Image reconstruction method and device, computer equipment, storage medium and program product
By using the network parameter sharing method of multiple reconstruction processes in magnetic resonance imaging image reconstruction, the problem of poor convergence effect of neural network models is solved, and the image reconstruction quality and model training efficiency are improved.
Patent Information
- Application Number
- CN202311800252.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-25
- Publication Date
- 2025-06-27
AI Technical Summary
In the prior art, neural network models have poor convergence effects during the reconstruction of magnetic resonance imaging images, resulting in low quality of reconstruction images.
An image reconstruction method is adopted to reconstruct the multi-phase initial reconstruction image through the first data fidelity layer and the three-dimensional neural network layer, determine the intermediate reconstruction image, and use it as the initial image of the next reconstruction process, realizing network parameters sharing of the multiple reconstruction process.
The convergence speed and effect of the image reconstruction model are improved, the quality of the reconstruction image is improved, and the data generation and memory usage are reduced during the training process.
Smart Images
Figure CN120219600A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical imaging technology, and particularly to an image reconstruction method, apparatus, computer device, storage medium, and program product. Background Art
[0002] Magnetic Resonance Imaging (MRI) is a widely used medical imaging technology. Usually, a patient is first scanned by magnetic resonance to obtain scan data (such as K-space data), and then a magnetic resonance image is reconstructed based on the scan data.
[0003] To improve the imaging speed and quality, in related technologies, undersampled scan data is usually combined with the prior constraint of a neural network model for image reconstruction.
[0004] However, limited by the number of iterations in the training process of the neural network model, the model convergence effect is poor, thus reducing the quality of the reconstructed image obtained based on this neural network model. Summary of the Invention
[0005] Based on this, it is necessary to provide an image reconstruction method, apparatus, computer device, storage medium, and program product for the above technical problems.
[0006] In a first aspect, this application provides an image reconstruction method, including:
[0007] Reconstructing multi-phase initial reconstructed images based on multi-phase K-space data;
[0008] Performing at least one reconstruction process on the multi-phase initial reconstructed images according to an image reconstruction model to obtain dynamic reconstructed images; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer;
[0009] The at least one reconstruction process includes:
[0010] Reconstructing the multi-phase initial reconstructed images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstructed images, and using the multi-phase intermediate reconstructed images of the current reconstruction as the multi-phase initial reconstructed images of the next reconstruction process;
[0011] Among them, the first data fidelity layer is used to perform K-space data fidelity processing on the input images; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process.
[0012] In one embodiment, at least two adjacent-phase K-space data in the multi-phase K-space data are obtained by interleaved acquisition.
[0013] In one embodiment, at least one reconstruction process is performed on the multi-phase initial reconstruction images according to the image reconstruction model to obtain dynamic reconstruction images, including:
[0014] Performing a first preset number of reconstruction processes on the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to obtain multi-phase target intermediate reconstruction images;
[0015] Determining the dynamic reconstruction images according to the multi-phase target intermediate reconstruction images.
[0016] In one embodiment, the image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer; determining the dynamic reconstruction images according to the multi-phase target intermediate reconstruction images includes:
[0017] Inputting the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images into the second data fidelity layer to perform fidelity processing on the K-space data;
[0018] Inputting the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images after the fidelity processing into the two-dimensional neural network layer, and correcting the corresponding-phase target intermediate reconstruction images according to the initial reconstruction images of each phase through the two-dimensional neural network layer to obtain the corrected multi-phase target intermediate reconstruction images;
[0019] Obtaining the dynamic reconstruction images according to the corrected multi-phase target intermediate reconstruction images.
[0020] In one embodiment, reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine the multi-phase intermediate reconstruction images, including:
[0021] Inputting the multi-phase initial reconstruction images into the first data fidelity layer to obtain multi-phase reference images;
[0022] Inputting the multi-phase reference images into the three-dimensional neural network layer to obtain multi-phase intermediate reconstruction images.
[0023] In one embodiment, the training process of the image reconstruction model includes:
[0024] Training the initial image reconstruction model with multiple groups of multi-phase sample images to obtain the image reconstruction model; wherein, the training process of each group of multi-phase sample images includes:
[0025] Inputting the multi-phase sample images into the initial image reconstruction model to perform the reconstruction process, and updating the network parameters of the three-dimensional neural network layer every second preset number of times until the images output by two adjacent reconstruction processes meet the preset difference condition.
[0026] In a second aspect, the present application further provides an image reconstruction apparatus, including:
[0027] An initial reconstruction module, configured to reconstruct multi-phase initial reconstruction images based on multi-phase K-space data;
[0028] An iterative reconstruction module, configured to perform at least one reconstruction process on the multi-phase initial reconstruction images according to an image reconstruction model to obtain dynamic reconstruction images; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer; the at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine intermediate reconstruction images, and using the intermediate reconstruction images of the current reconstruction as the multi-phase initial reconstruction images of the next reconstruction process; wherein, the first data fidelity layer is configured to perform fidelity processing on the input images for K-space data; the three-dimensional neural network layer is configured to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process.
[0029] In a third aspect, the present application further provides a computer device, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of any one of the above methods are implemented.
[0030] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the above methods are implemented.
[0031] In a fifth aspect, the present application further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of any one of the above methods are implemented.
