A fast magnetic resonance reconstruction method and system based on a multi-level wavelet convolution network

Abstract

The application provides a fast magnetic resonance reconstruction method and system based on a multi-level wavelet convolution network, and particularly relates to the technical field of magnetic resonance imaging and image reconstruction. The scheme comprises the following steps: acquiring multi-channel undersampled K-space data; performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain an undersampled magnetic resonance image; inputting the undersampled magnetic resonance image into a trained multi-level wavelet convolution network for reconstruction to obtain a reconstructed magnetic resonance image, wherein the multi-level wavelet convolution network is constructed based on multi-level wavelet transform and a convolutional neural network. The scheme can shorten the scanning time by using an undersampling mode to acquire magnetic resonance K-space data, and the multi-level wavelet convolution network can effectively retain the high-frequency details of the image in the reconstruction process, thereby significantly improving the reconstruction quality and efficiency of the multi-channel magnetic resonance image as a whole.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of magnetic resonance imaging and image reconstruction, and particularly relates to a fast magnetic resonance reconstruction method and system based on a multi-level wavelet convolution network. BACKGROUND

[0002] Magnetic resonance imaging (MRI) is one of the important imaging technologies in modern medicine and has very important clinical value. In the MRI imaging process, a static magnetic field, a gradient magnetic field and a specific frequency radio frequency (RF) pulse magnetic field are applied to make the imaging object produce a magnetic resonance (MR) signal associated with the spatial position.

[0003] From the perspective of magnetic resonance imaging, the phase encoding step in traditional magnetic resonance imaging must be strictly executed according to the imaging sequence order by step-by-step conversion of the magnetic field gradient, which restricts the data acquisition speed. In conventional scanning, the magnetic resonance scanning of a single part takes tens of minutes or even 1-2 hours, which may produce motion artifacts in imaging tasks with higher real-time performance, such as heart, perfusion and functional imaging, etc., and the patient keeps relatively still in the closed device for a long time, which is easy to cause involuntary movement, resulting in lower imaging quality.

[0004] Therefore, fast magnetic resonance reconstruction has become a research hotspot. Although the fast magnetic resonance reconstruction methods based on deep learning such as ADMM-Net and MoDL have achieved certain results, the reconstruction of high-frequency details, local textures and microstructures of the image is often too blurred or smooth due to the influence of the traditional down-sampling structure and the design defects of the reconstruction loss function, which affects the overall accuracy of the reconstructed image. Therefore, there is an urgent need for a fast magnetic resonance reconstruction method that can meet the high-precision imaging in clinical practice. SUMMARY

[0005] In view of the above deficiencies of the prior art, the purpose of the present application is to provide a fast magnetic resonance reconstruction method and system based on a multi-level wavelet convolution network, which aims to solve the problems of the reconstructed magnetic resonance image being too smooth and the reconstruction accuracy being not high in the prior art.

[0006] In order to achieve the above-mentioned purpose, the first aspect of the present application provides a fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network, comprising:

[0007] acquiring multi-channel undersampled K-space data;

[0008] performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain an undersampled magnetic resonance image;

[0009] inputting the undersampled magnetic resonance image into the trained multi-level wavelet convolutional network to perform reconstruction, to obtain a reconstructed magnetic resonance image, the multi-level wavelet convolutional network being constructed based on a multi-level wavelet transform and a convolutional neural network.

[0010] Optionally, the trained multi-level wavelet convolutional network comprises a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network, and the inputting the undersampled magnetic resonance image into the trained multi-level wavelet convolutional network to perform reconstruction, to obtain a reconstructed magnetic resonance image, comprises:

[0011] performing feature extraction and decomposition on the undersampled magnetic resonance image by using the multi-level wavelet decomposition network, to obtain a feature map;

[0012] performing reconstruction on the feature map by using the multi-level wavelet reconstruction network, to obtain a reconstructed magnetic resonance image.

[0013] Optionally, the multi-level wavelet decomposition network is composed of a plurality of encoding convolutional layers and wavelet decomposition layers stacked alternately, and the performing feature extraction and decomposition on the undersampled magnetic resonance image by using the multi-level wavelet decomposition network, to obtain a feature map, comprises:

[0014] performing feature extraction on the undersampled magnetic resonance image by using the encoding convolutional layer, to obtain an initial encoding feature map;

[0015] performing wavelet decomposition on the initial encoding feature map by using the wavelet decomposition layer, to obtain an initial encoding feature frequency component, and inputting the initial encoding feature frequency component into the next encoding convolutional layer and the wavelet decomposition layer for iterative processing in turn, to obtain a feature map.

