Multi-Task Phase Preprocessing Method and System Based on Bidirectional Learning 3D Network

Abstract

The present invention discloses a multi-task phase preprocessing method and system based on a bidirectional learning three-dimensional network. This method can perform one-stop processing of integrated phase unwrapping and background field removal on the original phase map. The bidirectional learning three-dimensional network consists of an evolution branch and a degradation branch. The evolution branch uses the local field map as the training label to learn the mapping relationship between the phase map and the field map. The degradation branch uses the wrapped phase map and the background field map as the training labels, aiming to learn the degradation process of the phase map. The wrapped number and the background field estimated by the degradation branch are used to inversely calculate the second local field map from the original phase; it is fused with the first local field map predicted by the evolution branch to achieve phase preprocessing based on bidirectional learning. By introducing amplitude information verification, the ability of the network to suppress artifacts and perturbations in the phase data is improved. The present invention can show more accurate and reliable processing effects in areas with complex wrapped distributions, uneven backgrounds, large signal fluctuations, severe lesions, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic resonance image processing based on convolutional neural networks, and particularly relates to a multi-task phase preprocessing method and system based on a bidirectional learning three-dimensional network. Background Art

[0002] Magnetic resonance imaging is a non-invasive medical imaging technology developed with the development of computer technology, superconducting technology, and electronic circuit technology. Through magnetic resonance technology, clear amplitude images and phase images can be obtained. Based on these images with different contrasts, the tissue structure information can be analyzed in depth, thereby providing effective auxiliary information for various diseases. Quantitative susceptibility mapping is a post-processing technology of MRI for calculating tissue susceptibility values. Susceptibility is an inherent property of a substance used to measure the degree of magnetization of the substance under the action of an external magnetic field. Quantitative susceptibility technology uses the relationship between the dipole field and susceptibility to inversely solve the susceptibility from the phase image and obtain the susceptibility distribution of the tissue, providing auxiliary information for the diagnosis of diseases such as the central nervous system, heart, and abdomen.

[0003] Generally, magnetic resonance imaging technology only uses amplitude information, and the phase image containing susceptibility information is discarded due to the complex processing process. Susceptibility imaging technology can fully exploit the information in the phase image to obtain the susceptibility map of the tissue, thereby overcoming the non-local effect of the dipole field and providing an effective quantitative evaluation tool for various brain diseases. Since the phase information is obtained by using the four-quadrant arctangent function in the phase image, its value range is (-π, π], and as the echo time increases, the actual phase value will exceed this range but be wrapped into this range, that is, phase wrapping occurs. Therefore, a phase unwrapping algorithm is first required to restore the phase wrapping. In addition, the perturbation magnetic field generated by brain tissue within the brain tissue region is defined as the local field, and there are many strong susceptibility sources outside the brain tissue, such as the sinuses and the junction of the skull and brain tissue. These strong susceptibility sources, eddy currents, and the perturbation fields generated by the inhomogeneous main magnetic field are the background fields, and the background fields will interfere with the local field. Therefore, it is necessary to distinguish the background field from the local field to obtain a high-quality local field map generated only by brain tissue. Currently, many algorithms have been proposed for phase preprocessing, including the path method of unwrapping, the Laplace method, the dipole field projection method for removing the background field, and a series of methods based on the Laplace equation, etc. However, these traditional methods need to introduce time and space limitations for unwrapping and need to trade off between time and accuracy; while for the background field, boundary conditions usually need to be assumed, and suitable optimization parameters need to be selected according to different data characteristics, and the processing time is generally relatively long.

[0004] Finally, a susceptibility distribution map is obtained through the susceptibility reconstruction algorithm. Susceptibility reconstruction is a process of solving an ill-posed equation problem. The dipole field is expressed as the product of susceptibility and the dipole kernel in the k-space. However, due to the existence of singular angles, the reconstruction process is an ill-conditioned inverse process. Therefore, it is usually solved by adding a regularization term to transform it into a convex optimization problem.

[0005] Nowadays, with the rapid development of hardware devices and artificial intelligence technologies, deep neural networks have also been rapidly developed and applied. Among them, deep convolutional neural networks have been widely used in the fields of image and computer vision and have demonstrated powerful performance. Convolutional neural networks have received great attention and achieved good results in solving three-dimensional image reconstruction problems, including inverse problems and MRI reconstruction. Deep learning and convolutional neural networks can automatically learn the features required in the image processing process without the need for manual design and selection of features, and have achieved better results than many traditional methods and machine learning methods in many image reconstruction tasks. With sufficient support from training data, a convolutional neural network structure suitable for the task is designed, labels are made, and appropriate parameters are selected to train the network parameters, enabling the network to learn the mapping relationship from the original phase image to the local field map and automatically complete the phase preprocessing of three-dimensional images.

[0006] In recent years, many deep learning algorithms for image reconstruction have been proposed. In particular, the proposed UNet has been widely used in the field of medical image processing. It connects the feature maps between the encoder and the decoder, reuses the features, improves the performance of the network, and enhances the efficiency of network training. In the field of QSM, in 2018, Johnson KM et al. first proposed a phase unwrapping method based on convolutional neural networks. In 2019, Steffen Bollmann et al. applied UNet to the step of removing the background field. However, these networks can only solve one step of phase preprocessing, either phase unwrapping or background field removal. Their structures are basically based on the UNet network and do not fully utilize the advantages of the multi-modal GRE sequence to supplement information.

[0007] Due to the inconsistent sizes of the winding and the background field, it is difficult to select a convolution kernel of an appropriate size. Multi-scale information is important for unwrapping and removing the background field. Visual tasks require multi-scale information to describe objects with a larger receptive field by a feature extractor. The convolutional neural network obtains multi-scale features of different granularities by stacking convolution kernels. The Res2Net network structure generates a combination of multiple receptive fields at a finer granularity by grouping feature channels. After grouping the channels, the feature maps of different groups are each subjected to convolution calculation and then fused to obtain multi-scale information, while increasing the width and depth of the network without introducing too many parameters. In addition, the theory and methods of multi-task learning deepen the connection between phase unwrapping and background field removal, enhancing the network's ability to solve complex problems.