[0032] The above-mentioned image reconstruction method, device, computer equipment, storage medium and program product reconstruct multi-phase initial reconstruction images based on K-space data of multiple phases, and perform at least one reconstruction process on the multi-phase initial reconstruction images according to an image reconstruction model to obtain dynamic reconstruction images. Among them, the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer. The at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstruction images, and using the multi-phase intermediate reconstruction images reconstructed in the current time as the multi-phase initial reconstruction images for the next reconstruction process. Among them, the first data fidelity layer is used to perform fidelity processing on the input images for K-space data; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process. In the above method, the image reconstruction model is trained based on sharing the network parameters of the three-dimensional neural network layer, which reduces the data generated during the training process, reduces the video memory occupancy of the device carrying the model, realizes the full training of the model under limited video memory conditions, speeds up the convergence speed of the entire image reconstruction model, improves the convergence effect, and correspondingly improves the image quality of the images reconstructed based on the entire image reconstruction model. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is the internal structure diagram of a computer device in one embodiment;
[0034] Figure 2 is the flowchart of the image reconstruction method in one embodiment;
[0035] Figure 3 is the schematic diagram of the K-space data mask of multiple phases in one embodiment;
[0036] Figure 4 is the flowchart of the image reconstruction method in another embodiment;
[0037] Figure 5 is the flowchart of the image reconstruction method in another embodiment;
[0038] Figure 6 is the flowchart of the image reconstruction operation in one embodiment;
[0039] Figure 7 is the flowchart of the image reconstruction method in another embodiment;
[0040] Figure 8 is the schematic diagram of the graphical process of the image reconstruction method in one embodiment;
[0041] Figure 9 is the structural block diagram of the image reconstruction device in one embodiment. Detailed implementation manners
[0042] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0043] The image reconstruction method provided by the embodiments of the present application can be applied to a computer device as shown in Figure 1 The computer device may be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (Near Field Communication) or other technologies. When the computer program is executed by the processor, an image reconstruction method is implemented. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a button, a trackball or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.
[0044] Those skilled in the art can understand that Figure 1 the structure shown in
[0045] is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements. Figure 2 In one embodiment, as shown in Figure 1 taking the method applied to the computer device in
[0046] S210. Reconstruct multi-phase initial reconstruction images according to multi-phase K-space data.
[0047] Among them, K-space refers to the Fourier space of the rectangular coordinate space, that is, the frequency space of the Fourier transform, also known as the Fourier space. K-space data is used to represent the data converted from magnetic resonance signals.
[0048] Optionally, the computer device may communicate with a magnetic resonance device to obtain multi-phase K-space data during the magnetic resonance scanning of a scanned object by the magnetic resonance device, and convert the obtained K-space data into the image domain for image reconstruction to obtain multi-phase initial reconstructed images. Among them, the computer device may use an image reconstruction algorithm to reconstruct the multi-phase K-space data to obtain multi-phase initial reconstructed images. Exemplarily, the image reconstruction algorithm may be an Iterative Self-consistent Parallel Imaging Reconstruction (SPIRiT) algorithm.
[0049] It should be noted that the computer device may also be a magnetic resonance device to directly obtain multi-phase K-space data. Exemplarily, the computer device may convert the multi-phase K-space data obtained from a whole-body scan of the scanned object to obtain multi-phase initial reconstructed images corresponding to the whole body of the scanned object, and then use an image reconstruction model to reconstruct the whole-body dynamic image of the scanned object.
[0050] In practical applications, the obtained multi-phase K-space data is usually undersampled K-space data. To improve the quality of the multi-phase initial reconstructed images obtained by reconstruction, the computer device may determine combined K-space data based on the multi-phase undersampled K-space data, and reconstruct the multi-phase initial reconstructed images according to the combined K-space data. Among them, the combined K-space data is more complete and comprehensive than the undersampled K-space data.
[0051] Optionally, the computer device may merge the multi-phase undersampled K-space data and then take the average value to obtain combined K-space data, and reconstruct the multi-phase initial reconstructed images based on the combined K-space data. For example, a coil sensitivity map (CSM) is determined based on the combined K-space data, and the multi-phase initial reconstructed images are reconstructed by combining the coil sensitivity map and the multi-phase undersampled K-space data.
[0052] S220. Perform at least one reconstruction process on the multi-phase initial reconstructed images according to the image reconstruction model to obtain dynamic reconstructed images.
[0053] The image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer; the above at least one reconstruction process includes:
[0054] Reconstruct the multi-phase initial reconstructed images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstructed images, and use the multi-phase intermediate reconstructed images of the current reconstruction as the multi-phase initial reconstructed images of the next reconstruction process.
[0055] Among them, the first data fidelity layer is used to perform fidelity processing on the input image in the K-space data; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase image according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared during at least one reconstruction process.
[0056] Continuous multi-phase images can form a dynamic image. The reconstruction process is a process of denoising and artifact removal for the input image of the image reconstruction model. When the input image is a multi-phase image, the obtained reconstructed image is correspondingly a dynamic reconstructed image.
[0057] The data fidelity layer can also be called the Data Consistency (DC) layer. The three-dimensional neural network (3D Convolutional Neural Network, 3D CNN) layer represents a prior constraint and is used to perform Artificial Intelligence (AI) processing on the input image to constrain the reconstruction process to obtain a reconstructed image that meets the requirements.
[0058] It should be noted that network parameter sharing means that during the training process of the image reconstruction model, it is not necessary to update the network parameters of the three-dimensional neural network every time the reconstruction is trained. Instead, the training is carried out by sharing the same set of network parameters for multiple training reconstructions. Among them, network parameter sharing can be achieved based on the Deep Equilibrium (DEQ) model.
[0059] The reconstruction process implemented by the image reconstruction model satisfies the following formula:
[0060]
[0061] represents the magnetic resonance image updated by each reconstruction, represents the finally sought magnetic resonance image, represents the fidelity term coefficient, A represents the forward process of magnetic resonance undersampling, b represents the undersampled K-space data, is the prior constraint term or regularization term for magnetic resonance rapid imaging. The fidelity coefficient can be set by human experience or obtained by convolutional learning.