[0016] Optionally, the multi-level wavelet reconstruction network is composed of a plurality of decoding convolutional layers and wavelet reconstruction layers stacked alternately, and the performing reconstruction on the feature map by using the multi-level wavelet reconstruction network, to obtain a reconstructed magnetic resonance image, comprises:

[0017] performing analysis on the feature map by using the decoding convolutional layer, to obtain a plurality of initial decoding feature frequency components;

[0018] performing wavelet reconstruction on all the initial decoding feature frequency components by using the wavelet reconstruction layer, to obtain an initial decoding feature map, and inputting the initial decoding feature map into the next decoding convolutional layer and the wavelet reconstruction layer for iterative processing in turn, to obtain a reconstructed magnetic resonance image.

[0019] Optionally, the training process of the multi-level wavelet convolutional network comprises:

[0020] obtaining a plurality of groups of undersampled magnetic resonance image samples and corresponding fully sampled magnetic resonance images based on a fully sampled K-space data set;

[0021] constructing a multi-level wavelet convolution network and initializing network parameters, the multi-level wavelet convolution network being constructed by a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network;

[0022] inputting the undersampled magnetic resonance image sample into the multi-level wavelet convolution network to obtain a network reconstruction image;

[0023] calculating a reconstruction loss based on the network reconstruction image and the corresponding fully sampled magnetic resonance image;

[0024] iteratively optimizing the parameters of the multi-level wavelet convolution network using a back propagation algorithm until the reconstruction loss reaches a preset loss threshold or the number of iterations reaches a preset iteration number threshold, to obtain a trained multi-level wavelet convolution network.

[0025] Optionally, the process of calculating the reconstruction loss comprises:

[0026] calculating a reconstruction loss value of each pixel based on the network reconstruction image and the corresponding fully sampled magnetic resonance image;

[0027] calculating a unit loss drop probability based on the reconstruction loss value of each pixel;

[0028] obtaining a total number of pixels of the network reconstruction image, and calculating a loss drop probability of each pixel according to a preset loss drop ratio, the total number of pixels and the unit loss drop probability;

[0029] calculating a loss weight of each pixel based on the loss drop probability of each pixel;

[0030] Optionally, the process of calculating a unit loss drop probability based on the reconstruction loss value of each pixel comprises:

[0031] normalizing the loss value of each pixel using a preset normalization correction factor to obtain a unit loss drop probability, the normalization correction factor being used to adjust the difference in loss drop probability of different reconstruction difficulty pixels in the network reconstruction image.

[0032] The second aspect of the present application provides a fast magnetic resonance reconstruction system based on a multi-level wavelet convolution network, the system comprising:

[0033] an information acquisition module for acquiring multi-channel undersampled K-space data;

[0034] an undersampled image generation module for performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain an undersampled magnetic resonance image;

[0035] The image reconstruction module is configured to input the undersampled magnetic resonance image into the trained multi-level wavelet convolution network to reconstruct the undersampled magnetic resonance image and obtain a reconstructed magnetic resonance image.

[0036] The third aspect of the present application provides a magnetic resonance device, which comprises a memory configured to store executable instructions, and a processor configured to invoke and run the executable instructions in the memory to perform the steps of any one of the above-mentioned fast magnetic resonance reconstruction methods based on a multi-level wavelet convolution network.

[0037] The fourth aspect of the present application provides a computer readable storage medium, which stores program instructions, and when the program instructions are run by a processor, the steps of any one of the above-mentioned fast magnetic resonance reconstruction methods based on a multi-level wavelet convolution network are implemented.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] The present application adopts an undersampling method to collect magnetic resonance K-space data, which can shorten the scanning time and improve the imaging speed; the trained multi-level wavelet convolution network can extract multi-scale features in the undersampled magnetic resonance image and better protect the high-frequency details of the image, thus significantly improving the reconstruction quality and efficiency of the multi-channel magnetic resonance image. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0041] Figure 1 The flow chart of the fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network of the present application;

[0042] Figure 2 The schematic diagram of the multi-channel undersampled K-space data and the undersampled magnetic resonance image acquisition process of the present application;

[0043] Figure 3 The schematic diagram of the multi-level wavelet convolution network module of the present application;

[0044] Figure 4A module schematic diagram of a fast magnetic resonance reconstruction system based on a multi-level wavelet convolution network of the present application;

[0045] Figure 5 A structure schematic diagram of a magnetic resonance device of the present application. DETAILED DESCRIPTION

[0046] In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular sequences of steps, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods, devices, and circuits are omitted so as not to obscure the description of the present application with unnecessary detail.

