pixel values in the undersampled image data and the pixel values in the preliminary reconstructed image data, and replace the pixel values in the preliminary reconstructed image data with the processed pixel values; another One processing method is to directly replace the pixel values in the under-sampled image data with the pixel values in the preliminary reconstructed image data.

The K-space consistent layer of the first iteration way is expressed by a specific formula, then the K-space consistent layer is expressed as:

Where S is the set of all sampled pixel positions of the undersampled image data of the jth channel; k is the position of any pixel in the preliminary reconstructed image data output by the comprehensive model of the neural network; f _l,j (k) is the jth The pixel value of the K space of the preliminary reconstructed image data of each channel at position k; λ is the preset weighted weight; f _0,j (k) is the K space of the undersampled image data input from the jth channel at position k Pixel values.

Using a specific formula to express the K-space consistent layer of the second generation, the K-space consistent layer is expressed as:

As for why the above generation processing is required, it is because the reconstruction effect of the undersampled image data by the neural network is not good enough, and the effect is insufficient for two reasons: one is that the neural network for learning and training may not be very accurate, and the second is undersampled image data These pixels already existing in the image will be distorted after being reconstructed by the neural network. After using the under-sampled image data that has not undergone the reconstruction process to perform the above-mentioned generation processing on the preliminary reconstructed image data, the reconstruction effect of the preliminary reconstructed image data can be improved.

It should be noted that the image data processed by the above K-space consistent layer are all image data in K-space. In addition, this step is to process the undersampled image data of each channel separately to obtain the target reconstructed image data corresponding to each channel respectively.

S104: Combine target reconstruction image data of all channels to obtain single-channel image data.

Among them, the target reconstructed image data of all channels can be combined through adaptive channel merging or square sum merging, etc., to obtain single-channel image data.

It can be seen from the above technical solutions that the present invention provides a parallel magnetic resonance imaging method, which obtains undersampled image data acquired in parallel by the coils of multiple channels in K-space, and inputs the undersampled image data of each channel to the A comprehensive model of the constructed neural network. In addition to the neural network, the model also includes a K-space consistent layer. The neural network can reconstruct the undersampled image data. The K-space consistent layer can use the undersampled image data to reconstruct the image data of the neural network. After performing the generation process, the target reconstructed image data is obtained, and finally, the target reconstructed image data of all channels are merged to obtain a single-channel image. Compared with the prior art, the present invention adds a K-space consistent layer to the neural network model, which can use the under-sampled image data before reconstruction to perform post-processing on the reconstructed image data, thereby retaining more images The details improve the image reconstruction effect.

Furthermore, the present invention designs a novel cascaded multi-channel network based on Convolutional Neural Network (CNN) for parallel magnetic resonance imaging. The learning of the image domain and the consistency of the collected K-space data are considered in the network structure, which can progressively improve the image quality in the cascade network without losing the information of the collected data. At the same time, the present invention adopts a complex convolution operation for rationally processing real and imaginary parts of complex magnetic resonance data.

The following specifically describes the training process of the neural network in the neural network synthesis model.

Among them, the neural network is specifically a convolutional neural network. Referring to FIG. 3, a specific training process of the convolutional neural network includes the following steps S301 to S304.

S301: Obtain the fully sampled image data of each channel.

Among them, the full-sampled image data is used to make the under-sampled image data during training, so the full-sampled image data is obtained first, and the full-sampled data is used as the label data in the training process of the convolutional neural network.

For example, you can use a 12-channel head coil on a 3T magnetic resonance instrument (SIEMENS MANGETOM Trio Tim) to acquire more than 3000 two-dimensional magnetic resonance brain images, which can include sagittal, coronal, and cross-sectional images, or you can mix T1 Weighted, T2-weighted, and PD-weighted images. More than 3000 two-dimensional magnetic resonance brain images are fully sampled image data. It should be noted that a two-dimensional magnetic resonance brain image contains fully sampled image data acquired by each of 12 channels, that is, 3000*12 pieces of fully sampled image data are obtained.