[0008] In summary, the phase preprocessing mainly includes two major steps: phase unwrapping and background field removal. The phase winding situation is variable, the background field interference is complex and is also restricted by the unwrapping result. The strong interference regions such as the air-tissue boundary regions of the sinuses and skulls are difficult to process. At the same time, the traditional algorithms also have disadvantages such as numerous parameters and long processing time. In these aspects, the deep learning method has significant performance advantages and application prospects, but a reasonable convolutional neural network structure needs to be designed to effectively learn the phase winding and background field information simultaneously. Summary of the Invention

[0009] Aiming at the above-mentioned technical problems, the purpose of the embodiments of the present application is to propose a multi-task phase preprocessing method and system based on a bidirectional learning three-dimensional network to solve the technical problems mentioned in the above background art section.

[0010] In a first aspect, the present invention provides a multi-task phase preprocessing method based on a bidirectional learning three-dimensional network, including the following steps:

[0011] S1, construct a bidirectional learning three-dimensional network, the bidirectional learning three-dimensional network includes an evolution branch, a degradation branch and a fusion module. The evolution branch includes a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The degradation branch includes a wrapped phase sub-network and a background field sub-network, and shares the shared sub-network;

[0012] S2, obtain an original phase map. In the evolution branch, input the original phase map into the shared sub-network for feature extraction, and then pass it through the field map reconstruction sub-network, introduce the amplitude map and fuse the wrapped phase distribution feature map in the degradation branch to obtain a first local field map. Input the first local field map into the structural feature sub-network to extract the structural information of the first local field map;

[0013] S3. In the degradation branch, feature extraction is performed through a shared sub-network, and then through a wrapped phase sub-network to obtain a wrapped phase distribution map. The total field map obtained by adding the wrapped phase distribution map and the original phase map is input into the background field sub-network to learn background field information and superimpose the structural information of the first local field map, thereby obtaining a second local field map.

[0014] S4. The first local field map and the second local field map are input into a fusion module for fusion to obtain a final local field map.

[0015] Preferably, the shared sub-network includes a gradient prior module and three grouped multi-scale residual modules connected in sequence, with downsampling performed once between every two grouped multi-scale residual modules; the field map reconstruction sub-network includes two residual modules, an amplitude map fusion module, a set of learnable weights, and two upsamplings; the structural feature sub-network includes a Laplacian operation, grouped multi-scale residual modules, and a residual module; the original phase map is input into the gradient prior module and passes through three grouped multi-scale residual modules in sequence. The output of the second grouped multi-scale residual module in the shared sub-network is cascaded with the output of the third grouped multi-scale residual module and input into the first residual module of the field map reconstruction sub-network. The output of the first residual module is fused with the amplitude map input into the amplitude map fusion module. The fused result is cascaded with the wrapped phase distribution feature map in the degradation branch and passes through a set of learnable weights and the second residual module to obtain the first local field map; the first local field map is input into the structural feature sub-network, first undergoes a Laplacian filtering operation, and then passes through two grouped multi-scale residual modules and a residual module in sequence to obtain the structural information of the first local field map.

[0016] Preferably, in the gradient prior module, after calculating the gradient distribution maps in three directions based on the input original phase map and adding them, the result is cascaded with the input to obtain the output:

[0017] Y = C(G x (X) + G y (X) + G z (X), X);

[0018] where X represents the input, Y represents the output, C represents cascading, and G x , G y , G z represent the gradients in three directions;

[0019] The amplitude map fusion module passes the input, the corresponding amplitude map, and their concatenation through a 3×3×3 convolutional layer. Then, the concatenated feature map and the other two feature maps are concatenated respectively and passed through a 3×3×3 convolutional layer and a Sigmoid function for learning to obtain a weight matrix. Then, the weight matrix is multiplied element-wise with the feature maps of the input and the amplitude map to select features. Finally, the results are concatenated and passed through a 3×3×3 convolutional layer and a LeakyReLU activation function to obtain the output.

[0020] Preferably, the wrapped phase sub-network includes four residual modules, an output convolutional layer, and upsampling and downsampling; the background field sub-network includes two grouped multi-scale residual modules and two residual modules, with downsampling between the two grouped multi-scale residual modules and upsampling between the two residual modules; the output of the third grouped multi-scale residual module of the shared sub-network passes through the first two residual modules of the wrapped phase sub-network, is concatenated with the output of the second grouped multi-scale residual module, then passes through a residual module, is concatenated with the output of the first grouped multi-scale residual module, and then passes through the last residual module and a convolutional layer to obtain the wrapped phase distribution map; the wrapped phase distribution map is added to the original phase map to obtain the total field map. The total field map is input into the background field sub-network and passes through two grouped multi-scale residual modules and two residual modules in sequence to obtain the background field information. The total field map is subtracted from the background field information to obtain the degraded local field map, and the structural information of the first local field map is superimposed to obtain the second local field map.

[0021] Preferably, after the input of the grouped multi-scale residual module passes through a 1×1×1 convolutional layer, all its channels are divided into 4 groups, denoted as x = {x1, x2, x3, x4}. The operations performed on each group after grouping are as follows:

[0022] y1 = x1;

[0023] y2 = f1(x2);

[0024] y3 = f2(f1(x2) + x3);

[0025] y4 = f3(f2(f1(x2) + x3) + x4);

[0026] where f1, f2, and f3 represent three different 3×3×3 convolutional layers. The four groups of outputs are concatenated and denoted as y = {y1, y2, y3, y4}, and then they are processed through a 1×1×1 convolutional layer and the LeakyReLU activation function.

[0027] The residual module includes two 3×3×3 convolutional layers and a LeakyReLU activation function in the middle. The result after passing through two 3×3×3 convolutional layers and a LeakyReLU activation function is connected with the input of the residual module by a residual connection.