[0047] It is to be understood that the terminology "including", "comprising", "consisting" and "consisting essentially of" used in the specification and the appended claims, indicates the presence of the stated features, integers, steps, operations, elements, and / or components but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0048] It is also to be understood that the terminology used in the present specification and the appended claims is for the purpose of describing the particular embodiments only and is not intended to be limiting.

[0049] It is further to be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, and that the term "at least one of" followed by a list of two or more items means any single one of the listed items individually, as well as any combination of two or more of the listed items.

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort belong to the scope of protection of the present application.

[0051] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details presented herein. In other instances, well-known methods, procedures, components, and circuits have not been described in detail as not to unnecessarily obscure aspects of the present application.

[0052] Magnetic resonance imaging (MRI) is one of the most important imaging techniques in modern medicine. Fast magnetic resonance reconstruction techniques can significantly improve the speed and quality of magnetic resonance imaging, and have very important clinical value. In recent years, the development of deep learning and its wide application in image processing has provided a new way for fast magnetic resonance reconstruction. On the one hand, deep learning networks allow network layers to extract more abstract and representative features without the need for manual design, so they often obtain more stable and higher quality results compared to traditional magnetic resonance reconstruction algorithms. On the other hand, the highly parallel computing and processing capabilities of deep learning networks make real-time magnetic resonance imaging possible, providing more timely and effective image information for clinical diagnosis. This is particularly important for applications such as cardiac and perfusion imaging that require fast and dynamic imaging. Therefore, fast magnetic resonance reconstruction based on deep learning has important clinical significance and brings new opportunities and prospects for the development of medical imaging.

[0053] Fast magnetic resonance reconstruction algorithms based on deep learning perform well in capturing global image features, but due to the limitations of traditional down-sampling structures and the design of reconstruction loss functions, the current mainstream deep learning reconstruction methods such as ADMM-Net and MoDL still have suboptimal performance in reconstructing high-frequency details. There are still certain limitations in reconstructing local textures and small structures, which may lead to blurred and overly smooth reconstructed images.

[0054] Based on this, the application provides a fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network, which first combines a multi-level wavelet transform with a convolution neural network to construct a multi-level wavelet convolution network; then uses paired input data and reference data to determine all parameters in the network structure through a back propagation algorithm; and finally reconstructs test data (i.e. undersampled magnetic resonance images) through a forward propagation process. In the training stage, the multi-level wavelet convolution network is input with paired undersampled multi-channel K-space data and corresponding fully sampled K-space data to train the network to establish an "end-to-end" mapping relationship, that is, the network can automatically and accurately map the input data to the reference data. In the test stage, only the test data needs to be provided as input to the network, and the trained multi-level wavelet convolution network can be used to perform the image reconstruction process. Since the present scheme uses an undersampling method to collect magnetic resonance K-space data, the scanning time can be shortened, which is conducive to improving the imaging speed; since the wavelet transform has reversibility, it can better avoid the problem of loss of detail information compared with the traditional downsampling operation (such as maximum pooling), so the present scheme can better protect the high-frequency details of the image and improve the image reconstruction quality; at the same time, the application proposes a new focal reconstruction loss function (FRL), which can adaptively adjust the reconstruction loss weight of different regions according to the loss value, so as to make the model focus more attention on the high-frequency region with higher reconstruction difficulty, thereby effectively improving the reconstruction performance of the multi-level wavelet convolution network for high-frequency details, and further improving the overall image reconstruction quality.

[0055] The embodiment of the application provides a fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network, which is deployed on a computer, a server or the like electronic device, applied to a scene of reconstructing a magnetic resonance image, and aims at the case of reconstructing an undersampled magnetic resonance image. Figures 1-3 As shown in the figure, the steps of the method of the embodiment include:

[0056] Step S100: collecting multi-channel undersampled K-space data;

[0057] Specifically, since under-sampling is a sparse sampling means of reducing the sampling frequency in the case that the bandwidth capacity of the test equipment is insufficient or in order to shorten the scanning time. In magnetic resonance imaging, multiple receiving coils are used to collect K-space partial data points simultaneously in an under-sampling manner. Since different coils correspond to different receiving channels, it is beneficial to improve the imaging speed and image quality. Therefore, in the process of data acquisition based on a multi-channel magnetic resonance system, first, the signal is selected by using a slice selection gradient field and a radio frequency excitation to determine the imaging layer and position, and the layer thickness and layer position are controlled by adjusting the strength and direction of the slice selection gradient field and the bandwidth and center frequency of the radio frequency field. Then, data is collected simultaneously by each receiving coil, and the collected data is filled into the K-space in the order of phase encoding and frequency encoding to form a two-dimensional complex matrix, i.e. multi-channel under-sampling K-space data K u , wherein the multi-channel under-sampling K-space data K u includes an under-sampling region and a full-sampling region. The under-sampling mode used in the embodiment can be a random under-sampling mode or an equal-interval under-sampling mode, as shown in Figure 2 , which is multi-channel under-sampling K-space data constructed by an equal-interval under-sampling mode. The data size can be represented as Nx*Ny*Nc, wherein Nx represents the number of rows of collected data, Ny represents the number of columns of data, and Nc represents the number of receiving channels.