It should be noted that big data is very important for deep learning, so if you want to get better training results, you must ensure a certain amount of data. When the amount of original data is not large enough, data can be augmented by data augmentation. In the present invention, for each original image, by rotating 90°, 180°, 270°, and mirroring and flipping along the x-axis and y-axis, the data amount is 8 times that of the original data. In other words, the above expanded data operation is performed on the full-sampled data of each channel to obtain 3000*12*8 full-sampled image data, where each channel corresponds to 3000*8 full-sampled image data. Of course, there may be other data expansion methods, which are not limited to the above methods.

It should be noted that the expanded fully sampled image data is also used as tag data.

S302: Undersampling the fully sampled image data of each channel to obtain the undersampled image data samples of each channel.

This step can be referred to as the production step of undersampled image data.

Specifically, the undersampling processing is retrospective undersampling processing, which can be expressed as: X _u =F ^H P ^H PFX; where, X _u is an undersampled image data sample, and F represents a Fourier coding matrix (after Fourier coding after the transformation matrix, can be converted to the undersampled image data space K), P represents the channel mode sampling cycle undersampling diagonal matrix, and P ^{H H} F. superscript H indicates the Hermitian transpose operation , X represents the label data.

Undersampling processing is the process of simulating the acquisition of undersampling image data. The full sampling image data of each channel must be subjected to the above undersampling processing to obtain the undersampling image data corresponding to each full sampling image data. Since the obtained under-sampled image data is used as a training sample for training the convolutional neural network, it can be called an under-sampled image data sample.

It should be noted that before performing step S303 using the convolutional neural network training algorithm to train the undersampled image data samples, each undersampled image data can also be formulated according to the formula

Normalize the modulus.

Where X _k represents the modulus of the pixel value of the pixel at position k after the normalization of the undersampled image data; I _k represents the modulus of the pixel value of the pixel at the position k of the undersampled image data; max{abs (X)} represents the maximum modulus of all pixel values of undersampled image data.

Specifically, the under-sampled image data collected by the magnetic resonance coil is a complex image, that is, the pixel values of pixels in the image are of a complex type. Each pixel in the undersampled image data is subjected to the normalization of the modulus value to normalize the modulus value of each pixel in the undersampled image data to (0,1).

Between each under-sampled image data, the pixel value of the pixels may be relatively large, and the normalized processing of the modulus value can unify the magnitude of the pixel value of each under-sampled image data. For the comprehensive model of deep learning neural network, the normalization of the modulus value can ensure the effectiveness and convergence of training, which is a very important step in data processing. In step S303, training processing is performed on the undersampled image data that has undergone the normalization of the modulus.

S303: Use a convolutional neural network training algorithm to train the undersampled image data samples to obtain the weight parameter with the smallest loss value.

Among them, the convolutional neural network training algorithm is an existing training algorithm, and the present invention will not repeat them in detail. The convolutional neural network training process is the process of calculating the most suitable weight parameters in the convolutional neural network.

Specifically, the training process includes multiple forward propagation steps and multiple back propagation (BP) steps. It can be known that the weight parameters need to be set before training. The forward propagation step is to use the known weight parameters to calculate the output value of the convolutional neural network in the forward direction (the output value is the reconstruction data). In the propagation step, a loss function is used to compare the reconstructed data with the label data, and the weight parameters in the convolutional neural network are updated through the comparison result.

Backpropagation constantly updates the weight parameters by calculating the gradient of the loss function. The result of the update is that the difference between the output value of the convolutional neural network calculated by the loss function and the label data is the minimum value. It can be seen that the back propagation process needs to use a loss function. The function of the loss function is to compare the difference between the processing result of the undersampled image data of the convolutional neural network and the fully sampled image data corresponding to the undersampled image data. The value is the loss value in this step. The larger the loss value, the greater the gap between the two image data, which means that the reconstruction result of the convolutional neural network is not good enough, so the reconstruction result needs to be fed back to the convolutional neural network. If the loss value reaches the minimum value, the final value of the weight parameter can be determined, and the training process ends.

The back propagation process can be expressed as:

among them:

For the weight parameters obtained by training, J(x,y) is the loss function, x is the undersampled image data input to the convolutional neural network, and y is the fully sampled image data corresponding to the undersampled image data. It should be noted that the convolutional neural network is a multi-layer, each layer has corresponding weight parameters, so the weight parameters finally obtained

Includes multiple sets of values.