[0028] Preferably, the fusion module in step S4 includes:

[0029] Combining the first local field map and the second local field map for joint learning to obtain a weight matrix, screening the field map features of the two branches respectively, cascading and fusing through a residual module to obtain the final local field map:

[0030] M = Sigmoid(f 12 (f 11 (F up )) + f 22 (f 21 (F de ))) ;

[0031] Y = RB(C(f 11 (F up ) × (1 + M), f 21 (F de ))) ;

[0032] Among them, F up , F de respectively represent the first local field map and the second local field map, M represents the weight matrix, Y represents the final local field map, f 11 , f 12 , f 21 , f 22 represent different 3×3×3 convolutional layers, C represents cascading, and RB represents the residual module.

[0033] Preferably, the bidirectional learning three-dimensional network adopts step-by-step training, which is divided into the following three steps:

[0034] Step 1: Train the shared sub-network, the field map reconstruction sub-network of the evolution branch, and the wrapped phase sub-network of the degradation branch. The loss function is defined as the sum of the first local field map loss of the evolution branch and the wrapped phase loss of the degradation branch;

[0035] Step 2: Train the structural feature sub-network of the evolution branch and the background field sub-network of the degradation branch. The loss function is defined as the sum of the background field of the degradation branch and the second local field loss;

[0036] Step 3: Train the fusion module. The loss function is defined as the final local field loss;

[0037] The loss function L(θ) includes mean square error and gradient loss:

[0038]

[0039] Among them, N represents the batch size, χ represents the output and the label respectively, and α represents the hyperparameter used to balance the two losses.

[0040] In a second aspect, the present invention provides a multi-task phase preprocessing system based on a bidirectional learning three-dimensional network, including:

[0041] A network construction unit configured to construct a bidirectional learning three-dimensional network. The bidirectional learning three-dimensional network includes an evolution branch, a degradation branch, and a fusion module. The evolution branch includes a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The degradation branch includes a wrapped phase sub-network and a background field sub-network, and they share the shared sub-network;

[0042] An evolution unit configured to obtain an original phase map. In the evolution branch, the original phase map is input into the shared sub-network for feature extraction, and then through the field map reconstruction sub-network, an amplitude map is introduced and the wrapped phase distribution feature map in the degradation branch is fused to obtain a first local field map. The first local field map is input into the structural feature sub-network to extract the structural information of the first local field map;

[0043] A degradation unit configured to, in the degradation branch, perform feature extraction through the shared sub-network, and then through the wrapped phase sub-network to obtain a wrapped phase distribution map. The total field map result obtained by adding the wrapped phase distribution map and the original phase map is input into the background field sub-network to learn the background field information and superimpose the structural information of the first local field map to obtain a second local field map;

[0044] A final local field map acquisition unit configured to input the first local field map and the second local field map into the fusion module for fusion to obtain a final local field map.

[0045] In a third aspect, the present invention provides an electronic device, including one or more processors; a storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner of the first aspect.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method described in any implementation manner of the first aspect.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention aims at two important phase preprocessing steps in the field of susceptibility reconstruction, namely phase unwrapping and background field removal. For the first time, the two phase preprocessing steps are jointly processed using deep learning methods, reducing the transmission of errors during the processing of different methods and improving the processing efficiency. According to the task characteristics, a bidirectional learning three-dimensional network structure with an evolutionary branch and a degradation branch is constructed, and a specific loss function is designed. By using the bidirectional learning three-dimensional network of evolution and degradation, while the evolutionary branch learns the mapping relationship from the phase image to the local field map, the degradation branch learns the winding information and background field information in the phase image. Since the phase map is composed of the superposition of the background field phase and the local phase of the brain tissue and is caused by the winding phenomenon, the degradation information and the evolutionary information are complementary. Through the bidirectional learning of evolution and degradation and multiple information interactions of evolution and degradation in the network, the reconstruction of the local field map is assisted and improved, strengthening the learning ability in areas with complex winding and severe background fields, and reducing the background field residue in these difficult areas in the finally reconstructed local field map.

[0049] In the method of the present invention, the encoder part in the network all adopts a grouped multi-scale residual module, namely the Res2Net network module, which improves the network's representation ability for features of different scales without increasing the network parameters. At the same time, amplitude map information is introduced into the evolutionary branch to supplement detailed information, making full use of the characteristics of multi-modal magnetic resonance images. Compared with other traditional methods and deep learning methods, a better reconstruction effect is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0051] Figure 1 It is a schematic flowchart of a multi-task phase preprocessing method based on a bidirectional learning three-dimensional network according to an embodiment of the present application;

[0052] Figure 2 It is a schematic structural diagram of a bidirectional learning three-dimensional network of a multi-task phase preprocessing method based on a bidirectional learning three-dimensional network according to an embodiment of the present application;

[0053] Figure 3 It is a schematic structural diagram of different network modules of a multi-task phase preprocessing method based on a bidirectional learning three-dimensional network according to an embodiment of the present application; wherein, (a) is a residual module RB, (b) is a gradient prior module GPB, (c) is a multi-scale residual module R2B, (d) is an amplitude map fusion module MFB, and (e) is a fusion module Fusion;

[0054] Figure 4 In (a), the influence of separately removing the GPB module (PPNet_NoGPB), MFB module (PPNet_NoMFB), and R2B module (PPNet_NoR2B) on the first step is shown through the loss curve; in (b), the influence of separately removing the GPB module, MFB module, and R2B module on the second step is shown through the loss curve; in (c), the influence of separately removing the evolutionary branch (PPNet_NoUpgrade) and degradation branch (PPNet_NoDegrade) on the reconstruction result is shown through the loss curve.