[0058] Step S200: performing inverse Fourier transform on the under-sampling scan data to obtain an under-sampling magnetic resonance image;

[0059] Specifically, the two-dimensional Fourier transform (2D FFT) is performed on the collected multi-channel under-sampling K-space data K u to convert it into an under-sampling magnetic resonance image I u .

[0060] Step S300: inputting the under-sampling magnetic resonance image into a trained multi-level wavelet convolution network for reconstruction to obtain a reconstructed magnetic resonance image, wherein the multi-level wavelet convolution network is constructed based on a convolutional neural network and a multi-level wavelet transform.

[0061] In the training phase, a large number of high-quality full-sampling magnetic resonance image samples and corresponding under-sampling magnetic resonance image samples are needed as training data. The network parameters are optimized through a back propagation algorithm to enable the multi-level wavelet convolution network to learn the ability to restore high-quality images from under-sampling magnetic resonance images. In the inference phase, the under-sampling magnetic resonance image is directly input into the trained multi-level wavelet convolution network, and the reconstructed high-quality magnetic resonance image can be obtained through network prediction.

[0062] In this embodiment, the K-space data of magnetic resonance is collected in an undersampling manner, which can shorten the scanning time and improve the imaging speed; the trained multi-level wavelet convolution network can reconstruct the undersampling magnetic resonance image into a high-quality magnetic resonance image while effectively preserving the high-frequency details of the image, thereby significantly improving the reconstruction efficiency and quality of the multi-channel magnetic resonance image.

[0063] In a preferred embodiment, the trained multi-level wavelet convolution network comprises a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network, and the inputting of the undersampling magnetic resonance image into the trained multi-level wavelet convolution network for reconstruction in step S300 to obtain the reconstructed magnetic resonance image comprises:

[0064] Step S310: performing feature extraction and decomposition on the undersampling magnetic resonance image using the multi-level wavelet decomposition network to obtain a feature map;

[0065] Step S320: reconstructing the feature map using the multi-level wavelet reconstruction network to obtain the reconstructed magnetic resonance image.

[0066] Specifically, the trained multi-level wavelet convolution network in this embodiment comprises a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network. The multi-level wavelet decomposition network uses multi-level wavelet transform and convolution operation to gradually decompose the input undersampling magnetic resonance image into feature maps of different frequencies and resolutions, which contain not only spatial information of the original image but also frequency information of the image at different scales, such as high-frequency components (edges, textures) and low-frequency components (smooth areas), thereby providing rich feature information for the subsequent reconstruction process. The multi-level wavelet reconstruction network gradually recovers the resolution and details of the image through multi-level wavelet inverse transform and convolution operation, so as to reconstruct the feature map extracted by the multi-level wavelet decomposition network into a high-quality magnetic resonance image. In actual application, the parameters and structure of the multi-level wavelet decomposition network and the multi-level wavelet reconstruction network can be designed and adjusted according to specific magnetic resonance imaging tasks and data.

[0067] In a preferred embodiment, the multi-level wavelet decomposition network is composed of a plurality of encoding convolution layers and wavelet decomposition layers stacked alternately, and the feature extraction and decomposition of the undersampling magnetic resonance image using the multi-level wavelet decomposition network in step S310 to obtain a feature map comprises:

[0068] Step S311: performing feature extraction on the undersampling magnetic resonance image using the encoding convolution layer to obtain an initial encoding feature map;

[0069] Step S312: wavelet decomposition transformation is performed on the initial coding feature map using the wavelet decomposition layer to obtain initial coding feature frequency components, and the initial coding feature frequency components are input into the next coding convolution layer and the wavelet decomposition layer for iterative processing in turn to obtain a feature map.

[0070] Specifically, as shown in the figure, Figure 3 The multi-level wavelet decomposition network in this embodiment is stacked by alternating a plurality of coding convolution layers and discrete wavelet transform (DWT) layers. The process of using the multi-level wavelet decomposition network to extract and decompose the undersampled magnetic resonance image includes: first, using the first coding convolution layer to perform preliminary feature extraction on the undersampled magnetic resonance image, specifically, sliding a series of learnable convolution kernels (or filters) on the image, calculating the local dot product of the input image and the convolution kernel, and then obtaining the initial features in the image through normalization and nonlinear activation processing, and representing the initial features in the form of a feature map, which contains low-level features of the image, such as edges, textures and color information. Then, the discrete wavelet transform layer is used to perform discrete wavelet transform on the initial features to obtain different frequency components of the initial feature map. Iterative processing is performed according to the above steps, and the layers are optimized layer by layer until the last coding convolution layer outputs a feature map containing multi-scale features and rich detail information. It should be noted that in actual application, the parameters and structure of the multi-level wavelet decomposition network can be adjusted and optimized according to specific tasks and data to obtain better performance.