[0062] According to the Proximal Gradient Descent (PGD) algorithm, the above iterative formula can be expanded as follows:
[0063]
[0064] The DC layer in the image reconstruction model is used to implement the part in the above expansion. The 3D CNN layer is used to implement the part in the above expansion. The above expansion represents that the processing of the DC layer is performed first, followed by the processing of the 3D CNN layer. In practical applications, the processing of the 3D CNN layer can also be performed first, followed by the processing of the DC layer.
[0065] Among them, k represents the number of reconstruction times, represents the image input to the DC layer, represents the image output by the DC layer and is also the image input to the 3D CNN layer, represents the image output by the 3D CNN layer.
[0066] The DC layer strengthens the data consistency between the input image and the output image through methods such as the proximity operator or data backfill to achieve the fidelity processing of the K-space data, so that the input image and the output image of the DC layer retain the same K-space data to the greatest extent, ensuring the consistency of the K-space data during the reconstruction process and enhancing the stability and robustness of the entire image reconstruction model.
[0067] The 3D CNN layer extracts the features of the input multi-phase images, and outputs the processed multi-phase images according to the features of other phase images combined with the features of its own phase image, correspondingly forming multi-phase images to achieve the comprehensive determination of the output multi-phase images. The comprehensive determination of the output using multi-phase images speeds up the convergence of the image reconstruction model, reduces reconstruction artifacts at the same time, and improves the quality of the reconstructed images. For example, the multi-phase images input to the 3D CNN layer include images 1 to 4. The three-dimensional network model can determine the output image 1' corresponding to image 1 based on the features of image 2 combined with the features of image 1, determine the output image 2' corresponding to image 2 based on the features of images 1 and 3 combined with the features of image 2, determine the output image 3' corresponding to image 3 based on the features of images 2 and 4 combined with the features of image 3, and determine the output image 4' corresponding to image 4 based on the features of image 3 combined with the features of image 4, that is, output multi-phase images 1' to 4'.
[0068] Optionally, the computer device may first input the multi-phase initial reconstruction images into the first data fidelity layer for K-space data fidelity processing, and then the three-dimensional neural network layer comprehensively determines the multi-phase intermediate reconstruction images based on the multi-phase initial reconstruction images after the fidelity processing, and uses the multi-phase intermediate reconstruction images of the current reconstruction as the multi-phase initial reconstruction images for the next reconstruction process, and performs multiple reconstruction processes in this way. The dynamic reconstruction images are determined based on the multi-phase intermediate reconstruction images (i.e., multi-phase target intermediate reconstruction images) finally output by the three-dimensional neural network layer. The computer device may also first input the multi-phase initial reconstruction images into the three-dimensional neural network layer, comprehensively determine the multi-phase intermediate reconstruction images based on the multi-phase initial reconstruction images, and then the first data fidelity layer performs K-space data fidelity processing on the multi-phase intermediate reconstruction images, and performs multiple reconstruction processes in this way. The dynamic reconstruction images are determined based on the multi-phase intermediate reconstruction images finally output by the first data fidelity layer. Exemplarily, the computer device may directly generate the dynamic reconstruction images using the multi-phase intermediate reconstruction images finally output by the three-dimensional neural network layer / the first data fidelity layer.
[0069] In the embodiments of the present application, multi-phase initial reconstruction images are reconstructed based on the K-space data of multiple phases, and at least one reconstruction process is performed on the multi-phase initial reconstruction images according to the image reconstruction model to obtain the dynamic reconstruction images. Among them, the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer. The at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine the multi-phase intermediate reconstruction images, and using the multi-phase intermediate reconstruction images of the current reconstruction as the multi-phase initial reconstruction images for the next reconstruction process. Among them, the first data fidelity layer is used to perform K-space data fidelity processing on the input images; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images based on the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process. In the above method, the image reconstruction model is trained based on sharing the network parameters of the three-dimensional neural network layer, which reduces the data generated during the training process, reduces the video memory occupancy of the device carrying the model, realizes the full training of the model under limited video memory conditions, speeds up the convergence speed of the entire image reconstruction model, improves the convergence effect, and correspondingly improves the image quality of the images reconstructed based on the entire image reconstruction model.
[0070] In one of the embodiments, the K-space data of at least two adjacent phases in the above-mentioned K-space data of multiple phases are acquired in an interleaved manner.
[0071] Among them, the K-space data of at least two adjacent phases obtained by interleaved acquisition characterize that the encoding positions of the K-space data of these two adjacent phases are different, that is, the sampling data corresponding to different positions of the scanned object. If the K-space data of adjacent phases are all acquired in an interleaved manner, then after fusing the K-space data of multiple phases, the sampling data of most positions of the scanned object are obtained, even though these data are not acquired in the same phase.
[0072] Optionally, the K-space data of multiple phases can be obtained by data mask sampling. For example, as Figure 3 shown, each column in the figure represents a mask schematic diagram of the K-space data of one phase, and the K-space data of each phase are under-sampled. It can be seen from the figure that the K-space data sets of at least two adjacent phases are acquired in an interleaved manner. To improve the reconstruction efficiency, in one embodiment, as Figure 4 shown, the above S220, performing at least one reconstruction process on the multi-phase initial reconstruction image according to the image reconstruction model to obtain a dynamic reconstruction image, includes:
[0073] S410, performing a reconstruction process on the multi-phase initial reconstruction image a first preset number of times through the first data fidelity layer and the three-dimensional neural network layer to obtain a multi-phase target intermediate reconstruction image.
[0074] Among them, the reconstruction process is essentially to denoise and remove artifacts from the image.
[0075] Optionally, the computer device can input the multi-phase initial reconstruction image into the image reconstruction model, perform a reconstruction process on the multi-phase initial reconstruction image through the first data fidelity layer and the three-dimensional neural network layer, and count the number of executions until the number of executions reaches the first preset number to obtain a multi-phase target intermediate reconstruction image.