[0055] Figure 5 are the local field map reconstruction results of different methods on healthy human brain simulation data, including the original phase map, amplitude map, local field map reconstruction result, and the corresponding difference map;

[0056] Figure 6 are the local field map reconstruction results of different branches and fusion results of the bidirectional learning three-dimensional network on healthy human brain simulation data, including the original phase map, amplitude map, local field map reconstruction result, and the corresponding difference map;

[0057] Figure 7 are the local field reconstruction results of different methods on patient human brain data, including the original phase map, amplitude map, and local field map reconstruction result;

[0058] Figure 8 are the local field reconstruction results of different methods on the 2019 QSM Challenge data, including the original phase map, amplitude map, local field map reconstruction result, and the corresponding difference map;

[0059] Figure 9 is a schematic diagram of a multi-task phase preprocessing system based on a bidirectional learning three-dimensional network according to an embodiment of the present application. Detailed implementation manners

[0060] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] Figure 1 Illustrates a multi-task phase preprocessing method based on a bidirectional learning three-dimensional network provided by an embodiment of the present application, including the following steps:

[0062] S1. Construct a three-dimensional bidirectional learning network. The three-dimensional bidirectional learning network includes an evolution branch, a degradation branch, and a fusion module. The evolution branch includes a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The degradation branch includes a wrapped phase sub-network and a background field sub-network, and they share the shared sub-network.

[0063] Specifically, the original phase map is obtained by taking the arctangent of the MRI complex signal. The phase information of the signal is obtained using the four-quadrant arctangent function. However, due to the periodicity of the four-quadrant arctangent function, its value range is (-π, π]. In actual imaging, as the echo time increases, there will be phase values exceeding this interval and being wrapped into this range. After phase unwrapping to obtain the total field map, since the total field map is formed by the superposition of the local field and the background field generated by the external interference susceptibility source of the brain, it is necessary to exclude the residual background field in the brain region to obtain the local field map. Therefore, in the learning degradation process, it is necessary to first learn the degradation information of phase wrapping and then learn the degradation information of the background field. Therefore, according to the characteristics of the task, the embodiment of this application designs a three-dimensional bidirectional learning network structure, which can perform one-stop processing of integrated phase unwrapping and background field removal on the original phase map.

[0064] Reference Figure 2 The three-dimensional bidirectional learning network consists of two parts: an evolution branch and a degradation branch. The evolution branch is composed of three parts: a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The evolution branch uses the local field map as the training label to learn the mapping relationship between the phase map and the field map. The degradation branch is composed of a wrapped phase sub-network and a background field sub-network. The degradation branch uses the wrapped phase map and the background field map of the phase as the training labels, aiming to learn the degradation process of the phase map. The degradation branch is composed of a wrapped phase sub-network and a background field sub-network, and shares the shared sub-network of the evolution branch, with a smaller network scale. The wrapped number and background field estimated by the degradation branch can be used to invert the local field from the original phase; fusing it with the field information predicted by the evolution branch realizes phase preprocessing based on bidirectional learning. In addition, by introducing amplitude map information verification, the ability of the network to suppress artifacts and perturbations in phase data is further improved.

[0065] S2. Obtain the original phase map. In the evolution branch, input the original phase map into the shared sub-network for feature extraction, and then pass it through the field map reconstruction sub-network. Introduce the amplitude map and fuse the wrapped phase distribution feature map in the degradation branch to obtain the first local field map. Input the first local field map into the structural feature sub-network to extract the structural information of the first local field map.

[0066] In a specific embodiment, the shared sub-network includes a gradient prior module, a grouped multi-scale residual module, and downsampling; the field map reconstruction sub-network includes a residual module, an amplitude map fusion module, a set of learnable weights, and upsampling; the structural feature sub-network includes a Laplacian operation, a grouped multi-scale residual module, and a residual module. Specifically, the shared sub-network includes a gradient prior module and three grouped multi-scale residual modules connected in sequence. The original phase map is input into the gradient prior module and passes through the three grouped multi-scale residual modules in sequence. The output of the second grouped multi-scale residual module is cascaded with the output of the third grouped multi-scale residual module and input into the first residual module of the field map reconstruction sub-network. The input of the first residual module is fused with the amplitude map input into the amplitude map fusion module. The fused result is cascaded with the wrapped phase distribution map in the degradation branch and passes through a set of learnable weights and the second residual module to obtain the first local field map. The first local field map is input into the structural feature sub-network, first undergoes a Laplacian operation, and then passes through two grouped multi-scale residual modules and a residual module in sequence to obtain the structural information of the first local field map.

[0067] Specifically, in the evolution branch, first, feature extraction is performed through the shared sub-network composed of a gradient prior module, three grouped multi-scale residual blocks, and secondary downsampling. Then, it enters the field map reconstruction sub-network to introduce amplitude map information to supplement the structural information in the field map and fuse the wrapped phase distribution feature map from the degradation branch to help identify the wrapped phase. The first local field map of the evolution branch is obtained through a residual module. In addition, the obtained first local field map will pass through the structural feature sub-network, which includes a Laplacian operator, two grouped multi-scale residual modules, and a residual module. The structural feature sub-network extracts the structural information of the first local field map and inputs it as supplementary information into the degradation branch to enhance the field map reconstruction effect of the degradation branch.

[0068] In a specific embodiment, referring to Figure 3 (a), the residual module includes two layers of 3×3×3 convolutional layers and a LeakyReLU activation function in the middle. The result after passing through the two convolutional layers and the LeakyReLU activation function is connected to the input of the residual module through a residual connection.

[0069] Referring to Figure 3 (b), in the gradient prior module, after calculating the gradient distribution maps in three directions based on the input original phase map and adding them together and cascading with the input, the output is obtained:

[0070] Y = C(G x (X) + G y (X) + G z (X), X);

[0071] where X represents the input, Y represents the output, C represents cascading, and G x,G y ,G z Indicates the gradients in three directions.

[0072] Reference Figure 3 (c), After the input of the grouped multi-scale residual module passes through a 1×1×1 convolutional layer, all its channels are divided into 4 groups, labeled as x = {x1, x2, x3, x4}. The operations performed by each group after grouping are as follows:

[0073] y1 = x1;

[0074] y2 = f1(x2);

[0075] y3 = f2(f1(x2) + x3);

[0076] y4 = f3(f2(f1(x2) + x3) + x4);

[0077] Among them, f1, f2, and f3 represent three different 3×3×3 convolutional layers. After the four groups of outputs are concatenated, they are labeled as y = {y1, y2, y3, y4}, and then they are processed through a 1×1×1 convolutional layer and the activation function LeakyReLU. The grouping operation can achieve multiple receptive fields without introducing too many parameters and obtain multi-scale feature information.