An L-layer convolutional neural network

It can be specifically expressed as follows:

Among them, C ₀ is the image data reconstructed by the first layer of convolutional neural network; X _u is the undersampled image data sample;

C _l is the reconstructed image data of the first layer convolutional neural network; σ _l is the nonlinear activation function; Ω _l is the convolution kernel with dimensions FW _l ×FH _l ×K _l-1 ×K _l , FW _l ×FH _l is the size of the convolution kernel of the layer 1 convolutional neural network, K _l-1 is the number of feature maps of the l-1 layer convolutional neural network, and K _l is the number of feature maps of the layer 1 convolutional neural network; C _l-1 is the image data reconstructed by the convolutional neural network of layer l-1; b _l is the offset of K _l dimension; L is the number of layers of the convolutional neural network;

C _L is the reconstructed image data of the Lth layer convolutional neural network; Ω _L is the convolution kernel of dimension FW _L ×FH _L ×K _L-1 ×K _L , and FW _L ×FH _L is the Lth layer convolutional nerve The size of the network convolution kernel, K _L-1 is the number of L-1 layer convolutional neural network feature maps, K _L is the Lth layer convolutional neural network feature maps; C _L-1 is the L-1th layer Image data reconstructed by the layer convolutional neural network; b _L is the offset of K _L dimension.

Convolutional Neural Network

Training, ie training weight parameters

the process of. among them

Including the value of L group Ω, b, namely

The specific loss function used in the present invention will be described below. The loss function will supervise the network training process, and different loss functions have different effects on the network training.

Specifically, the loss value is calculated by a loss function, and the loss function used in the present invention is an average absolute value error function

Where M is the number of undersampled image data input to the same channel in the convolutional neural network by one batch; m is the serial number of undersampled image data input to the same channel in the convolutional neural network by one batch; C(x _m ; θ) is the image data reconstructed by the convolutional neural network for the undersampled image data with sequence number m; y _m is the fully sampled image data corresponding to the undersampled image data with sequence number m.

It should be noted that, during the training process, input training samples may be performed in batches, that is, each input training sample is multiple training samples for each channel. For example, if the number of channels is 12, and a batch contains 4 training samples of one channel, the input training samples each time are 12*4 undersampled image data.

The loss function J(x, y) used in the present invention can be specifically an average absolute value error function (Mean Absolution Error, MAE). Compared with the loss function of mean square error, MAE can not only ensure that the training results can well suppress noise , And slightly better maintain image detail.

S304: Construct a convolutional neural network using weight parameters.

Among them, the weight parameters in the convolutional neural network are set to the final training

You can get the constructed convolutional neural network

Where x is the undersampled image data to be input,

It is the weight parameter of the convolutional neural network. The trained convolutional neural network can reconstruct any input undersampled image data in practical applications.

In order to test the feasibility and effectiveness of the neural network integrated model constructed by the present invention, the inventors carried out a comparative implementation. Specifically, for the same experimental sample data obtained by the one-dimensional random (1D random) 3 times undersampling method, the traditional parallel imaging algorithm and the algorithm provided by the present invention are used for the reconstruction process, and the reproduction effect is shown in FIG. 4.

As shown in Fig. 4, from left to right, the top and bottom images are: original image and sampling mode, SPIRiT reconstruction result and corresponding error map, L1-SPIRiT reconstruction result and corresponding error map, obtained using the algorithm of the present invention Reconstruction results and corresponding error graphs.

It can be clearly seen that the parallel magnetic resonance imaging method provided by the present invention is relatively thorough in removing aliasing artifacts caused by one-dimensional undersampling, and has the best noise suppression effect, which is closest to the original image. It can be seen that, compared with the traditional parallel imaging algorithm, the magnetic resonance parallel imaging method proposed by the present invention has obvious advantages.

See FIG. 5, which shows a structure of a magnetic resonance parallel imaging device provided by the present invention. As shown in FIG. 5, the magnetic resonance parallel imaging apparatus may include: an image data obtaining module 501, an integrated model obtaining module 502, an image data reconstruction module 503, and an image data merging module 504.

