Introduction of a pre-estimated lightweight susceptibility map reconstruction method, system and electronic device

Abstract

The application discloses a light-weight magnetization rate map reconstruction method, system and electronic device, constructs a three-dimensional multi-scale convolution network of magnetization rate estimation, and comprises an encoding module, a magnetization rate estimation module and a decoding module; a field map is acquired, the field map is input into the encoding module to extract three feature maps with different sizes and channel numbers; a magnetization rate estimation result is acquired, the magnetization rate estimation result is output after a convolution layer and a down-sampling layer to output an estimation feature map; and the feature map output by the encoding module and the estimation feature map output by the magnetization rate estimation module are output after feature extraction fusion and up-sampling to output a magnetization rate image. The light-weight network design can realize intelligent and accurate reconstruction of quantitative magnetization rate maps of tissues and organs, provides reliable quantitative information basis for clinical quantitative evaluation of the magnetization rates of tissues and organs, and has application value in real-time imaging and other scenes.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance image processing technology based on convolutional neural networks, and in particular to a lightweight magnetic susceptibility map reconstruction method, system, and electronic device that incorporates pre-estimation. Background Technology

[0002] Magnetic resonance imaging (MRI) is a technique capable of imaging the internal structure of objects. It offers advantages such as no radiation damage, high resolution, and multi-parameter imaging, making it an indispensable tool for clinical diagnosis. Currently, MRI technology is widely used in scientific research in important fields such as neuroscience, brain science, and mental illness. Under the influence of an external magnetic field, different substances exhibit different magnetization reactions, known as magnetic susceptibility, which can be measured using magnetic susceptibility values. As an emerging MRI technique, quantitative susceptibility mapping (QSM) can measure the magnetic susceptibility characteristics of tissues, thereby enabling precise quantitative analysis of indicators such as iron content, calcification level, and blood oxygen saturation. These applications are of great significance for the research and diagnosis of neurodegenerative diseases, brain injuries, and other organ diseases.

[0003] Most magnetic resonance imaging (MRI) techniques typically extract information from the amplitude map, while the magnetic susceptibility information contained in the phase map is often ignored. In contrast, QSM (Quantitative Susceptibility Mapping) technology can effectively utilize information from the phase map, obtaining a quantitative magnetic susceptibility image after multiple processing steps. QSM technology includes the following key processes: First, there is phase ambiguity, or phase entanglement, between the phase recorded by the imaging system and the true phase; therefore, phase unwinding is required to obtain the original, unentangled phase. Second, because the magnetic susceptibility of air is much greater than that of human tissue, the background magnetic field fluctuates significantly at the tissue-air interface, resulting in lower contrast in local tissue magnetic susceptibility and more pronounced phase entanglement. Therefore, after the phase unwinding step, background field interference needs to be removed to preserve a high-quality local field distribution; this is the background field removal step. Finally, based on a specific reconstruction algorithm, a quantitative magnetic susceptibility image can be obtained from the local field map; this is the dipole inversion step.

[0004] In magnetic susceptibility imaging, dipole inversion is the most crucial step. Due to the presence of singular angles in the dipole nucleus, obtaining the magnetic susceptibility map from the local field map obtained after data preprocessing is an ill-conditioned inverse problem. Although multi-directional data acquisition for magnetic susceptibility reconstruction (COSMOS) is a relatively effective method, it is time-consuming, and multiple data acquisitions may cause discomfort to patients, making it difficult to implement in clinical practice. Therefore, many unidirectional reconstruction methods have been proposed, such as threshold-based methods and regularization methods combined with amplitude map structure priors. However, these methods suffer from significant artifacts, leading to underestimation of magnetic susceptibility, or require manual design of feature extraction rules, specific regularization terms, and parameter adjustments during algorithm implementation, making the process cumbersome.

[0005] In recent years, thanks to advancements in hardware and related theories, especially the emergence of high-performance GPUs and the implementation of effective training methods, deep neural networks have experienced rapid development and application. Convolutional neural networks (CNNs), in particular, have been widely used in image processing and computer vision, achieving excellent results. Deep learning methods directly learn the feature distribution transformation between input and labeled data by training deep neural networks, eliminating the need for manually designed feature extraction rules. Deep learning methods have achieved good results and demonstrated enormous development potential in complex inverse problems such as image denoising, super-resolution, and compressed sensing reconstruction. In the field of QSM (Quantitative Susceptibility Modeling), Jongho Lee et al. first proposed a quantitative magnetic susceptibility reconstruction network based on convolutional neural networks, while Steffen Bollmann et al. successfully reconstructed the magnetic susceptibility image of a real human brain after training on simulated data.