[0078] Reference Figure 3 (d), The amplitude map fusion module passes the input, the corresponding amplitude map, and their concatenation through a 3×3×3 convolutional layer respectively. Then, the concatenated feature map of the two and the other two feature maps are concatenated and passed through a 3×3×3 convolutional layer and the Sigmoid function to learn a weight matrix, and then multiplied point by point with the feature maps of the input and the amplitude map to select features. Finally, they are concatenated and passed through a 3×3×3 convolutional layer and the LeakyReLU activation function to obtain the output. By introducing the amplitude map, the structural information in the local field is enhanced.

[0079] S3. In the degradation branch, feature extraction is performed through a shared sub-network, and then through the wrapped phase sub-network to obtain the wrapped phase distribution map. The total field map result obtained by adding the wrapped phase distribution map and the original phase map is input into the background field sub-network to learn the background field information and superimpose the structural information of the first local field map to obtain the second local field map.

[0080] In a specific embodiment, the winding phase sub-network includes four residual modules and one convolutional layer; the background field sub-network includes two grouped multi-scale residual modules and two residual modules, with downsampling performed between the two grouped multi-scale residual modules and upsampling performed between the two residual modules. Specifically, the output of the third grouped multi-scale residual module of the shared sub-network passes through the first two residual modules of the winding phase sub-network, is concatenated with the output of the second grouped multi-scale residual module, then passes through one residual module, is concatenated with the output of the first grouped multi-scale residual module, and then passes through the last residual module and the convolutional layer to obtain the winding phase distribution map. The winding phase distribution map is added to the original phase map to obtain the total field map, which is input into the background field sub-network and passes through two grouped multi-scale residual modules and two residual modules in sequence to obtain the background field information. The total field map is subtracted from the background field information to obtain the degraded local field map, and the structural information of the first local field map is superimposed to obtain the second local field map. The winding phase distribution feature map output by the last residual module is used to be concatenated with the fusion result output by the amplitude map fusion module in the evolutionary branch.

[0081] S4. Input the first local field map and the second local field map into the fusion module for fusion to obtain the final local field map.

[0082] In a specific embodiment, the fusion module in step S4 specifically includes:

[0083] Combine and learn the first local field map and the second local field map to obtain a weight matrix, screen the field map features of the two branches respectively, concatenate them and fuse them through one residual module to obtain the final local field map:

[0084] M = Sigmoid(f 12 (f 11 (F up )) + f 22 (f 21 (F de ))) ;

[0085] Y = RB(C(f 11 (F up ) × (1 + M), f 21 (F de ) × (1 + M))) ;

[0086] Among them, F up , F de respectively represent the first local field map and the second local field map, M represents the weight matrix, Y represents the final local field map, f 11 , f 12 , f 21 , f 22represents different 3×3×3 convolutional layers, C represents cascade, and RB represents residual module.

[0087] The downsampling in the bidirectional learning three-dimensional network is achieved by maximum pooling, and the upsampling is achieved by trilinear interpolation. The number of upsampling and downsampling in the network is consistent, and the step size is 2, ensuring that the network input and output scales are the same. GPB represents the gradient prior module, R2B represents the grouped multi-scale residual module, MFB represents the amplitude map fusion module, RB represents the residual module, Fusion represents the fusion module, and the numbers {1,32,64,128} represent the number of channels. Other abbreviations are marked in the figure. The input of the bidirectional learning three-dimensional network is the original phase map, the output of the field map reconstruction subnetwork is the first local field map, and the output of the entangled phase subnetwork is the entangled phase distribution map; the input of the background field subnetwork is the unentangled phase, and the output is the local field, which is combined with the structural information of the first local field map output by the structural feature subnetwork of the evolutionary branch to obtain the second local field map; the first local field map and the second local field map are input into the fusion module to obtain the final local field map.

[0088] In a specific embodiment, in order to distinguish the two tasks of phase unwrapping and background field removal, the two-way learning three-dimensional network adopts step-by-step training, which is divided into the following three steps:

[0089] Step 1: train the shared subnetwork, the field graph reconstruction subnetwork of the evolutionary branch, and the winding phase subnetwork of the degenerate branch. The loss function is defined as the sum of the first local field graph loss of the evolutionary branch and the winding phase loss of the degenerate branch.

[0090] Step 2: Train the structural feature sub-network of the evolutionary branch and the background field sub-network of the degenerate branch. The loss function is defined as the sum of the background field of the degenerate branch and the second local field loss.

[0091] Step 3: Train the fusion module, and the loss function is defined as the final local field loss.

[0092] In order to preserve the structural information of the local field as much as possible, the loss function L(θ) includes the mean square error and gradient loss:

[0093]

[0094] Where N is the batch size, χ represents the output and label respectively, and α represents the hyperparameter used to balance the two losses.

[0095] The training data of the embodiments of the present application come from 9 healthy volunteers. Each volunteer scans 9 - 16 echoes in 4 - 5 directions, with TR / TE / ΔTE = 45 / 2 / 2 ms, the imaging field of view being 220×220×110, and the matrix size being 224×224×110. Since there is no completely correct data as labels in the phase pre - processing step, first, the COSMOS susceptibility image of each sample and the dipole kernel are used to calculate the local field map as the label using the forward model. Then, a large susceptibility source is set at the sinus position to simulate the strong interference background field at the sinus position. Additionally, several ellipsoidal susceptibility sources with random sizes and susceptibility values are placed at random positions outside the brain tissue to simulate the background field. Similarly, the background field is obtained using the forward model with the dipole kernel and is superimposed on the local field to synthesize the total field map. Then, referring to the true TE of the reference data, 4 TE values (0.005 s, 0.01 s, 0.015 s, 0.02 s) are set to simulate different echo data. Finally, the total phase is calculated to the range of [-π, π] to simulate the wrapped phase, and then it is used as the input to the network together with the corresponding normalized amplitude. Figure 1 To increase the training sample size, during training, the phase map, amplitude map, and the corresponding local field map are cut into 3D chunks of 64×64×32. A 64×64 image can contain large - scale structures in the original image, and 32 - layer images can also provide rich inter - layer information.