The image data obtaining module 501 is used to obtain the under-sampled image data acquired by the parallel multiple channel magnetic resonance coils in the K space;

The integrated model obtaining module 502 is used to obtain a pre-built neural network integrated model; wherein the neural network integrated model includes interconnected neural networks and a K-space consistent layer;

The image data reconstruction module 503 is used to input the undersampled image data of each channel into the neural network integrated model, so that the neural network integrated model performs the following steps:

The neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data to obtain preliminary reconstructed image data corresponding to each channel;

The K-space consistent layer performs the generation processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel, to obtain the target reconstructed image data corresponding to each channel;

The image data merging module 504 is used to merge the target reconstructed image data of all channels to obtain single channel image data.

In one implementation, the comprehensive model of the neural network includes a plurality of cascaded residual modules, and each residual module includes a connected neural network and a K-space consistent layer.

In one implementation, the neural network specifically includes a convolutional neural network.

Referring to FIG. 6, the present invention also provides another structure of a parallel magnetic resonance imaging device. As shown in FIG. 6, the magnetic resonance parallel imaging device may further include a neural network training module 505 based on the structure shown in FIG. 5. The neural network training module 505 is used to train the convolutional neural network in the integrated model of the neural network. Specifically, the neural network training module 505 includes:

Full sampling data acquisition sub-module, used to obtain the full sampling image data of each channel;

Undersampling data acquisition sub-module, used to undersampling the fully sampled image data of each channel to obtain the undersampled image data samples of each channel;

A weight parameter training sub-module, used to train the undersampled image data samples using a convolutional neural network training algorithm to obtain weight parameters with the smallest loss value;

Convolutional network determination sub-module for constructing a convolutional neural network using the weighting parameters

Where x is the undersampled image data to be input,

It is the weight parameter of the convolutional neural network.

In one implementation, the convolutional neural network

Specifically:

Among them, C ₀ is the image data reconstructed by the first layer of convolutional neural network; X _u is the undersampled image data sample;

C _l is the reconstructed image data of the first layer convolutional neural network; σ _l is the nonlinear activation function; Ω _l is the convolution kernel with dimensions FW _l ×FH _l ×K _l-1 ×K _l , FW _l ×FH _l is the size of the convolution kernel of the layer 1 convolutional neural network, K _l-1 is the number of feature maps of the l-1 layer convolutional neural network, and K _l is the number of feature maps of the layer 1 convolutional neural network; C _l-1 is the image data reconstructed by the convolutional neural network of layer l-1; b _l is the offset of K _l dimension; L is the number of layers of the convolutional neural network;

C _L is the reconstructed image data of the Lth layer convolutional neural network; Ω _L is the convolution kernel of dimension FW _L ×FH _L ×K _L-1 ×K _L , and FW _L ×FH _L is the Lth layer convolutional nerve The size of the network convolution kernel, K _L-1 is the number of L-1 layer convolutional neural network feature maps, K _L is the Lth layer convolutional neural network feature maps; C _L-1 is the L-1th layer Image data reconstructed by the layer convolutional neural network; b _L is the offset of K _L dimension.

In an implementation manner, the loss value is calculated by a loss function, and the loss function is an average absolute value error function

Where M is the number of undersampled image data input to the same channel in the convolutional neural network by one batch; m is the serial number of undersampled image data input to the same channel in the convolutional neural network by one batch; C(x _m ; θ) is the image data reconstructed by the convolutional neural network for the undersampled image data with sequence number m; y _m is the fully sampled image data corresponding to the undersampled image data with sequence number m.

In one implementation, the K-space consistent layer is specifically:

Where S is the set of all sampled pixel positions of the undersampled image data of the jth channel; k is the position of any pixel in the preliminary reconstructed image data output by the comprehensive model of the neural network; f _l,j (k) is the jth The pixel value of the K space of the preliminary reconstructed image data of each channel at position k; λ is the preset weighted weight; f _0,j (k) is the K space of the undersampled image data input from the jth channel at position k Pixel values.

In one implementation, the neural network training module 505 further includes: a module normalization processing sub-module.

Modular normalization processing sub-module, which is used to normalize the modular values of each of the undersampled image data according to the following formula before using the convolutional neural network training algorithm to train the undersampled image data samples:

Where X _k represents the modulus of the pixel value of the pixel at position k after the normalization of the undersampled image data; I _k represents the modulus of the pixel value of the pixel at the position k of the undersampled image data; max {abs(X)} represents the maximum modulus of all pixel values of undersampled image data.