[0006] Deep learning-based magnetic susceptibility reconstruction methods have effectively reduced reconstruction time, but existing methods have not been effectively integrated with traditional methods, limiting the improvement of reconstruction accuracy. Furthermore, existing magnetic susceptibility reconstruction models contain significant parameter redundancy, resulting in a large number of model parameters, complex and inefficient computation, and failing to meet the real-time imaging requirements of practical clinical applications. Summary of the Invention

[0007] The main objective of this invention is to overcome the aforementioned deficiencies in the prior art and propose a lightweight magnetic susceptibility map reconstruction method, system, and electronic device that incorporates pre-estimation, effectively reducing the number of parameters and computational load, and improving the network's operating speed.

[0008] The present invention adopts the following technical solution:

[0009] A lightweight magnetic susceptibility map reconstruction method incorporating pre-estimation is characterized by pre-constructing a three-dimensional multi-scale convolutional network incorporating magnetic susceptibility pre-estimation. The three-dimensional multi-scale convolutional network includes an encoding module, a magnetic susceptibility pre-estimation module, and a decoding module. The specific reconstruction method includes the following steps:

[0010] S1, Obtain the field map, and input the field map into the encoding module to extract feature maps of three different sizes and number of channels;

[0011] S2, obtain the magnetic susceptibility pre-estimation result, and input the magnetic susceptibility pre-estimation result into the magnetic susceptibility pre-estimation module and output the pre-estimation feature map;

[0012] S3, the feature map output by the encoding module and the pre-estimated feature map output by the magnetic susceptibility pre-estimation module are processed by feature extraction, fusion and upsampling to output a magnetic susceptibility image.

[0013] Preferably, the encoding module includes two ResIB blocks and two downsampling layers; each ResIB block includes a first convolutional layer, an Inception module, a second convolutional layer, and a third convolutional layer. The output of the second convolutional layer is input to the third convolutional layer and element-wise added to the output of the third convolutional layer to form a residual structure. The Inception module includes three sizes of convolutional kernels, and the channel ratio of the three sizes of convolutional kernels is 1:1:2.

[0014] Preferably, the first ResIB block takes the field map as input and outputs a first encoded feature map; the second ResIB block takes the downsampled first encoded feature map as input and outputs a second encoded feature map. The downsampling layer connecting the ResIB blocks uses trilinear interpolation, and the size of the output feature map is halved after downsampling, while the number of channels remains unchanged.

[0015] Preferably, the magnetic susceptibility pre-estimation module includes two Conv3 convolutional layers and one downsampling layer. The implementation method of magnetic susceptibility pre-estimation is as follows:

[0016]

[0017] Where T is the set threshold, and B(k), D(k), and χ(k) are the field diagram, dipole nucleus, and magnetic susceptibility diagram in k-space, respectively. k x k x k x χ represents the coordinate values ​​of the three dimensions in k-space; intThe result of (k) is passed through a Conv3 convolutional layer to obtain a first pre-estimated feature map. The first pre-estimated feature map and the downsampled first encoded feature map are concatenated as the input of the second ResIB block. The first pre-estimated feature map is passed through a downsampled layer and a Conv3 convolutional layer to obtain a second pre-estimated feature map. The second pre-estimated feature map and the downsampled second encoded feature map are concatenated as the input of the decoding module.

[0018] Preferably, the decoding module includes three RB modules, two upsampling layers, and one Conv1 layer. Each RB module contains a first decoding convolutional layer and a second decoding convolutional layer. The output of the first decoding convolutional layer is input to the second decoding convolutional layer and element-wise added to the output of the second decoding convolutional layer to form a residual block. The input of the first RB module is the concatenation result of the downsampled second encoded feature map and the second pre-estimated feature map, and the output of the first RB module is the first decoded feature map. The input of the second RB module is the concatenation result of the upsampled first decoded feature map and the second encoded feature map after the Conv3 convolutional layer, and the output of the second RB module is the second decoded feature map. The input of the third RB module is the concatenation result of the upsampled second decoded feature map and the first encoded feature map after the Conv3 convolutional layer, and the output of the third RB module is the third decoded feature map. The upsampling layers connecting the RB modules all use trilinear interpolation. The Conv1 layer after the third RB module is used to extract and fuse features from the third decoded feature map.