[0096] The hematoma human brain data comes from one volunteer, with a total of 8 echo data. The first echo time is 3.6 ms, the echo interval is 5 ms, the imaging field of view is 256×256×146, and the matrix size is 240×240×73. The amplitude map is also normalized. This set of data is used to test the generalization performance of the network.

[0097] Pytorch is used as the deep - learning framework to implement the three - dimensional network for bidirectional learning of evolution and degradation. The number of channels in the first layer of the network is 32, and the number of channels doubles after each down - sampling. In the loss function, by adjusting the values of w1 and w2, the ratio of the mean - square error to the gradient error is approximately 1:1. Finally, the values of w1 and w2 are set to 1 and 0.5. The training batch size is 8, and ADAM is used as the network optimizer. The initial learning rate for step one and step two is 10 -4 , and it is adjusted to 10 after 20 epochs of training -5 Then, the training data is traversed 10 more times. Since step three is relatively simple, 5×10 -5 is used for training for 5 epochs and then adjusted to 10 -5 for another 5 epochs of training. The computing platform used is the NVIDIA GeForce GTX TiTan X GPU. Using this platform, the time required to train the three - dimensional network for bidirectional learning of evolution and degradation in three steps is approximately 24 hours.

[0098] Figure 4 In (a), the training loss curve in Step 1 shows the effects of removing the gradient prior module, grouped multi-scale residual module, amplitude map fusion module, and the three-level downsampling of the degradation branch in the method of the present invention on the field map reconstruction effect respectively; Figure 4 (b) shows the effects of removing the gradient prior module, grouped multi-scale residual module, amplitude map fusion module, and the three-level downsampling of the degradation branch on the training loss in Step 2. It can be seen from the figure that the use of each main functional module can improve the reconstruction effect to varying degrees. Figure 4 (c) shows the loss function curves of obtaining local field maps by only using the evolution branch and only using the degradation branch. The results show that the dual evolution and degradation branches of the present invention can effectively improve the reconstruction effect of the field map.

[0099] The reconstruction results of the bidirectional learning three-dimensional network proposed in the embodiments of the present application are compared and analyzed with those of Laplacian unwrapping and iRSHARP, Laplacian unwrapping and VSHARP, and UNet methods on healthy human brain simulation data, healthy human brain real data, hematoma patient data, and 2019 QSM challenge calcification data. The truncated singular value decomposition threshold of iRSHARP is 0.1, the Gaussian kernel weight is 0.25, and the convolution kernel radii of iRSHARP and VSHARP are both 6. The method of the present invention is labeled as PPNet.

[0100] Figure 5 are the local field map reconstruction results of different methods on simulated healthy human brain data. The first two rows in the figure show the reconstruction results of the cross-section and coronal plane, and the last two rows are the difference images between the corresponding method results and the labels. The arrows in the figure indicate the areas with obvious differences. The areas pointed by the arrows show that due to the bidirectional learning of evolution and degradation in the present method, it is more accurate in areas with severe wrapping and background field interference. Table 1 below gives the objective evaluation indicators of PPNet and other methods on the tested human brain simulation data, including Root Mean Squared Error (RMSE), High Frequency Error Norm (HFEN), Peak Signal Noise Ratio (PSNR), and Structural Similarity (SSIM). The four indicators of the proposed PPNet method are all superior to other comparison methods. It can be seen from the difference map and the following table that the results of the present invention are closer to the label data.

[0101] Table 1 Comparison of objective evaluation indicators of different methods

[0102]

[0103] Figure 6 The comparison results and difference maps of the first local field map of the evolutionary branch, the second local field map of the degenerative branch, the final local field map, and the label obtained from the two-way learning three-dimensional network PPNet. It can be clearly seen from the difference map that there are obvious differences and a certain complementary relationship between the features learned by the evolutionary branch and the degenerative branch. The final local field map has obvious improvements in the difficult positions of the field map learning of both the evolutionary and degenerative branches, which proves the effectiveness of the network's evolutionary and degenerative branch design.

[0104] Figure 7 The local field map reconstruction results of the proposed PPNet and other comparison methods on the hematoma human brain data. To test the generalization ability of the method, the deep learning-based methods were all trained on healthy human brain data without fine-tuning. It can be observed from the figure that due to the PPNet method's full utilization of the complementarity of positive and negative information and the introduction of the amplitude map as information supplement, there are fewer artifacts and the display of the lesion area is more accurate.

[0105] The results of different methods on the calcification data of the QSM 2019 challenge are as Figure 8 shown. The test data is simulated data containing calcification areas. The deep learning-based methods were first trained on healthy human brain data and then fine-tuned using the simulated data without calcification in QSM 2019. It can be seen from the enlarged area map in the figure that the reconstruction results of the method proposed in the present invention more accurately display the calcification area.

[0106] Further referring to Figure 9 As an implementation of the methods shown in the above figures, the present application provides an embodiment of a multi-task phase preprocessing system based on a two-way learning three-dimensional network. This system embodiment corresponds to the Figure 1 method embodiment shown, and this system can be specifically applied to various electronic devices.