See FIG. 7, which shows a magnetic resonance parallel imaging device provided by the present application, which specifically includes: a memory 701, a receiver 702, a processor 703, a display 704, and a communication bus 705.

Among them, the memory 701, the receiver 702, the processor 703, and the display 704 communicate with each other through the communication bus 705.

The memory 701 is used to store programs; the memory 501 may include a high-speed RAM memory, or may further include a non-volatile memory (non-volatile memory), for example, at least one magnetic disk memory.

The receiver 702 is used to connect with the image acquisition device, and is used to receive the under-sampled image data acquired in the K-space by the magnetic resonance coils of multiple parallel channels in the image acquisition device.

The processor 703 is configured to execute a program. The program may include a program code, and the program code includes an operation instruction of the processor. Among them, the program can be specifically used for:

Obtain a pre-built neural network integrated model; wherein the neural network integrated model includes interconnected neural networks and a K-space consistent layer; input the undersampled image data of each channel into the neural network integrated model, so that the The comprehensive model of the neural network performs the following steps: the neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data to obtain the preliminary reconstructed image data corresponding to each channel; K The spatially consistent layer performs back-processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel to obtain the target reconstructed image data corresponding to each channel; the target reconstructed image data of all channels is merged to obtain Single-channel image data.

The processor 703 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

It should be noted that the processor may execute various steps related to the above-mentioned parallel magnetic resonance imaging method, which is not repeated here.

The display 704 is used to display the single-channel image data.

The present application also provides a readable storage medium on which a computer program is stored, and the computer program can be executed by a processor to implement various steps in a method of processing virtual asset data.

It should be noted that the embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments refer to each other. can.

It should also be noted that in this article, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply these entities or operations There is any such actual relationship or order. Moreover, the terms "include", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also those not explicitly listed Or other elements that are inherent to this process, method, article, or equipment. Without more restrictions, the elements defined by the sentence "include one..." do not exclude that there are other identical elements in the process, method, article or equipment that includes the above elements.

The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims (11)

1. A parallel magnetic resonance imaging method, characterized in that it includes:

   Obtain under-sampled image data acquired in parallel by K-space for multiple channels of magnetic resonance coils;

   Obtain a pre-built neural network integrated model; wherein the neural network integrated model includes interconnected neural networks and a K-space consistent layer;

   Input the undersampled image data of each channel into the neural network integrated model, so that the neural network integrated model performs the following steps:

   The neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data to obtain preliminary reconstructed image data corresponding to each channel;

   The K-space consistent layer performs the generation processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel, to obtain the target reconstructed image data corresponding to each channel;

   The target reconstructed image data of all channels is merged to obtain single-channel image data.
2. The magnetic resonance parallel imaging method according to claim 1, wherein the neural network synthesis model includes a plurality of cascaded residual modules, each residual module includes a connected neural network and a K-space consistent layer .
3. The parallel magnetic resonance imaging method of claim 1, wherein the neural network specifically comprises a convolutional neural network.
4. The parallel magnetic resonance imaging method of claim 3, wherein the neural network training process in the neural network synthesis model includes:

   Obtain the fully sampled image data of each channel;

   Undersampling the fully sampled image data of each channel to obtain the undersampled image data samples of each channel;

   Use a convolutional neural network training algorithm to train the undersampled image data samples to obtain the weight parameter with the smallest loss value;

   Use the weight parameters to construct a convolutional neural network

   

   Where x is the undersampled image data to be input,

   

   It is the weight parameter of the convolutional neural network.
5. The parallel magnetic resonance imaging method of claim 4, wherein:

   The convolutional neural network

   

   Specifically:

   

   Among them, C ₀ is the image data reconstructed by the first layer of convolutional neural network; X _u is the undersampled image data sample;