[0019] Preferably, the loss function of the three-dimensional multi-scale convolutional network is:

[0020]

[0021] Where χ represents the result of the COSMOS method, which is the label image used for training. The magnetic susceptibility image output by the three-dimensional multi-scale convolutional network is Grad, where Grad represents the edge information extraction using a 3D gradient operator, and k1 and k2 are respectively... and The weight values ​​of these two losses.

[0022] A lightweight magnetic susceptibility map reconstruction system incorporating pre-estimated values, characterized by comprising:

[0023] The network construction module is configured to build a three-dimensional multi-scale convolutional network that incorporates magnetic susceptibility pre-estimation. The three-dimensional multi-scale convolutional network includes an encoding module, a magnetic susceptibility pre-estimation module, and a decoding module.

[0024] The encoding module is configured to acquire a field map and extract feature maps of three different sizes and number of channels from the input field map;

[0025] The magnetic susceptibility pre-estimation module is configured to acquire the magnetic susceptibility pre-estimation result, and output the pre-estimation feature map after passing the magnetic susceptibility pre-estimation result through a convolutional layer and a downsampling layer.

[0026] The decoding module is configured to acquire feature maps and pre-estimated feature maps, perform feature extraction and fusion, upsample, and then output a magnetization image.

[0027] An electronic device, comprising:

[0028] One or more processors;

[0029] Storage device for storing one or more programs.

[0030] When the one or more programs are executed by the one or more processors, the one or more processors implement a lightweight magnetic susceptibility map reconstruction method with a pre-estimated value as described above.

[0031] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0032] This invention addresses the ill-conditioned inverse problem of reconstructing quantitative magnetic susceptibility images from field maps. It employs a deep convolutional neural network, constructs a three-dimensional multi-scale convolutional network based on task characteristics, and designs a specific loss function. While achieving accurate reconstruction of quantitative magnetic susceptibility images, it effectively reduces the number of parameters in the network, improves execution speed, and has greater application value.

[0033] The proposed method incorporates a 3D multi-scale convolutional network that pre-estimates magnetic susceptibility during reconstruction to accelerate network convergence, improve training stability, and reduce computational overhead. The encoding module employs multi-scale convolutional layers to enhance feature extraction capabilities. The entire network uses only two levels of downsampling, with a maximum convolutional kernel size of 5, and a maximum output channel count of only 128. This effectively reduces the number of parameters and computational load, improves network speed, and demonstrates potential for real-time imaging scenarios such as intraoperative navigation. Attached Figure Description

[0034] Figure 1 This is a schematic flowchart of the method of the present invention;

[0035] Figure 2 This is a schematic diagram of the three-dimensional multi-scale convolutional network of the present invention;

[0036] Figure 3To illustrate the impact of replacing the Inception module in ResIB with a regular convolution with the same number of parameters (SP_NoIB), replacing the initial magnetic susceptibility pre-estimation result with a field map (SP_deltaB), and removing the magnetic susceptibility pre-estimation (SP_NoPreEst) on the model using loss-epoch curves, (a) is the loss curve on the training set, and (b) is the loss curve on the validation set.

[0037] Figure 4 The images show the magnetic susceptibility reconstruction results of different methods of the present invention on healthy human brain data. The top two rows are cross-sectional and coronal views of the reconstruction results, respectively, and the bottom two rows are the difference graphs between the top two rows and the gold standard results.

[0038] Figure 5 For the comparison of single-row data of magnetic susceptibility reconstruction results on healthy human brain data using the method of the present invention, the selected single-row data positions are in Figure 4 The white horizontal line in the TKD results indicates this.

[0039] Figure 6 The reconstruction results of different methods of the present invention on calcification data from the 2016 QSM Challenge include two views and corresponding magnified views;

[0040] Figure 7 This is a diagram showing the composition of the system of the present invention.