[0071] In this embodiment, by combining the advantages of discrete wavelet transform and convolutional neural network, the image features are extracted and decomposed step by step through iterative processing, which not only effectively extracts the multi-scale features of the image, but also better protects the high-frequency details, local textures and microstructure information of the image, and improves the quality and accuracy of image reconstruction.

[0072] In a preferred embodiment, the multi-level wavelet reconstruction network is composed of a plurality of decoding convolution layers and wavelet reconstruction layers stacked alternately, and the reconstruction of the feature map using the multi-level wavelet reconstruction network in step S320 to obtain the reconstructed magnetic resonance image includes:

[0073] Step S321: using the decoding convolution layer to analyze the feature map to obtain a plurality of initial decoding feature frequency components;

[0074] Step S322: using the wavelet reconstruction layer to perform wavelet reconstruction on all the initial decoding feature frequency components to obtain initial decoding features, and inputting the initial decoding features into the next decoding convolution layer and the wavelet reconstruction layer for iterative processing in turn to obtain the reconstructed magnetic resonance image.

[0075] Specifically, as shown in Figure 3 The multi-level wavelet reconstruction network in this embodiment is stacked alternately by decoding convolutional layers and wavelet reconstruction layers, where the wavelet reconstruction layer refers to an inverse wavelet transform (IWT) layer. The process of reconstructing the feature map by using the multi-level wavelet reconstruction network includes: first, the decoding convolutional layer is used to analyze the feature map output by the multi-level wavelet decomposition network to obtain initial feature frequency components. Then, the wavelet reconstruction layer uses the inverse wavelet transform to reconstruct the initial feature frequency components to obtain the initial decoded feature map after reconstruction. The above steps are iteratively processed, and the optimization is performed layer by layer until the last decoding convolutional layer outputs the final reconstructed high-quality magnetic resonance image.

[0076] In this embodiment, the reconstruction capability of the inverse wavelet transform and the image restoration advantage of the convolutional neural network are fully combined, the image features are gradually analyzed and the image details are gradually restored through multiple iterations, and high-precision image reconstruction is realized.

[0077] In a preferred embodiment, the training process of the multi-level wavelet convolutional network includes:

[0078] Step M100: obtaining a plurality of groups of undersampled magnetic resonance image samples and corresponding fully sampled magnetic resonance images based on a fully sampled K-space data set;

[0079] Specifically, based on the multi-channel fully sampled K-space data, undersampled K-space data is generated in a retrospective undersampling manner to obtain a plurality of pairs of undersampled K-space data and corresponding fully sampled K-space data. Then, the inverse Fourier transform is used to generate undersampled magnetic resonance image samples corresponding to the undersampled K-space data, and fully sampled magnetic resonance images corresponding to the corresponding fully sampled K-space data.

[0080] Step M200: constructing a multi-level wavelet convolutional network and initializing network parameters, the multi-level wavelet convolutional network being constructed by a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network;

[0081] Specifically, the network is composed of a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network. The multi-level wavelet decomposition network is stacked alternately by encoding convolutional layers (CNN) and wavelet decomposition layers (DWT). Among them, the encoding convolutional layer contains convolution, normalization, nonlinear activation and other calculation units, which are used to extract image features; the wavelet decomposition layer refers to the discrete wavelet transform, which is used to decompose the image features by wavelet to obtain feature components of different frequencies or scales, which are transmitted to the next convolutional layer. In this application, the type of wavelet can be any, and this embodiment takes Haar wavelet (Haar Wavelet) as an example to illustrate the calculation process of the wavelet decomposition layer. Assuming that the feature output by the convolutional layer is F, where N, C, H and W are the batch size, number of channels, length and width of the feature, respectively. The Haar wavelet contains a low-pass filter f LL and three high-pass filters f LH , f HL and f HH , expressed as follows:

[0082]

[0083] The four filters are used to filter the feature F batch by batch and channel by channel to obtain four frequency components X LL , X LH , X HL and X HH , which are spliced in the channel direction to obtain a new feature After processing by the multi-level wavelet decomposition network, the output feature map will be further input into the multi-level wavelet reconstruction network.