[0076] S420, determining a dynamic reconstruction image according to the multi-phase target intermediate reconstruction image.
[0077] Optionally, after the computer device iteratively executes the image reconstruction operation to obtain a multi-phase target intermediate reconstruction image, it can directly generate a dynamic reconstruction image based on the multi-phase target intermediate reconstruction image, or can further correct the artifacts of the multi-phase target intermediate reconstruction image to generate a dynamic reconstruction image according to the corrected multi-phase target intermediate reconstruction image.
[0078] In the embodiments of the present application, through the first data fidelity layer and the three-dimensional neural network layer, a reconstruction process is performed on the multi-phase initial reconstruction image a first preset number of times to obtain a multi-phase target intermediate reconstruction image, and then a dynamic reconstruction image is determined according to the multi-phase target intermediate reconstruction image. In the above method, by setting the first preset number, a limited number of reconstruction processes are performed, and the result is output in a timely manner, thereby improving the reconstruction efficiency.
[0079] The image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer. Thus, in one embodiment, as Figure 5 shown, S420 above, determining the dynamic reconstruction image according to the multi-phase target intermediate reconstruction image, includes:
[0080] S510, inputting the multi-phase target intermediate reconstruction image and the multi-phase initial reconstruction image into the second data fidelity layer to perform fidelity processing on the K-space data.
[0081] Among them, the second data fidelity layer has the same function as the first data fidelity layer, and is used to perform fidelity processing on the K-space data of the input image.
[0082] Optionally, the computer device may input the multi-phase target intermediate reconstruction image obtained through the first preset number of reconstruction processes into the second data fidelity layer to perform fidelity processing on the K-space data of the multi-phase target intermediate reconstruction image, obtain the multi-phase target intermediate reconstruction image after fidelity processing, and input the multi-phase initial reconstruction image into the second data fidelity layer to perform fidelity processing on the K-space data of the multi-phase initial reconstruction image, obtain the multi-phase initial reconstruction image after fidelity processing.
[0083] S520, inputting the multi-phase target intermediate reconstruction image after fidelity processing and the multi-phase initial reconstruction image into the two-dimensional neural network layer, and the two-dimensional neural network layer corrects the target intermediate reconstruction image of the corresponding phase according to the initial reconstruction image of each phase to obtain the corrected multi-phase target intermediate reconstruction image.
[0084] Among them, the two-dimensional neural network layer, i.e., the 2D CNN layer, is used to extract the features of the input multi-phase initial reconstruction images, and correct the target intermediate reconstruction image of the corresponding phase according to the independent features of the initial reconstruction images of each phase to obtain the corrected multi-phase target intermediate reconstruction image..
[0085] Optionally, the computer device may input the multi-phase target intermediate reconstruction image after fidelity processing and the multi-phase initial reconstruction image into the 2D CNN layer together, and the 2D CNN layer corrects the target intermediate reconstruction image of the corresponding phase according to the initial reconstruction image of each phase to obtain the corrected multi-phase target intermediate reconstruction image. For example, the multi-phase initial reconstruction images input into the 2D CNN layer include images 1 to 3, and the multi-phase target intermediate reconstruction images input into the 2D CNN layer include images a to c. The 2D CNN layer can correct image a based on the features of image 1 to output image a', correct image b based on the features of image 2 to output image b', and correct image c based on the features of image 3 to output image c', that is, obtain the corrected multi-phase target intermediate reconstruction images a' to c'.
[0086] S530. Obtain a dynamic reconstruction image based on the corrected multi-phase target intermediate reconstruction image.
[0087] Optionally, after obtaining the corrected multi-phase target intermediate reconstruction image output by the 2D CNN layer, the computer device can directly generate a dynamic reconstruction image based on the multi-phase target intermediate reconstruction image.
[0088] In the embodiments of the present application, the multi-phase target intermediate reconstruction image and the multi-phase initial reconstruction image are input into the second data fidelity layer for K-space data fidelity processing, and the multi-phase target intermediate reconstruction image and the multi-phase initial reconstruction image after the fidelity processing are input into the two-dimensional neural network layer. The two-dimensional neural network layer corrects the target intermediate reconstruction image of the corresponding phase according to the initial reconstruction image of each phase to obtain the corrected multi-phase target intermediate reconstruction image, and further obtains a dynamic reconstruction image based on the corrected multi-phase target intermediate reconstruction image. In the above method, the two-dimensional neural network layer realizes the correction of the target intermediate reconstruction image of each phase based on the characteristics of the initial reconstruction image of each phase to remove the artifact interference of the reconstruction images of other phases, further improving the image quality of the obtained dynamic reconstruction image.
[0089] To improve the image restoration degree, in one embodiment, as Figure 6 shown, the above reconstruction of the multi-phase initial reconstruction image through the first data fidelity layer and the three-dimensional neural network layer to determine the multi-phase intermediate reconstruction image includes:
[0090] S610. Input the multi-phase initial reconstruction image into the first data fidelity layer to obtain a multi-phase reference image.
[0091] Optionally, the computer device can first input the multi-phase initial reconstruction image into the first data fidelity layer to perform K-space data fidelity processing on the multi-phase initial reconstruction image to obtain a multi-phase reference image.
[0092] S620. Input the multi-phase reference image into the three-dimensional neural network layer to obtain a multi-phase intermediate reconstruction image.
[0093] Optionally, the computer device can input the multi-phase reference image output by the first data fidelity layer into the three-dimensional neural network layer for AI processing to comprehensively determine the output multi-phase image as the multi-phase intermediate reconstruction image to complete a reconstruction process.