   C _l is the reconstructed image data of the first layer convolutional neural network; σ _l is the nonlinear activation function; Ω _l is the convolution kernel with dimensions FW _l ×FH _l ×K _l-1 ×K _l , FW _l ×FH _l is the size of the convolution kernel of the layer 1 convolutional neural network, K _l-1 is the number of feature maps of the l-1 layer convolutional neural network, and K _l is the number of feature maps of the layer 1 convolutional neural network; C _l-1 is the image data reconstructed by the convolutional neural network of layer l-1; b _l is the offset of K _l dimension; L is the number of layers of the convolutional neural network;

   C _L is the reconstructed image data of the Lth layer convolutional neural network; Ω _L is the convolution kernel of dimension FW _L ×FH _L ×K _L-1 ×K _L , and FW _L ×FH _L is the Lth layer convolutional nerve The size of the network convolution kernel, K _L-1 is the number of L-1 layer convolutional neural network feature maps, K _L is the Lth layer convolutional neural network feature maps; C _L-1 is the L-1th layer Image data reconstructed by the layer convolutional neural network; b _L is the offset of K _L dimension.
6. The parallel magnetic resonance imaging method of claim 4, wherein:

   The loss value is calculated by a loss function, and the loss function is an average absolute value error function

   

   Where M is the number of undersampled image data input to the same channel in the convolutional neural network by one batch; m is the serial number of undersampled image data input to the same channel in the convolutional neural network by one batch; C(x _m ; θ) is the image data reconstructed by the convolutional neural network for the undersampled image data with sequence number m; y _m is the fully sampled image data corresponding to the undersampled image data with sequence number m.
7. The parallel magnetic resonance imaging method of claim 1, wherein the K-space consistent layer is specifically:

   

   Where S is the set of all sampled pixel positions of the undersampled image data of the jth channel; k is the position of any pixel in the preliminary reconstructed image data output by the comprehensive model of the neural network; f _l,j (k) is the jth The pixel value of the K space of the preliminary reconstructed image data of each channel at position k; λ is the preset weighted weight; f _0,j (k) is the K space of the undersampled image data input from the jth channel at position k Pixel values.
8. The parallel magnetic resonance imaging method according to claim 4, wherein, before using the convolutional neural network training algorithm to train the undersampled image data samples, the method further comprises:

   Normalize each of the under-sampled image data according to the following formula:

   

   Where X _k represents the modulus of the pixel value of the pixel at position k after the normalization of the undersampled image data; I _k represents the modulus of the pixel value of the pixel at the position k of the undersampled image data; max {abs(X)} represents the maximum modulus of all pixel values of undersampled image data.
9. A magnetic resonance parallel imaging device is characterized by comprising:

   An image data acquisition module for acquiring undersampled image data acquired in parallel by K-space magnetic resonance coils of multiple channels;

   A comprehensive model obtaining module, used to obtain a pre-built neural network comprehensive model; wherein the neural network comprehensive model includes interconnected neural networks and a K-space consistent layer;

   The image data reconstruction module is used to input the undersampled image data of each channel into the neural network integrated model, so that the neural network integrated model performs the following steps:

   The neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data to obtain preliminary reconstructed image data corresponding to each channel;

   The K-space consistent layer performs the generation processing on the preliminary reconstructed image data of each channel according to the undersampled image data of each channel, to obtain the target reconstructed image data corresponding to each channel;

   The image data merging module is used to merge the target reconstruction image data of all channels to obtain single channel image data.
10. A parallel magnetic resonance imaging device, characterized in that it includes:

    The receiver is used to receive the under-sampled image data collected by the parallel multiple channel magnetic resonance coils in the K space;

    A processor for obtaining a pre-built neural network integrated model; wherein the neural network integrated model includes interconnected neural networks and a K-space consistent layer; input the undersampled image data of each channel into the neural network integrated model , So that the comprehensive model of the neural network performs the following steps: the neural network reconstructs each of the under-sampled image data according to the correlation between each of the under-sampled image data, and obtains the preliminary corresponding to each channel Reconstructed image data; the K-space consistent layer reprocesses the preliminary reconstructed image data of each channel according to the undersampled image data of each channel to obtain the target reconstructed image data corresponding to each channel; merges the target reconstruction of all channels Image data to get single-channel image data;

    A display is used to display the single-channel image data.
11. A readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the magnetic resonance parallel imaging method according to any one of claims 1-8 is realized.

Images