[0107] The embodiment of the present application provides a multi-task phase preprocessing system based on a two-way learning three-dimensional network, including:

[0108] A network construction unit 1 configured to construct a two-way learning three-dimensional network. The two-way learning three-dimensional network includes an evolutionary branch, a degenerative branch, and a fusion module. The evolutionary branch includes a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The degenerative branch includes a wrapped phase sub-network and a background field sub-network, and they share the shared sub-network;

[0109] The evolution unit 2 is configured to obtain the original phase map. In the evolution branch, the original phase map is input into the shared sub-network for feature extraction, and then through the field map reconstruction sub-network, the amplitude map is introduced and the wrapped phase distribution feature map in the degradation branch is fused to obtain the first local field map. The first local field map is input into the structural feature sub-network to extract the structural information of the first local field map;

[0110] The degradation unit 3 is configured to, in the degradation branch, perform feature extraction through the shared sub-network, and then through the wrapped phase sub-network, obtain the wrapped phase distribution feature map. The total field map result obtained by adding the wrapped phase distribution feature map and the original phase map is input into the background field sub-network to learn the background field information and superimpose the structural information of the first local field map to obtain the second local field map;

[0111] The final local field map acquisition unit 4 is configured to input the first local field map and the second local field map into the fusion module for fusion to obtain the final local field map.

[0112] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium or a computer-readable medium or any combination of the above two. The computer-readable medium can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, devices, or components, or any combination of the above. More specific examples of the computer-readable medium can include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this application, the computer-readable medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution device, device, or component. And in this application, the computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries the computer-readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable medium, and this computer-readable medium can send, propagate, or transmit a program for use by or in combination with an instruction execution device, device, or component. The program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0113] Computer program code for performing the operations of this application can be written in one or more programming languages or combinations thereof. The programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or it can also be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0114] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram can represent a module, a segment of a program, or a part of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they can sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based device that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0115] The units described in the embodiments of this application can be implemented in software or in hardware. The described units can also be provided in a processor.

[0116] As another aspect, the present application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or may exist separately without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by the electronic device, the electronic device is caused to: construct a bidirectional learning three-dimensional network, the bidirectional learning three-dimensional network includes an evolution branch, a degradation branch and a fusion module, the evolution branch includes a shared sub-network, a field map reconstruction sub-network, a structural feature sub-network, the degradation branch includes a wrapped phase sub-network and a background field sub-network, and shares the shared sub-network; obtain an original phase map, in the evolution branch, input the original phase map into the shared sub-network for feature extraction, then pass through the field map reconstruction sub-network, introduce an amplitude map and fuse the wrapped phase distribution feature map in the degradation branch to obtain a first local field map, input the first local field map into the structural feature sub-network to extract the structural information of the first local field map; in the degradation branch, perform feature extraction through the shared sub-network, then pass through the wrapped phase sub-network to obtain a wrapped phase distribution map, input the total field map result obtained by adding the wrapped phase distribution map and the original phase map into the background field sub-network, learn the background field information and superimpose the structural information of the first local field map to obtain a second local field map; input the first local field map and the second local field map into the fusion module for fusion to obtain a final local field map.

[0117] The above description is only a preferred embodiment of the present application and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solution formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the present application.

Claims

1. A multi-task phase preprocessing method based on a bidirectional learning three-dimensional network, characterized in that It includes the following steps: S1. Construct a bidirectional learning three-dimensional network. The bidirectional learning three-dimensional network includes an evolution branch, a degradation branch, and a fusion module. The evolution branch includes a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The degradation branch includes a wrapped phase sub-network and a background field sub-network, and they share the shared sub-network. The bidirectional learning three-dimensional network adopts step-by-step training, which is divided into the following three steps: Step 1. Train the shared sub-network, the field map reconstruction sub-network of the evolution branch, and the wrapped phase sub-network of the degradation branch. The loss function is defined as the sum of the first local field map loss of the evolution branch and the wrapped phase loss of the degradation branch. Step 2. Train the structural feature sub-network of the evolution branch and the background field sub-network of the degradation branch. The loss function is defined as the sum of the background field of the degradation branch and the second local field loss. Step 3. Train the fusion module. The loss function is defined as the final local field loss. The loss function L(θ) includes mean square error and gradient loss: where N represents the batch size, χ represents the output and the label respectively, and α represents the hyperparameter used to balance the two losses; S2. Obtain the original phase map. In the evolution branch, input the original phase map into the shared sub-network for feature extraction, and then pass it through the field map reconstruction sub-network. Introduce the amplitude map and fuse the wrapped phase distribution feature map in the degradation branch to obtain the first local field map. Input the first local field map into the structural feature sub-network to extract the structural information of the first local field map. S3. In the degradation branch, perform feature extraction through the shared sub-network, and then pass it through the wrapped phase sub-network to obtain the wrapped phase distribution map. Input the total field map obtained by adding the wrapped phase distribution map and the original phase map into the background field sub-network to learn the background field information and superimpose the structural information of the first local field map to obtain the second local field map. S4. Input the first local field map and the second local field map into the fusion module for fusion to obtain the final local field map. The fusion module in step S4 includes: Combine and learn the first local field map and the second local field map to obtain a weight matrix, screen the field map features of the two branches respectively, cascade and fuse them through a residual module to obtain the final local field map: M = Sigmoid(f 12 (f 11 (F up )) + f 22 (f 21 (F de ))); Y = RB(C(f 11 (F up ) × (1 + M), f 21 (F de ) × (1 + M))); Among them, F up , F de respectively represent the first local field map and the second local field map, M represents the weight matrix, Y represents the final local field map, f 11 , f 12 , f 21 , f 22 represent different 3×3×3 convolutional layers, C represents concatenation, and RB represents the residual module.

2. The multi-task phase preprocessing method based on a two-way learning three-dimensional network according to claim 1, wherein The shared sub-network includes a gradient prior module and three grouped multi-scale residual modules connected in sequence. Downsampling is performed once between every two grouped multi-scale residual modules. The field map reconstruction sub-network includes two residual modules, an amplitude map fusion module, a group of learnable weights, and two upsamplings. The structural feature sub-network includes a Laplacian operation, a grouped multi-scale residual module, and a residual module; the original phase map is input to the gradient prior module and passes through three grouped multi-scale residual modules in sequence. The output of the second grouped multi-scale residual module in the shared sub-network is cascaded with the output of the third grouped multi-scale residual module and input to the first residual module of the field map reconstruction sub-network. The output of the first residual module and the amplitude map are input to the amplitude map fusion module for fusion. The fused result is cascaded with the wrapped phase distribution feature map in the degradation branch and passes through a set of learnable weights and the second residual module to obtain the first local field map; the first local field map is input to the structural feature sub-network, first undergoes a Laplacian filtering operation, and then passes through two grouped multi-scale residual modules and a residual module in sequence to obtain the structural information of the first local field map.