[0094] In the embodiments of the present application, by inputting multi-phase initial reconstruction images into the first data fidelity layer, multi-phase reference images are obtained, and then the multi-phase reference images are input into the three-dimensional neural network layer to obtain multi-phase intermediate reconstruction images. In the above method, the reconstruction process first performs the fidelity processing of K-space data, and then uses the multi-phase images to comprehensively determine the output to implement AI processing, avoiding the influence of AI processing on the fidelity processing and improving the image restoration degree of the image process.
[0095] In one of the embodiments, the training process of the image reconstruction model includes:
[0096] Using multiple groups of multi-phase sample images to train the initial image reconstruction model in the image reconstruction model to obtain the image reconstruction model; wherein, the training process of each group of multi-phase sample images includes:
[0097] Inputting the multi-phase sample images into the initial image reconstruction model to execute the reconstruction process, and updating the network parameters of the three-dimensional neural network layer once every second preset number of executions until the images output by two adjacent reconstruction processes meet the preset difference condition.
[0098] Among them, the preset difference condition is the stopping condition. In the embodiments of the present application, the stopping condition is that the output result of the image reconstruction model reaches a fixed point, that is, the images output by the image reconstruction model in two adjacent reconstruction processes are basically the same. Exemplarily, whether the preset difference condition is satisfied can be judged based on the fixed point determined by the Find FixedPoint (FFP) algorithm.
[0099] The FFP algorithm is described as follows:
[0100]
[0101] Optionally, when the computer device trains the initial image reconstruction model, multiple groups of multi-phase sample images can be created in advance, and the multiple groups of multi-phase sample images are divided into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to evaluate the current performance of the model, and the test set is used to test the generalization performance of the model, so as to realize the training of the initial image reconstruction model and obtain the image reconstruction model.
[0102] For each group of sample images, the computer device inputs the multi-phase sample images into the initial image reconstruction model to execute the reconstruction process, and counts the number of executions. When the number of executions reaches the second preset number, the network parameters of the three-dimensional neural network layer are updated once to continue the reconstruction process and re-count the number of executions, and continue to update the network parameters of the three-dimensional neural network layer once when the number of executions reaches the second preset number, and so on, until the images output by the image reconstruction model in two adjacent reconstruction processes meet the preset difference condition.
[0103] Exemplarily, the preset difference condition includes that the image similarity is greater than a preset threshold. Based on this, during the training process of the image reconstruction model, the computer device can continuously obtain the image similarity between the images output by the image reconstruction model in two adjacent reconstruction processes, so as to determine whether the preset difference condition is met according to the image similarity. Among them, when the image similarity is greater than or equal to the preset threshold, it indicates that the preset difference condition is met, which means that the difference between the current reconstruction result and the previous reconstruction result is small and almost the same. It can be considered that the current model has converged, and the current result is recorded as a fixed point. At the same time, it is defaulted that the subsequent loop has little effect on improving the reconstruction effect, and the reconstruction process can be terminated; otherwise, when the image similarity is less than the preset threshold, it indicates that the preset difference condition is not met, which means that the difference between the current reconstruction result and the previous reconstruction result is large, and the current model has not converged, and the reconstruction process needs to continue.
[0104] Each reconstruction process can be encapsulated into an unrolling block, which includes the fidelity processing of the DC layer and the AI processing of the 3D CNN layer. The k-th reconstruction process corresponds to k unrolling blocks. The DEQ that realizes network parameter sharing greatly enriches the choice of the number of unrolling blocks. In theory, DEQ allows the selection and use of an infinite number of unrolling blocks, especially when supplemented with the "Jacobian-free" method, and can also solve the problem of out-of-memory explosion during training. To a certain extent, the introduction of DEQ significantly improves the model reconstruction ability. In order to improve the training efficiency, in the embodiments of the present application, DEQ and FFP are used to realize the adaptive adjustment of the number of unrolling blocks.
[0105] Optionally, when the image reconstruction model includes a three-dimensional neural network layer and a two-dimensional neural network layer, the three-dimensional neural network layer and the two-dimensional neural network layer can be trained synchronously. The specific process is as follows:
[0106] Input the initial multi-phase sample images into the first DC layer for fidelity processing of K-space data, and then input the multi-phase sample images after fidelity processing into the 3D CNN layer to comprehensively determine the output using the multi-phase sample images after fidelity processing, obtaining the multi-phase sample images output by one reconstruction process; then input the multi-phase sample images output by one reconstruction process into the first DC layer to implement the loop execution of the reconstruction process. When the number of executions reaches the second preset number, input the multi-phase sample images output by the 3D CNN layer and the initial multi-phase sample images into the second DC layer for fidelity processing, and then input the multi-phase sample images after fidelity processing and the initial multi-phase sample images into the 2D CNN layer to correct the multi-phase sample images of the corresponding phase output by the 3D CNN layer using the initial multi-phase sample images after fidelity processing, obtaining the corrected multi-phase sample images, and synchronously updating the network parameters of the 3D CNN layer and the 2D CNN layer at the same time. Repeat this process until the images output by the image reconstruction model in two adjacent reconstruction processes meet the preset difference condition.
[0107] In the embodiments of the present application, multiple groups of sample images are used to train the initial image reconstruction model to obtain the image reconstruction model. Among them, the training process of each group of sample images includes: inputting the sample images into the initial image reconstruction model for iterative reconstruction, and updating the network parameters of the three-dimensional neural network layer once every second preset number of iterations until the images output by two adjacent reconstruction processes meet the preset difference condition. In the above method, updating the network parameters once in multiple reconstruction processes realizes the training process of network parameter sharing, which helps to achieve a sufficient number of reconstruction processes, improves the reconstruction ability of the entire image reconstruction image, makes the convergence effect of the entire image reconstruction model better, and correspondingly improves the reliability of the reconstruction result and the image quality.