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0043] See Figure 1 This invention proposes a lightweight magnetic susceptibility map reconstruction method incorporating pre-estimation. A three-dimensional multi-scale convolutional network with pre-estimation of magnetic susceptibility is pre-constructed. This network includes an encoding module, a magnetic susceptibility pre-estimation module, and a decoding module. The encoding module, primarily using ResIB blocks and corresponding downsampling blocks, extracts multiple sets of feature maps from the input field map. The magnetic susceptibility pre-estimation module extracts multiple sets of pre-estimated feature maps from the pre-estimation results. The decoding module concatenates the corresponding output feature maps from the encoding and pre-estimation modules, then passes them through an RB module and performs upsampling to finally reconstruct an accurate magnetic susceptibility image.

[0044] In the embodiments of this application, a three-dimensional convolutional kernel is used. The input of the three-dimensional multi-scale convolutional network is the field map (δB) and the magnetic susceptibility prediction result (χ). int The output is a magnetic susceptibility image (χ). Figure 2 The overall architecture of the network is as follows: the size of the input field map and the output image patch is 64×64×32, the size of the image patch of the input magnetic susceptibility pre-estimation result is 32×32×16, and the numbers {1, 32, 64, 128} represent the number of channels. The number of channels in both the input and output feature maps is 1. The channel number is only labeled in other positions in the figure when the number of channels in the output feature map changes. The upsampling and downsampling methods used by the three-dimensional multi-scale convolutional network are trilinear interpolation, and the activation function used is LeakeyReLU.

[0045] The specific reconstruction method of this invention includes the following steps:

[0046] S1, Obtain the field map, input the field map into the encoding module to extract feature maps of three different sizes and number of channels.

[0047] In a specific embodiment, the encoding module includes two ResIB blocks and two downsampling layers. Each ResIB block includes a first convolutional layer, an Inception module, a second convolutional layer (Conv1), and a third convolutional layer. The output of the second convolutional layer is input to the third convolutional layer and element-wise added to form a residual structure. The Inception module can include convolutional kernels of three sizes: 5×5×3, 3×3×5, and 3×3×3, with a channel ratio of 1:1:2. The first ResIB block takes a field map with 1 channel as input and outputs a first encoded feature map with 32 channels. The second ResIB block takes the downsampled first encoded feature map as input and outputs a second encoded feature map with 64 channels. The second encoded feature map is further halved in size after downsampling, while the number of channels remains unchanged. The downsampling method connecting the ResIB blocks is trilinear interpolation, which halves the size of the output feature map after downsampling while maintaining the same number of channels. The encoding module can then extract the first encoding feature map, the second encoding feature map, and the downsampled second encoding feature map.

[0048] S2, obtain the magnetic susceptibility pre-estimation result, and input the magnetic susceptibility pre-estimation result into the magnetic susceptibility pre-estimation module to output the pre-estimation feature map.

[0049] In this invention, the magnetic susceptibility pre-estimation module includes two Conv3 convolutional layers and one downsampling layer. The implementation method of magnetic susceptibility pre-estimation is as follows:

[0050]

[0051] Where T is the set threshold, and B(k), D(k), and χ(k) are the field diagram, dipole nucleus, and magnetic susceptibility diagram in k-space, respectively. k x k x k x The graphs represent the coordinate values ​​of the three dimensions in k-space, respectively, and the estimated magnetic susceptibility. χ int The result of (k) is passed through a Conv3 convolutional layer to obtain a first pre-estimated feature map. The first pre-estimated feature map and the downsampled first encoded feature map are concatenated as the input to the second ResIB block. The first pre-estimated feature map is then passed through a downsampling layer and a Conv3 convolutional layer to obtain a second pre-estimated feature map. The second pre-estimated feature map and the downsampled second encoded feature map are concatenated as the input to the decoding module. The downsampling method of the magnetic susceptibility pre-estimation module is trilinear interpolation. After downsampling, the output feature map size is halved, while the number of channels remains unchanged.

[0052] S3 outputs a magnetic susceptibility image after feature extraction, fusion, and upsampling of the feature map output by the encoding module and the pre-estimated feature map output by the magnetic susceptibility pre-estimation module.