[0084] The multi-level wavelet reconstruction network is stacked alternately by decoding convolutional layers (CNN) and wavelet reconstruction layers (IWT). Among them, the decoding convolutional layer contains convolution, normalization, nonlinear activation and other calculation units, which are used to analyze the features to obtain four different feature frequency components X LL , X LH , X HL and X HH ; the wavelet reconstruction layer refers to the inverse wavelet transform, which is used to perform wavelet reconstruction on the four feature frequency components, and continuously recover and reconstruct high-resolution features X. For the Haar wavelet, the inverse wavelet transform calculation process is as follows:

[0085]

[0086] Step M300: inputting the undersampled magnetic resonance image sample into the multi-level wavelet convolution network to obtain a network reconstructed image;

[0087] Specifically, the undersampled image is input into the multi-level wavelet convolution network to capture the multi-scale features in the image, and the resolution and detail information of the image are gradually recovered through multiple convolutional layers, and finally a reconstructed magnetic resonance image, i.e. a network reconstructed image, is output. The image should have less fold artifacts and noise than the input undersampled image, while retaining the high-frequency details, local textures and microstructure information in the original image, improving the quality and accuracy of image reconstruction.

[0088] Step M400: calculating a reconstruction loss based on the network reconstructed image and the corresponding fully sampled magnetic resonance image;

[0089] Specifically, in the calculation of the reconstruction loss, a preset reconstruction loss function is used to calculate the difference between the network reconstructed image and the corresponding fully sampled magnetic resonance image in terms of pixel value accuracy, structural similarity, texture details and other parameters.

[0090] Step M500: iteratively optimize the parameters of the multi-level wavelet convolution network using a back propagation algorithm until the reconstruction loss reaches a preset loss threshold or the number of iterations reaches a preset iteration number threshold, to obtain a trained multi-level wavelet convolution network.

[0091] In a preferred embodiment, the process of calculating the reconstruction loss in step M400 includes:

[0092] Step M410: based on the network reconstructed image and the corresponding fully sampled magnetic resonance image, the reconstruction loss value of each pixel is calculated;

[0093] Step M420: based on the reconstruction loss value of each pixel, a unit loss drop probability is calculated;

[0094] Step M430: obtain the total number of pixels of the network reconstructed image, and according to the preset loss drop ratio, the total number of pixels and the unit loss drop probability, calculate the loss drop probability of each pixel;

[0095] Step M440: based on the loss weight of each pixel and the reconstruction loss value of each pixel, a total reconstruction loss is calculated.

[0096] Specifically, for the magnetic resonance image reconstruction task, the embodiment proposes a new focal reconstruction loss function (FRL) which can be effectively combined with common pixel-wise type losses such as L1 loss, L2 loss and SSIM loss to improve the reconstruction performance of the model. The principle of FRL is mainly as follows: since there are differences in the reconstruction difficulty of different regions on the image by the neural network, the high-frequency region is theoretically more difficult to reconstruct than the low-frequency region, and the existing reconstruction loss function generally ignores this difference, resulting in the same loss weight for different regions, which leads to the low-frequency region with a larger proportion contributes most of the gradient, which is not conducive to the reconstruction of high-frequency details by the network. To solve this problem, FRL can adaptively adjust the reconstruction loss weight of different regions according to the loss value, so as to make the model focus more attention on the more difficult high-frequency region, and improve the reconstruction performance of the network on high-frequency details. In the embodiment, the weight value is set to only exist in two states of 0 or 1. When the weight is 0, the loss of the pixel is discarded and does not participate in the loss calculation, and when the weight is 1, the loss is calculated normally. Assuming that the network reconstruction result is P, and the corresponding fully sampled magnetic resonance image is t, where M represents the number of image pixels, for any Pixel-wise N type reconstruction loss function (such as L1 loss, L2 loss and SSIM loss, etc.), the calculation process of FRL is as follows:

[0097] Firstly, based on the network reconstructed image and the corresponding full-sampling magnetic resonance image, the loss value g of each pixel is calculated i = Q (P[i], t[i]), and then the following normalization processing is performed:

[0098]

[0099] Where h i is called the unit loss drop probability, which represents the probability that the loss of pixel i is dropped in the case of dropping the loss of a pixel, and β>0 is a normalization correction factor, which is used to adjust the difference degree of the loss drop probability of different reconstruction difficulty pixels in the reconstructed magnetic resonance image. In order to facilitate description, the pixels in the reconstructed magnetic resonance image are divided into easy-to-reconstruct pixels and difficult-to-reconstruct pixels in the embodiment, wherein the difficult-to-reconstruct pixels include high-frequency details, local textures and microstructures and other pixel regions in the reconstructed magnetic resonance image; the easy-to-reconstruct pixels refer to the pixel regions in the reconstructed magnetic resonance image except the difficult-to-reconstruct pixels.