[0108] For the convenience of understanding by those skilled in the art, the following provides a detailed introduction to the image reconstruction method provided in the present application, as Figure 7 shown, the method may include:
[0109] S701. Use multiple groups of multi-phase sample images to train the initial image reconstruction model to obtain the image reconstruction model;
[0110] S702. The training process of each group of multi-phase sample images includes:
[0111] Input the multi-phase sample images into the initial image reconstruction model to execute the reconstruction process, and update the network parameters of the three-dimensional neural network layer in the initial image reconstruction model once every second preset number of executions until the images output by two adjacent reconstruction processes meet the preset difference condition;
[0112] S703. Reconstruct the multi-phase initial reconstruction images according to the K-space data of multiple phases;
[0113] S704. Input the multi-phase initial reconstructed images into the first data fidelity layer in the image reconstruction model to obtain multi-phase reference images; input the multi-phase reference images into the three-dimensional neural network layer in the image reconstruction model to obtain multi-phase intermediate reconstructed images, and use the multi-phase intermediate reconstructed images of the current reconstruction as the multi-phase initial reconstructed images for the next reconstruction process, so as to perform the reconstruction process for the first preset number of times to obtain multi-phase target intermediate reconstructed images;
[0114] S705. Input the multi-phase target intermediate reconstructed images and the multi-phase initial reconstructed images into the second data fidelity layer in the image reconstruction model to perform the fidelity processing of the K-space data;
[0115] S706. Input the multi-phase target intermediate reconstructed images and the multi-phase initial reconstructed images after the fidelity processing into the two-dimensional neural network layer, and use the two-dimensional neural network layer to correct the target intermediate reconstructed images of the corresponding phases according to the initial intermediate reconstructed images of each phase to obtain the corrected multi-phase target intermediate reconstructed images;
[0116] S707. Obtain the dynamic reconstructed images according to the corrected multi-phase target intermediate reconstructed images.
[0117] It should be noted that for the descriptions in S701 - S707 above, reference can be made to the relevant descriptions in the above embodiments, and their effects are similar. Therefore, this embodiment will not be elaborated here.
[0118] As Figure 8 shown, the trained image reconstruction model includes a 3D CNN layer and a 2D CNN layer, as well as a DC layer connected to the 3D CNN layer and a DC layer connected to the 2D CNN layer. The computer device performs phase encoding on the magnetic resonance signals obtained by magnetic resonance scanning to obtain under-sampled K-space data of multiple phases, and performs time dimension averaging, such as performing weighted averaging to form combined K-space data. Based on the combined K-space data, the CSM is determined, and the multi-phase initial reconstructed images M L are calibrated by using the CSM. When M L is input into the image reconstruction model, the reconstruction process of the first preset number of times (the Jacobian-free matrix calculation loop is performed k times) is executed via the DC layer and the 3D CNN layer to obtain multi-phase target intermediate reconstructed images. Then, the multi-phase target intermediate reconstructed images and the multi-phase initial reconstructed images M L are input into the DC layer and the 2D CNN layer for processing, and finally the multi-phase target intermediate reconstructed images M H corrected by the 2D CNN layer are output. The dynamic reconstructed images are generated from the multi-phase target intermediate reconstructed images M H .
[0119] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.
[0120] Based on the same inventive concept, an embodiment of the present application further provides an image reconstruction device for implementing the above-mentioned image reconstruction method. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the following image reconstruction devices can refer to the limitations on the image reconstruction method in the above text, and will not be repeated here.
[0121] In one embodiment, as Figure 9 shown, an image reconstruction device is provided, including: an initial reconstruction module 901 and an iterative reconstruction module 902, where:
[0122] The initial reconstruction module 901 is used to reconstruct multi-phase initial reconstruction images based on the K-space data of multiple phases;
[0123] The iterative reconstruction module 902 is used to perform at least one reconstruction process on the multi-phase initial reconstruction images according to the image reconstruction model to obtain dynamic reconstruction images; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer; the at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstruction images, and using the multi-phase intermediate reconstruction images of the current reconstruction as the multi-phase initial reconstruction images of the next reconstruction process; where the first data fidelity layer is used to perform fidelity processing on the input images for K-space data; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process.
[0124] In one of the embodiments, the K-space data of at least two adjacent phases in the K-space data of multiple phases are acquired alternately.
[0125] In one of the embodiments, the iterative reconstruction module 902 includes:
[0126] The first reconstruction sub-module is configured to perform a first preset number of reconstruction processes on the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to obtain multi-phase target intermediate reconstruction images;
[0127] The second reconstruction sub-module is configured to determine dynamic reconstruction images based on the multi-phase target intermediate reconstruction images.
[0128] In one embodiment, the image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer; the second reconstruction sub-module includes:
[0129] The fidelity unit is configured to input the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images into the second data fidelity layer to perform fidelity processing on the K-space data;
[0130] The correction unit is configured to input the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images after the fidelity processing into the two-dimensional neural network layer, and correct the corresponding-phase target intermediate reconstruction images according to the initial reconstruction images of each phase through the two-dimensional neural network layer to obtain the corrected multi-phase target intermediate reconstruction images;
[0131] The reconstruction unit is configured to obtain dynamic reconstruction images based on the corrected multi-phase target intermediate reconstruction images.
[0132] In one embodiment, the iterative reconstruction module 902 includes:
[0133] The reference sub-module is configured to input the multi-phase initial reconstruction images into the first data fidelity layer to obtain multi-phase reference images;
[0134] The iterative sub-module is configured to input the multi-phase reference images into the three-dimensional neural network layer to obtain multi-phase intermediate reconstruction images.