[0053] Specifically, the decoding module includes three Regression Block (RB) modules, two upsampling layers, and one Conv1 layer. Each RB module has two Conv3 convolutional layers (the first and second decoding convolutional layers, respectively) and a skip connection structure. The output of the first decoding convolutional layer is input to the second decoding convolutional layer and element-wise added to form a residual block. The input of the first RB module is the concatenation result of the downsampled second encoded feature map and the second pre-estimated feature map. The output of the first RB module is the first decoded feature map. The number of input and output channels of the first RB module are 128 and 64, respectively. The input of the second RB module is the concatenation result of the upsampled first decoded feature map and the second encoded feature map after the Conv3 convolutional layer. The output of the second RB module is the second decoded feature map. The number of input and output channels of the second RB module are 128 and 32, respectively. The input of the third RB module is the concatenation result of the upsampled second decoded feature map and the first encoded feature map after the Conv3 convolutional layer. The output of the third RB module is the third decoded feature map. The number of input and output channels of the third RB module are 64 and 32, respectively. The upsampling layer connecting the RB modules uses trilinear interpolation. The Conv1 layer after the third RB module is used to extract and fuse the features of the third decoded feature map with 32 output channels from the third RB module.

[0054] In a specific embodiment, in order to obtain a good magnetic susceptibility reconstruction index while preserving the structural details of the reconstructed result, the loss function of the three-dimensional multi-scale convolutional network of the present invention is as follows:

[0055]

[0056] Where χ represents the result of the COSMOS method, which is the label image used for training. The magnetic susceptibility image output by the three-dimensional multi-scale convolutional network is Grad, where Grad represents the edge information extraction using a 3D gradient operator, and k1 and k2 are respectively... and The weight values ​​of these two losses.

[0057] In a specific embodiment, the network uses only two levels of downsampling and small convolutional kernels with a maximum kernel size of no more than 5, with a maximum number of channels of 128. This effectively reduces the number of network parameters and computational load, achieving a lightweight network. The introduction of the magnetic susceptibility pre-estimation module allows the model to maintain good reconstruction accuracy while being scaled down.

[0058] The training data used in this embodiment came from 9 healthy volunteers. Four volunteers scanned 5 echoes in 5 directions, with TR / TE1 / ΔTE = 28 / 5 / 5 ms, an imaging field of view of 220×220×110, and a matrix size of 224×224×126. The other 5 volunteers scanned 9 echoes in 4 directions, with TR / TE1 / ΔTE = 45 / 2 / 2 ms, an imaging field of view of 220×220×110, and a matrix size of 224×224×110. All data were generated into COSMOS tags for training. Due to GPU memory limitations, the field map and COSMOS were overlapped and cut into 64×64×32 three-dimensional blocks during training, resulting in a total of 6860 training samples.

[0059] The deep learning framework used to build the 3D multi-scale convolutional network with magnetic susceptibility pre-estimation is PyTorch, with a batch size of 8, and k1 and k2 set to 1 and 10 respectively. -3 ADAMW was used as the optimizer, with an initial learning rate set to 10. -4 After iterating through the training data 20 times, the learning rate was adjusted to 10. -5 The proposed network takes approximately 8.2 hours to train on an NVIDIA TITAN X (Pascal) graphics card.

[0060] refer to Figure 3The effects of replacing the Inception module in ResIB with a conventional convolution with the same number of parameters (SP_NoIB), replacing the initial magnetic susceptibility pre-estimation result with a field map (SP_deltaB), and removing the magnetic susceptibility pre-estimation (SP_NoPreEst) on the magnetic susceptibility reconstruction results are shown using Training Loss-Epoch and Validation-Epoch curves. The results show that the networks without the magnetic susceptibility pre-estimation result (SP_deltaB and SP_NoPreEst) have a larger initial loss, demonstrating the effectiveness of the magnetic susceptibility pre-estimation result in accelerating network convergence. However, after replacing the Inception module with a conventional convolution with the same number of parameters, the final convergence loss of SPnet is larger than that of the complete network, indicating the effectiveness of the Inception module design in ResIB. In summary, the design of each module in SPnet can improve reconstruction accuracy to varying degrees.

[0061] This application will also compare the reconstruction results of TKD, SFCR, and QSMnet on healthy human brain data and the 2016 QSM Challenge data. The threshold of TKD is set to 0.2, λ1 and λ2 in SFCR are set to 50 and 1, γ1 and γ2 are set to 2000 and 20, and the QSMnet model code is rewritten according to the original paper. The method of this invention is marked as SPnet.