[0100] Suppose the unit loss drop probabilities of the easy-to-reconstruct pixels and the difficult-to-reconstruct pixels are h m and h n , respectively, then the difference degree of the loss drop probabilities of the two is:

[0101]

[0102] In the above formula, g m < g n , therefore the greater β is, the greater the difference degree of the loss drop probabilities of the easy-to-reconstruct pixels and the difficult-to-reconstruct pixels, that is, the easy-to-reconstruct pixels are more likely to be dropped than the difficult-to-reconstruct pixels. Assuming that the total number of pixels in the image is M, and it is hoped that the proportion of pixels dropped in each training is γ, then the probability s i of each pixel being dropped is:

[0103] s i = γMh i (5)

[0104] In the above formula, γ ∈ [0, 1], which is set by human. When γ = 0, it means that all pixels are retained; when γ = 1, it means that all pixels are dropped. Next, the random drop step is executed, if a pixel is dropped, its loss weight is set to 0, so that it does not participate in the loss calculation, otherwise it is set to 1, that is:

[0105]

[0106] where x i is a random number sampled from a standard uniform distribution U(0,1), w i is the loss weight of pixel i, and the final focused reconstruction loss is:

[0107]

[0108] The focused reconstruction loss function FRL constructed in this embodiment can adaptively adjust the reconstruction loss weight of different regions according to the loss value size, prompting the model to focus more attention on the pixel region that is more difficult to reconstruct, and can significantly improve the reconstruction performance of the network on high-frequency details. At the same time, the focused reconstruction loss function FRL can combine various types of existing loss functions to meet different actual calculation precision, calculation amount and other needs, thereby improving the flexibility of the loss value of the reconstructed image.

[0109] As Figure 4 shown, corresponding to the above-mentioned fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network, the embodiment of the application also provides a fast magnetic resonance reconstruction system based on a multi-level wavelet convolution network, and the fast magnetic resonance reconstruction system based on a multi-level wavelet convolution network comprises:

[0110] An information acquisition module 410 is configured to acquire multi-channel undersampled K-space data.

[0111] An undersampled image generation module 420 is configured to perform inverse Fourier transform on the multi-channel undersampled K-space data to obtain an undersampled magnetic resonance image.

[0112] An image reconstruction module 430 is configured to input the undersampled magnetic resonance image into a trained multi-level wavelet convolution network for reconstruction to obtain a reconstructed magnetic resonance image, and the multi-level wavelet convolution network is constructed based on a multi-level wavelet transform and a convolutional neural network.

[0113] Specifically, in the embodiment, the specific functions of the fast magnetic resonance reconstruction system based on a multi-level wavelet convolution network can also be referred to the corresponding description in the fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network, which will not be described here.

[0114] Based on the above-mentioned embodiment, the application further provides a magnetic resonance device, and a principle block diagram thereof can be as Figure 5 shown. The magnetic resonance device can be used to execute the fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network provided by the above-mentioned embodiment, and for the sake of brevity, it will not be described here. The magnetic resonance device comprises a processor, the processor is coupled with a memory, the memory is used to store computer programs or instructions, and the processor is used to execute the computer programs or instructions stored in the memory, so that the method in the above-mentioned method embodiment is executed.

[0115] The application further provides a computer readable storage medium, which stores computer instructions for implementing the method in the method embodiments.

[0116] For example, the computer program is executed by a computer, so that the computer can implement the method in the method embodiments.

[0117] The application embodiments further provide a computer program product containing instructions, which are executed by a computer to make the computer implement the method in the method embodiments.

[0118] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0119] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0120] In several embodiments provided in the application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0121] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, some or all of the units can be selected to achieve the purpose of the embodiment.

[0122] In addition, each functional unit in each embodiment of the application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0123] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various media that can store program codes.