3. The multi-task phase preprocessing method based on a two-way learning three-dimensional network according to claim 2, wherein In the gradient prior module, after calculating the gradient distribution maps in three directions based on the input original phase map and adding them up, the result is cascaded with the input to obtain the output: Y = C(G x (X) + G y (X) + G z (X), X); Among them, X represents the input, Y represents the output, C represents the cascade, and G x , G y , G z represent the gradients in three directions; In the amplitude map fusion module, the input, the corresponding amplitude map, and their cascade are all passed through a 3×3×3 convolutional layer. Then, the feature map of their cascade and the other two feature maps are respectively cascaded and passed through a 3×3×3 convolutional layer and a Sigmoid function for learning to obtain a weight matrix. Then, the weight matrix is multiplied element-wise with the feature maps of the input and the amplitude map to select features. Finally, they are cascaded and passed through a 3×3×3 convolutional layer and a LeakyReLU activation function to obtain the output.

4. The multi-task phase preprocessing method based on a two-way learning three-dimensional network according to claim 2, characterized in that The wrapped phase sub-network includes four residual modules, an output convolutional layer, and upsampling and downsampling; the background field sub-network includes two grouped multi-scale residual modules and two residual modules. Downsampling is performed between the two grouped multi-scale residual modules, and upsampling is performed between the two residual modules; the output of the third grouped multi-scale residual module of the shared sub-network passes through the first two residual modules of the wrapped phase sub-network, is cascaded with the output of the second grouped multi-scale residual module, then passes through a residual module, is cascaded with the output of the first grouped multi-scale residual module, and then passes through the last residual module and a convolutional layer to obtain the wrapped phase distribution map; the wrapped phase distribution map is added to the original phase map to obtain the total field map. The total field map is input to the background field sub-network and passes through two grouped multi-scale residual modules and two residual modules in sequence to obtain the background field information. The total field map is subtracted from the background field information to obtain the degraded local field map, and the structural information of the first local field map is superimposed to obtain the second local field map.

5. The multi-task phase preprocessing method based on a two-way learning three-dimensional network according to claim 4, wherein The input of the grouped multi-scale residual module passes through a 1×1×1 convolutional layer and then all its channels are divided into 4 groups, marked as x = {x1, x2, x3, x4}. The operations performed on each group after grouping are as follows: y1 = x1; y2 = f1(x2); y3 = f2(f1(x2) + x3); y4 = f3(f2(f1(x2) + x3) + x4); Among them, f1, f2, and f3 represent three different 3×3×3 convolutional layers. After the four groups of outputs are concatenated, they are marked as y = {y1, y2, y3, y4}, and then processed through a 1×1×1 convolutional layer and the activation function LeakyReLU. The residual module includes two 3×3×3 convolutional layers and a LeakyReLU activation function in the middle. The result after passing through the two 3×3×3 convolutional layers and a LeakyReLU activation function is subjected to a residual connection with the input of the residual module.

6. A multi-task phase preprocessing system based on a bidirectional learning three-dimensional network, characterized in that, Including: A network construction unit configured to construct a bidirectional learning three-dimensional network. The bidirectional learning three-dimensional network includes an evolution branch, a degradation branch, and a fusion module. The evolution branch includes a shared sub-network, a field map reconstruction sub-network, and a structural feature sub-network. The degradation branch includes a wrapped phase sub-network and a background field sub-network, and they share the shared sub-network. The bidirectional learning three-dimensional network is trained in a step-by-step manner, which is divided into the following three steps: Step 1: Train the shared sub-network, the field map reconstruction sub-network of the evolution branch, and the wrapped phase sub-network of the degradation branch. The loss function is defined as the sum of the first local field map loss of the evolution branch and the wrapped phase loss of the degradation branch. Step 2: Train the structural feature sub-network of the evolution branch and the background field sub-network of the degradation branch. The loss function is defined as the sum of the background field of the degradation branch and the second local field loss. Step 3: Train the fusion module. The loss function is defined as the final local field loss. The loss function L(θ) includes mean square error and gradient loss: Among them, N represents the batch size, χ represents the output and the label respectively, and α represents the hyperparameter used to balance the two losses; An evolution unit configured to obtain an original phase map. In the evolution branch, the original phase map is input into the shared sub-network for feature extraction, and then passed through the field map reconstruction sub-network. An amplitude map is introduced and the wrapped phase distribution feature map in the degradation branch is fused to obtain a first local field map. The first local field map is input into the structural feature sub-network to extract the structural information of the first local field map. A degradation unit configured to, in the degradation branch, perform feature extraction through the shared sub-network, and then pass through the wrapped phase sub-network to obtain a wrapped phase distribution map. The total field map result obtained by adding the wrapped phase distribution map and the original phase map is input into the background field sub-network to learn background field information and superimpose the structural information of the first local field map to obtain a second local field map. A final local field map acquisition unit configured to input the first local field map and the second local field map into the fusion module for fusion to obtain a final local field map. The fusion module of the final local field map acquisition unit includes: Combining the first local field map and the second local field map to learn a weight matrix, screening the field map features of the two branches respectively, concatenating and fusing them through a residual module to obtain the final local field map: M = Sigmoid(f 12 (f 11 (f up )) + f 22 (f 21 (F de ))); Y = RB(C(f 11 (F up ) × (1 + M), f 21 (F de ) × (1 + M))); Among them, F up , F de respectively represent the first local field map and the second local field map, M represents the weight matrix, Y represents the final local field map, f 11 , f 12 , f 21 , f 23 represent different 3×3×3 convolutional layers, C represents concatenation, and RB represents the residual module.

7. An electronic device, including: One or more processors; A storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the method described in any one of claims 1-5.

Images