[0062] Figure 4 The figure shows the magnetization reconstruction results of different methods on healthy human brain data. The top two rows of the figure show the reconstruction results and local magnifications of the cross-sectional and coronal planes; the bottom two rows of the figure show the difference images between the results of different methods and the label (COSMOS); the arrows in the figure indicate the areas where the results of each method differ significantly. As can be seen from the positions indicated by the arrows in the figure, the present invention reconstructs the deep gray matter nucleus region more accurately. From the difference images of the cross-sectional and coronal planes, it can be seen that the results of the present invention are closer to the gold standard COSMOS. In the region around the globus pallidus, the detailed structure of the present invention is more completely preserved. The numbers in the figure are the Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity Index Measure (SSIM) of the layers shown. It can be seen that SPnet has better objective reconstruction metrics. Figure 5 The image shows a comparison of reconstructed data from a single row of healthy human brain data. The selected row is... Figure 4 The white horizontal line in the TKD results indicates that the results of this invention are closest to the results of the COSMOS label.

[0063] Table 1 lists the evaluation metrics for various methods on healthy human brain data, including Root Mean Squared Error (RMSE), High Frequency Error Norm (HFEN), Peak Signal-to-Noise Ratio (PSNR), and structural similarity. Lower RMSE and HFEN values ​​are better, while higher PSNR and structural similarity values ​​are better. Table 1 shows that SPnet achieved the best reconstruction results.

[0064] Table 1 Comparison of objective evaluation indicators for different methods

[0065]

[0066] Table 2 lists the basic performance parameters of each model. The deep learning model is configured with an Intel Xeon E5-2620 CPU, a 12GB NVIDIA Pascal Titan X GPU, and 64GB of RAM, while the traditional method is configured with an AMD Ryzen 7 4800H CPU and 32GB of RAM. Structurally, QSMnet has a more complex hierarchical structure with approximately 164.9M parameters, while SPnet has only 1.05M parameters. However, QSMnet performs worse than SPnet, indicating that the large number of parameters in QSMnet is redundant, and SPnet's structural design is more compact and efficient.

[0067] In terms of execution time, SPnet's training time is approximately 57% of QSMnet's. The test times given in the table represent the time spent by the model processing data, using data of size 192×192×108, obtained after background removal from the original 224×224×110 data. Taking the computational configuration used in this chapter as an example, SPnet takes approximately 1.69 seconds in the data loading-background removal-model data processing process, with data loading taking approximately 1.63 seconds, background removal taking 0.03 seconds, and model data processing taking 0.02 seconds. It is evident that SPnet's model processing time is on the order of magnitude smaller than its data loading time, and therefore will not become a bottleneck for real-time imaging tasks. QSMnet, on the other hand, takes approximately 3.11 seconds from data loading to obtaining the model output, with background removal and data cropping times being the same as SPnet's. QSMnet's model data processing time is approximately 1.40 seconds, on the same order of magnitude as its data loading time. The above comparison of execution time shows that SPnet consumes less computing resources, has higher running efficiency, and is a lightweight model.

[0068] Table 2 Comparison of performance parameters of each model

[0069]

[0070]

[0071] Figure 6 The results of direct testing of different deep learning-based methods on the 2016 QSM Challenge dataset are shown. As can be seen from the difference plot, the proposed method exhibits smaller reconstruction differences compared to other deep learning-based methods. The figures in the figure show the peak signal-to-noise ratio and structural similarity indices of each layer; SPnet demonstrates better objective reconstruction metrics. The reconstruction results in Table 3 also reflect that the proposed method achieves good reconstruction performance while employing a lightweight design.

[0072] Table 3 Comparison of Results Metrics in the 2019 QSM Challenge Data

[0073]

[0074] Further reference Figure 7 As an implementation of the methods shown in the above figures, this application provides an embodiment of a lightweight magnetic susceptibility map reconstruction system that introduces a pre-estimated model. This device embodiment is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0075] This application provides a lightweight magnetic susceptibility map reconstruction system that incorporates a pre-estimated value, comprising:

[0076] Network construction module 1 is configured to construct a three-dimensional multi-scale convolutional network that incorporates magnetic susceptibility pre-estimation. The three-dimensional multi-scale convolutional network includes an encoding module 2, a magnetic susceptibility pre-estimation module 3, and a decoding module 4.