Claims

1. A fast magnetic resonance reconstruction method based on a multi-level wavelet convolution network, characterized in that, The method comprises the following steps: acquiring multi-channel undersampled K-space data; performing inverse Fourier transform on the multi-channel undersampled K-space data to obtain an undersampled magnetic resonance image; inputting the undersampled magnetic resonance image into a trained multi-level wavelet convolution network for reconstruction to obtain a reconstructed magnetic resonance image, wherein the multi-level wavelet convolution network is constructed based on multi-level wavelet transform and a convolutional neural network; the trained multi-level wavelet convolution network comprises a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network, and the inputting of the undersampled magnetic resonance image into the trained multi-level wavelet convolution network for reconstruction to obtain a reconstructed magnetic resonance image comprises: performing feature extraction and decomposition on the undersampled magnetic resonance image by using the multi-level wavelet decomposition network to obtain a feature map; reconstructing the feature map by using the multi-level wavelet reconstruction network to obtain a reconstructed magnetic resonance image; the multi-level wavelet reconstruction network is composed of a plurality of decoding convolution layers and wavelet reconstruction layers which are alternately stacked, and the reconstructing of the feature map by using the multi-level wavelet reconstruction network to obtain a reconstructed magnetic resonance image comprises: analyzing the feature map by using the decoding convolution layer to obtain a plurality of initial decoding feature frequency components; performing wavelet reconstruction on all the initial decoding feature frequency components by using the wavelet reconstruction layer to obtain an initial decoding feature map, and inputting the initial decoding feature map into the next decoding convolution layer and the wavelet reconstruction layer for iterative processing in turn to obtain a reconstructed magnetic resonance image; the training process of the multi-level wavelet convolution network comprises: obtaining a plurality of groups of undersampled magnetic resonance image samples and corresponding fully sampled magnetic resonance images based on a fully sampled K-space data set; constructing a multi-level wavelet convolution network and initializing network parameters, wherein the multi-level wavelet convolution network is constructed by a multi-level wavelet decomposition network and a multi-level wavelet reconstruction network; inputting the undersampled magnetic resonance image samples into the multi-level wavelet convolution network to obtain network reconstruction images; calculating a reconstruction loss based on the network reconstruction images and the corresponding fully sampled magnetic resonance images; iteratively optimizing the parameters of the multi-level wavelet convolution network by using a back propagation algorithm until the reconstruction loss reaches a preset loss threshold or the number of iterations reaches a preset iteration number threshold to obtain a trained multi-level wavelet convolution network; the process of calculating the reconstruction loss comprises: calculating a reconstruction loss value of each pixel based on the network reconstruction images and the corresponding fully sampled magnetic resonance images; calculating a unit loss drop probability based on the reconstruction loss value of each pixel; obtaining the total number of pixels of the network reconstruction image, and calculating a loss drop probability of each pixel according to a preset loss drop ratio, the total number of pixels and the unit loss drop probability; calculating a loss weight of each pixel based on the loss drop probability of each pixel; calculating a total reconstruction loss based on the loss weight of each pixel and the reconstruction loss value of each pixel.

2. The fast magnetic resonance reconstruction method based on multi-level wavelet convolutional network according to claim 1, characterized in that, The multi-level wavelet decomposition network is composed of a plurality of encoding convolutional layers and wavelet decomposition layers stacked alternately, the multi-level wavelet decomposition network is used to extract features and decompose the undersampled magnetic resonance image, and a feature map is obtained, including: The encoding convolutional layer is used to extract features of the undersampled magnetic resonance image, and an initial encoding feature map is obtained. The wavelet decomposition layer is used to decompose the initial encoding feature map, and an initial encoding feature frequency component is obtained, and the initial encoding feature frequency component is input into the next encoding convolutional layer and wavelet decomposition layer for iterative processing in turn, and a feature map is obtained.

3. The multi-level wavelet convolution network based fast magnetic resonance reconstruction method of claim 1, wherein, The unit loss discard probability is calculated based on the reconstruction loss value of each pixel, including: The loss value of each pixel is normalized by using a preset normalization correction factor to obtain a unit loss discard probability, and the normalization correction factor is used to adjust the difference in loss discard probability of different reconstruction difficulty pixels in the network reconstruction image.

4. A fast magnetic resonance reconstruction system based on a multi-level wavelet convolution network, characterized by The system is used to implement the steps of the multi-level wavelet convolution network based fast magnetic resonance reconstruction method of any one of claims 1-3, and the system comprises: An information acquisition module is configured to acquire multi-channel undersampled K-space data; An undersampled image generation module is configured to perform inverse Fourier transform on the multi-channel undersampled K-space data to obtain an undersampled magnetic resonance image; An image reconstruction module is configured to input the undersampled magnetic resonance image into a trained multi-level wavelet convolution network for reconstruction to obtain a reconstructed magnetic resonance image, and the multi-level wavelet convolution network is constructed based on multi-level wavelet transform and convolutional neural network.

5. A magnetic resonance apparatus characterized by comprising: Including: A memory is configured to store executable instructions; A processor is configured to call and run the executable instructions in the memory to perform the steps of the multi-level wavelet convolution network based fast magnetic resonance reconstruction method of any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, when the program instructions are run by the processor, the multi-level wavelet convolution network based fast magnetic resonance reconstruction method of any one of claims 1-3 is implemented.

Images