[0135] In one embodiment, the image reconstruction device further includes:
[0136] The model training module is configured to train the initial image reconstruction model with multiple groups of multi-phase sample images to obtain the image reconstruction model; wherein, the training process of each group of multi-phase sample images includes: inputting the multi-phase sample images into the initial image reconstruction model to perform the reconstruction process, and updating the network parameters of the three-dimensional neural network layer every second preset number of times until the images output by two adjacent reconstruction processes meet the preset difference condition.
[0137] Each module in the above image reconstruction device can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor of the computer device in hardware form or be independent of it, or be stored in the memory of the computer device in software form so that the processor can call and execute the operations corresponding to the above-mentioned modules.
[0138] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:
[0139] Reconstructing multi-phase initial reconstruction images based on multi-phase K-space data; performing at least one reconstruction process on the multi-phase initial reconstruction images according to an image reconstruction model to obtain dynamic reconstruction images; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer;
[0140] The at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstruction images, and using the multi-phase intermediate reconstruction images reconstructed in the current time as the multi-phase initial reconstruction images for the next reconstruction process; wherein, the first data fidelity layer is used for performing fidelity processing on the K-space data of the input images; the three-dimensional neural network layer is used for comprehensively determining the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process.
[0141] In one embodiment, at least two adjacent-phase K-space data in the multi-phase K-space data are acquired in an interleaved manner.
[0142] In one embodiment, when the processor executes the computer program, the following steps are further implemented:
[0143] Performing a first preset number of reconstruction processes on the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to obtain multi-phase target intermediate reconstruction images; determining dynamic reconstruction images according to the multi-phase target intermediate reconstruction images.
[0144] In one embodiment, the image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer; when the processor executes the computer program, the following steps are further implemented:
[0145] Inputting the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images into the second data fidelity layer to perform fidelity processing on the K-space data; inputting the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images after the fidelity processing into the two-dimensional neural network layer, and correcting the corresponding-phase target intermediate reconstruction images according to the initial reconstruction images of each phase through the two-dimensional neural network layer to obtain corrected multi-phase target intermediate reconstruction images; obtaining dynamic reconstruction images according to the corrected multi-phase target intermediate reconstruction images.
[0146] In one embodiment, when the processor executes the computer program, the following steps are further implemented:
[0147] Input the multi-phase initial reconstruction images into the first data fidelity layer to obtain multi-phase reference images; input the multi-phase reference images into the three-dimensional neural network layer to obtain multi-phase intermediate reconstruction images.
[0148] In one embodiment, when the processor executes the computer program, the following steps are further implemented:
[0149] Use multiple groups of multi-phase sample images to train the initial image reconstruction model to obtain an image reconstruction model; among them, the training process of each group of multi-phase sample images includes:
[0150] Input the multi-phase sample images into the initial image reconstruction model to execute the reconstruction process, and update the network parameters of the three-dimensional neural network layer every time the second preset number of times is executed until the images output by two adjacent reconstruction processes meet the preset difference condition.
[0151] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:
[0152] Reconstruct multi-phase initial reconstruction images according to the K-space data of multiple phases; perform at least one reconstruction process on the multi-phase initial reconstruction images according to the image reconstruction model to obtain dynamic reconstruction images; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer;
[0153] The at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstruction images, and using the multi-phase intermediate reconstruction images reconstructed in the current time as the multi-phase initial reconstruction images for the next reconstruction process; among them, the first data fidelity layer is used to perform fidelity processing on the input images for K-space data; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process.
[0154] In one of the embodiments, the K-space data of at least two adjacent phases in the K-space data of multiple phases are acquired in an interleaved manner.
[0155] In one embodiment, when the computer program is executed by the processor, the following steps are further implemented:
[0156] Perform a first preset number of reconstruction processes on the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to obtain multi-phase target intermediate reconstruction images; determine dynamic reconstruction images according to the multi-phase target intermediate reconstruction images.
[0157] In one embodiment, the image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer; when the computer program is executed by a processor, the following steps are further implemented:
[0158] Input the multi-phase target intermediate reconstruction image and the multi-phase initial reconstruction image into the second data fidelity layer for K-space data fidelity processing; input the multi-phase target intermediate reconstruction image and the multi-phase initial reconstruction image after the fidelity processing into the two-dimensional neural network layer, and the two-dimensional neural network layer corrects the target intermediate reconstruction image of the corresponding phase according to the initial reconstruction image of each phase to obtain the corrected multi-phase target intermediate reconstruction image; obtain the dynamic reconstruction image according to the corrected multi-phase target intermediate reconstruction image.
[0159] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:
[0160] Input the multi-phase initial reconstruction image into the first data fidelity layer to obtain the multi-phase reference image; input the multi-phase reference image into the three-dimensional neural network layer to obtain the multi-phase intermediate reconstruction image.
[0161] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:
[0162] Use multiple groups of multi-phase sample images to train the initial image reconstruction model to obtain the image reconstruction model; wherein, the training process of each group of multi-phase sample images includes:
[0163] Input the multi-phase sample images into the initial image reconstruction model to execute the reconstruction process, and update the network parameters of the three-dimensional neural network layer every second preset number of executions until the images output by two adjacent reconstruction processes meet the preset difference condition.
[0164] In one embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the following steps are implemented:
[0165] Reconstruct the multi-phase initial reconstruction image according to the multi-phase K-space data; perform at least one reconstruction process on the multi-phase initial reconstruction image according to the image reconstruction model to obtain the dynamic reconstruction image; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer;
[0166] At least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through a first data fidelity layer and a three-dimensional neural network layer to determine the multi-phase intermediate reconstruction images, and using the multi-phase intermediate reconstruction images of the current reconstruction as the multi-phase initial reconstruction images of the next reconstruction process; wherein, the first data fidelity layer is used to perform fidelity processing on the K-space data of the input images; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in at least one reconstruction process.
[0167] In one embodiment, at least two adjacent phase K-space data in the multi-phase K-space data are acquired interleaved.