[0077] Encoding module 2 is configured to acquire a field map and extract three feature maps of different sizes and number of channels from the input field map, including a first encoded feature map, a second encoded feature map, and a downsampled second encoded feature map.

[0078] The magnetic susceptibility pre-estimation module 3 is configured to obtain the magnetic susceptibility pre-estimation result, and output the pre-estimation feature map after passing the magnetic susceptibility pre-estimation result through a convolutional layer and a downsampling layer, including a first pre-estimation feature map, a second pre-estimation feature map, etc.

[0079] Decoding module 4 is configured to acquire feature maps and pre-estimated feature maps, perform feature extraction and fusion, and upsampling to output a magnetic susceptibility image. Specifically, it acquires a first encoded feature map, a second encoded feature map, and the concatenated result of the downsampled second encoded feature map and the second pre-estimated feature map, performs feature extraction and fusion, and upsampling to output a magnetic susceptibility image.

[0080] The present invention also proposes an electronic device, comprising:

[0081] One or more processors;

[0082] Storage device for storing one or more programs.

[0083] When one or more programs are executed by one or more processors, the one or more processors implement the above-described lightweight magnetic susceptibility map reconstruction method that introduces a pre-estimated value.

[0084] In another aspect, this application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: construct a three-dimensional multi-scale convolutional network incorporating a magnetic susceptibility pre-estimation, the three-dimensional multi-scale convolutional network including an encoding module, a magnetic susceptibility pre-estimation module, and a decoding module; acquire a field map, input the field map into the encoding module to extract a first encoded feature map and a second encoded feature map; downsample the second encoded feature map and concatenate it with the second pre-estimation feature map, the concatenation result, the second encoded feature map, and the first encoded feature map respectively serving as inputs to the decoding module, and the decoding module finally outputting a magnetic susceptibility image.

[0085] It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium, a computer-readable medium, or any combination thereof. A computer-readable medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, or any combination thereof. More specific examples of a computer-readable medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution device, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than a computer-readable medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution device, apparatus, or apparatus. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0086] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and 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, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0087] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based means to perform the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0088] The units described in the embodiments of this application can be implemented in software or in hardware. The described units can also be located in a processor.

[0089] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A lightweight magnetic susceptibility map reconstruction method incorporating pre-estimated values, characterized in that, A three-dimensional multi-scale convolutional network with pre-estimated magnetic susceptibility is pre-constructed. The three-dimensional multi-scale convolutional network includes an encoding module, a magnetic susceptibility pre-estimation module, and a decoding module. The specific reconstruction method includes the following steps: S1, Obtain the field map, and input the field map into the encoding module to extract feature maps of three different sizes and number of channels; S2, obtain the magnetic susceptibility pre-estimation result, and input the magnetic susceptibility pre-estimation result into the magnetic susceptibility pre-estimation module and output the pre-estimation feature map; S3, the feature map output by the encoding module and the pre-estimated feature map output by the magnetic susceptibility pre-estimation module are processed by feature extraction, fusion and upsampling to output a magnetic susceptibility image; The encoding module includes two ResIB blocks and two downsampling layers; each ResIB block includes a first convolutional layer, an Inception module, a second convolutional layer, and a third convolutional layer. The output of the second convolutional layer is input into the third convolutional layer and element-wise added to the output of the third convolutional layer to form a residual structure. The Inception module includes three sizes of convolutional kernels, and the channel ratio of the three sizes of convolutional kernels is 1:1:

2. The magnetic susceptibility pre-estimation module includes two Conv3 convolutional layers and one downsampling layer. The implementation method of magnetic susceptibility pre-estimation is as follows: ； Where T is the set threshold, , and They are respectively Field diagrams, dipole nuclei, and magnetic susceptibility diagrams in space. , , , , These represent the coordinate values ​​in the three dimensions of the k-space; The result is passed through a Conv3 convolutional layer to obtain a first pre-estimated feature map. The first pre-estimated feature map and the downsampled first encoded feature map are concatenated as the input of the second ResIB block. The first pre-estimated feature map is passed through a downsampled layer and a Conv3 convolutional layer to obtain a second pre-estimated feature map. The second pre-estimated feature map and the downsampled second encoded feature map are concatenated as the input of the decoding module.