[0168] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:
[0169] Through the first data fidelity layer and the three-dimensional neural network layer, perform the reconstruction process for the first preset number of times on the multi-phase initial reconstruction images to obtain the multi-phase target intermediate reconstruction images; determine the dynamic reconstruction images according to the multi-phase target intermediate reconstruction images.
[0170] In one embodiment, the image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer; when the computer program is executed by a processor, the following steps are further implemented:
[0171] Input the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images into the second data fidelity layer to perform fidelity processing on the K-space data; input the multi-phase target intermediate reconstruction images and the multi-phase initial reconstruction images after the fidelity processing into the two-dimensional neural network layer, and use the two-dimensional neural network layer to correct the target intermediate reconstruction images of the corresponding phases according to the initial reconstruction images of each phase to obtain the corrected multi-phase target intermediate reconstruction images; obtain the dynamic reconstruction images according to the corrected multi-phase target intermediate reconstruction images.
[0172] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:
[0173] Input the multi-phase initial reconstruction images into the first data fidelity layer to obtain the multi-phase reference images; input the multi-phase reference images into the three-dimensional neural network layer to obtain the multi-phase intermediate reconstruction images.
[0174] In one embodiment, when the computer program is executed by a processor, the following steps are further implemented:
[0175] Train the initial image reconstruction model with multiple groups of multi-phase sample images to obtain the image reconstruction model; wherein, the training process of each group of multi-phase sample images includes:
[0176] Input multi-phase sample images into the initial image reconstruction model to perform the reconstruction process. Update the network parameters of the three-dimensional neural network layer every second preset number of times until the images output by two adjacent reconstruction processes meet the preset difference condition.
[0177] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium provided in the various embodiments of the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the various embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the various embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.
[0178] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.
[0179] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.
Claims
1. An image reconstruction method, characterized in that, The method includes: Reconstructing multi-phase initial reconstructed images based on multi-phase K-space data; Performing at least one reconstruction process on the multi-phase initial reconstructed images according to an image reconstruction model to obtain dynamic reconstructed images; the image reconstruction model includes a data fidelity layer and a three-dimensional neural network layer; The at least one reconstruction process includes: Reconstructing the multi-phase initial reconstructed images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstructed images, and using the multi-phase intermediate reconstructed images of the current reconstruction as the multi-phase initial reconstructed images of the next reconstruction process; Wherein, the first data fidelity layer is used to perform fidelity processing of K-space data on the input images; the three-dimensional neural network layer is used to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared in the at least one reconstruction process.
2. The method according to claim 1, wherein At least two adjacent-phase K-space data in the multi-phase K-space data are acquired interleaved.
3. The method according to claim 1, wherein The performing at least one reconstruction process on the multi-phase initial reconstructed images according to the image reconstruction model to obtain dynamic reconstructed images includes: Performing a first preset number of reconstruction processes on the multi-phase initial reconstructed images through the first data fidelity layer and the three-dimensional neural network layer to obtain multi-phase target intermediate reconstructed images; Determining the dynamic reconstructed images according to the multi-phase target intermediate reconstructed images.
4. The method according to claim 3, wherein The image reconstruction model further includes a second data fidelity layer and a two-dimensional neural network layer; the determining the dynamic reconstructed images according to the multi-phase target intermediate reconstructed images includes: Inputting the multi-phase target intermediate reconstructed images and the multi-phase initial reconstructed images into the second data fidelity layer to perform fidelity processing of K-space data; Inputting the multi-phase target intermediate reconstructed images and the multi-phase initial reconstructed images after the fidelity processing into the two-dimensional neural network layer, and correcting the target intermediate reconstructed images of the corresponding phases according to the initial reconstructed images of each phase through the two-dimensional neural network layer to obtain corrected multi-phase target intermediate reconstructed images; Obtaining the dynamic reconstructed images according to the corrected multi-phase target intermediate reconstructed images.
5. The method according to any one of claims 1-4, characterized in that, The reconstructing the multi-phase initial reconstructed images through the first data fidelity layer and the three-dimensional neural network layer to determine multi-phase intermediate reconstructed images includes: Inputting the multi-phase initial reconstructed images into the first data fidelity layer to obtain multi-phase reference images; Inputting the multi-phase reference images into the three-dimensional neural network layer to obtain the multi-phase intermediate reconstructed images.
6. The method according to any one of claims 1-4, characterized in that, The training process of the image reconstruction model includes: Training an initial image reconstruction model with multiple groups of multi-phase sample images to obtain the image reconstruction model; wherein, the training process of each group of multi-phase sample images includes: Inputting the multi-phase sample images into the initial image reconstruction model to perform the reconstruction process, and updating the network parameters of the three-dimensional neural network layer every second preset number of executions until the images output by two adjacent reconstruction processes meet the preset difference condition.
7. An image reconstruction device, characterized in that, The device includes: An initial reconstruction module, configured to reconstruct multi-phase initial reconstruction images based on multi-phase K-space data; An iterative reconstruction module, configured to perform at least one reconstruction process on the multi-phase initial reconstruction images according to an image reconstruction model to obtain dynamic reconstruction images; the image reconstruction model includes a first data fidelity layer and a three-dimensional neural network layer; the at least one reconstruction process includes: reconstructing the multi-phase initial reconstruction images through the first data fidelity layer and the three-dimensional neural network layer to determine intermediate reconstruction images, and using the intermediate reconstruction images of the current reconstruction as the multi-phase initial reconstruction images for the next reconstruction process; wherein, the first data fidelity layer is configured to perform fidelity processing on the input images for K-space data; the three-dimensional neural network layer is configured to comprehensively determine the output multi-phase images according to the input multi-phase images, and the network parameters of the three-dimensional neural network layer are shared during the at least one reconstruction process.
8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.