2. The lightweight magnetic susceptibility map reconstruction method according to claim 1, characterized in that, The first ResIB block takes the field map as input and outputs a first encoded feature map; the second ResIB block takes the downsampled first encoded feature map as input and outputs a second encoded feature map. The downsampling layer connected to the ResIB block uses trilinear interpolation, and the size of the output feature map is halved after downsampling, while the number of channels remains unchanged.

3. The lightweight magnetic susceptibility map reconstruction method according to claim 1, characterized in that, The decoding module includes three RB modules, two upsampling layers, and one Conv1 layer. Each RB module contains a first decoding convolutional layer and a second decoding convolutional layer. The output of the first decoding convolutional layer is input to the second decoding convolutional layer and element-wise added to the output of the second decoding convolutional layer to form a residual block. The input of the first RB module is the concatenation result of the downsampled second encoded feature map and the second pre-estimated feature map, and the output of the first RB module is the first decoded feature map. The input to the second RB module is the concatenation result of the upsampled first decoded feature map and the second encoded feature map after the Conv3 convolutional layer, and the output of the second RB module is the second decoded feature map; The input to the third RB module is the concatenation result of the upsampled second decoded feature map and the first encoded feature map after the Conv3 convolutional layer. The output of the third RB module is the third decoded feature map. The upsampling layers connected to the RB modules all use trilinear interpolation. The Conv1 layer after the third RB module is used to extract and fuse features from the third decoded feature map.

4. The lightweight magnetic susceptibility map reconstruction method according to claim 1, characterized in that, The loss function of the three-dimensional multi-scale convolutional network is: ； in, The result of the COSMOS method is the label image used for training. The magnetic susceptibility image output by the three-dimensional multi-scale convolutional network. This indicates that a 3D gradient operator is used for edge information extraction. and They are respectively and The weight values ​​of these two losses.

5. A lightweight magnetic susceptibility map reconstruction system incorporating a pre-estimated value, characterized in that, include: The network construction module is configured to build a three-dimensional multi-scale convolutional network that incorporates magnetic susceptibility pre-estimation. The three-dimensional multi-scale convolutional network includes an encoding module, a magnetic susceptibility pre-estimation module, and a decoding module. An encoding module is configured to acquire a field map and extract feature maps of three different sizes and channel numbers from the input field map. The encoding module includes two ResIB blocks and two downsampling layers. Each ResIB block includes a first convolutional layer, an Inception module, a second convolutional layer, and a third convolutional layer. The output of the second convolutional layer is input to the third convolutional layer and element-wise added to the output of the third convolutional layer to form a residual structure. The Inception module includes three convolutional kernels of three sizes, with a channel ratio of 1:1:

2. A magnetic susceptibility pre-estimation module is configured to acquire a magnetic susceptibility pre-estimation result, and output a pre-estimation feature map after passing the magnetic susceptibility pre-estimation result through a convolutional layer and a downsampling layer. The magnetic susceptibility pre-estimation module includes two Conv3 convolutional layers and one downsampling layer. The implementation method of magnetic susceptibility pre-estimation is as follows: ； Where T is the set threshold, , and They are respectively Field diagrams, dipole nuclei, and magnetic susceptibility diagrams in space. , , , , These represent the coordinate values ​​in the three dimensions of the k-space; The result is passed through a Conv3 convolutional layer to obtain a first pre-estimated feature map. The first pre-estimated feature map and the downsampled first encoded feature map are concatenated as the input of the second ResIB block. The first pre-estimated feature map is passed through a downsampled layer and a Conv3 convolutional layer to obtain a second pre-estimated feature map. The second pre-estimated feature map and the downsampled second encoded feature map are concatenated as the input of the decoding module. The decoding module is configured to acquire feature maps and pre-estimated feature maps, perform feature extraction and fusion, upsample, and then output a magnetization image.

6. An electronic device, comprising: One or more processors; 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 a lightweight magnetic susceptibility map reconstruction method that introduces a pre-estimated value as described in any one of claims 1-4.

Images