A CT image reconstruction method and apparatus

Abstract

Embodiments of the present application provide a CT image reconstruction method and device, the method comprising obtaining projection data of a first object, then the projection data can be input to a CT iterative unfolding reconstruction network for processing to obtain a CT image corresponding to the first object, wherein the CT iterative unfolding reconstruction network is obtained by neural network unfolding of a fidelity term and a compressed sensing regularization term or a total variation (TV) regularization term in a CT iterative reconstruction algorithm. In this way, the method directly performs neural network unfolding on the compressed sensing regularization term or the total variation (TV) regularization term in the CT iterative reconstruction algorithm, thereby obtaining a complete CT iterative unfolding reconstruction network, which can improve the explainability of the overall CT reconstruction network and thus improve the imaging quality of the reconstructed CT image.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a CT image reconstruction method and apparatus. Background Technology

[0002] Computed tomography (CT) technology detects X-rays passing through an object from different angles, reconstructs the linear attenuation coefficient of the entire object, and ultimately reflects the internal material and structure of the object, enhancing people's ability to observe the internal structure of objects. CT technology has advantages such as fast measurement speed, low detection cost, high image quality, and few limitations, and is therefore widely used in medical diagnosis and treatment, image-guided interventional procedures, and industrial non-destructive testing.

[0003] In recent years, with the widespread application of CT technology, CT reconstruction algorithms have also made significant progress. Especially with the rise of artificial intelligence technology, deep learning has gradually become a major reconstruction method in CT, and many problems existing in traditional CT reconstruction methods are expected to be better solved with the help of deep learning. The main idea of ​​applying deep learning technology to CT reconstruction is to integrate neural networks into the reconstruction process. For example, the weighting, filtering, and backprojection calculation steps in analytical reconstruction algorithms can be implemented using neural networks, or the iterative calculation steps in iterative reconstruction algorithms can be replaced by neural networks.

[0004] Applying deep learning technology to CT reconstruction has significantly improved the imaging quality of CT images. However, due to the black-box nature of deep neural networks, combining deep learning with CT reconstruction methods still faces challenges, namely the robustness and generalization of the reconstruction network. The lack of interpretability in deep neural networks means that the designed network heavily relies on the quality of the dataset, requiring massive amounts of high-precision data for training, and even slight changes in the distribution of the input data can lead to incorrect calculation results. To overcome the lack of interpretability in traditional deep learning methods, researchers have proposed constructing deep neural networks by unfolding model-based iterative algorithms. Several solutions have been proposed for this approach, but these solutions either require extensive approximation calculations, leading to a decrease in the computational accuracy of the CT reconstruction method, thus reducing the computational accuracy of the iteratively unfolded neural network, or they directly replace the compressed sensing regularization term with a neural network without unfolding the network, resulting in reduced interpretability of the constructed CT reconstruction network. Summary of the Invention

[0005] This application provides a CT image reconstruction method and apparatus to improve the interpretability of the overall CT reconstruction network, thereby improving the imaging quality of the reconstructed CT images.

[0006] In a first aspect, embodiments of this application provide a CT image reconstruction method, including:

[0007] Obtain the projection data of the first object;

[0008] The projection data is input into the CT iterative unfolding reconstruction network to obtain the target CT image corresponding to the first object;

[0009] The CT iterative unfolding reconstruction network is obtained by unfolding the fidelity term and the compressed sensing regularization term or the total variational TV regularization term in the CT iterative reconstruction algorithm into a neural network.

[0010] In the above technical solution, by directly expanding the neural network for the fidelity term and the compressed sensing regularization term or the total variational TV regularization term in the CT iterative reconstruction algorithm, a complete CT iterative expansion reconstruction network can be obtained. This can improve the interpretability of the overall CT reconstruction network, thereby improving the imaging quality of the reconstructed CT images.

[0011] In one possible implementation, the projection data is obtained by the CT system scanning the first object according to system parameters, or the projection data is obtained by the computing device simulating the first object according to the system parameters.

[0012] The above technical solution offers flexible and diverse methods for acquiring projection data, which can meet the needs of different users and thus make the solution applicable to a wide range of scenarios.

[0013] In one possible implementation, the CT iterative unfolding reconstruction network comprises m sub-networks processed sequentially in series, wherein the m sub-networks have the same network structure, and m is an integer greater than 1.

[0014] In the above technical solution, by using multiple serially connected sub-networks, effective iterative processing of projection data can be achieved to obtain reconstructed CT images, and the interpretability of the overall CT reconstruction network can be improved.

[0015] In one possible implementation, the projection data is input into a CT iterative unfolding reconstruction network to obtain a target CT image corresponding to the first object, including:

[0016] The projection data is sequentially passed through the m sub-networks to obtain the target CT image.

[0017] In the above technical solution, the projection data is iteratively processed by multiple serially connected sub-networks, which can effectively complete the reconstruction of the projection data and obtain a target CT image with relatively high visual quality.

[0018] In one possible implementation, the first sub-network located at the beginning of the m sub-networks is used as the input end of the CT iterative unfolding reconstruction network, and the m-th sub-network located at the end of the m sub-networks is used as the output end of the CT iterative unfolding reconstruction network. The input data acquired by the first sub-network includes the projection data, and the output data determined by the m-th sub-network includes the target CT image.

[0019] The i-th sub-network among the m sub-networks is used to determine multiple different types of second data based on the projection data and multiple different types of first data output by the (i-1)-th sub-network. The multiple different types of second data and the projection data are used as input data for the (i+1)-th sub-network. The multiple different types of second data correspond one-to-one with the multiple different types of first data. The i-th sub-network is any one of the other sub-networks among the m sub-networks, excluding the first sub-network and the m-th sub-network.

[0020] In the above technical solution, by inputting projection data and multiple different types of initial data into the first sub-network at the starting position, multiple different types of data can be obtained. The projection data and multiple different types of data are then input into the next sub-network, and so on, until the sub-network at the end position. In this way, high-quality CT images can be effectively iterated.

[0021] In one possible implementation, any sub-network includes an orthographic projection network layer, multiple back-projection network layers, and multiple convolutional neural network layers, wherein the multiple back-projection network layers are identical, and the multiple convolutional neural network layers are all different, and the multiple convolutional neural network layers include a first convolutional neural network layer, a second convolutional neural network layer, a third convolutional neural network layer, a fourth convolutional neural network layer, a fifth convolutional neural network layer, and a sixth convolutional neural network layer;

[0022] For the i-th sub-network, the multiple different types of first data acquired by the i-th sub-network include first data of a first type, first data of a second type, first data of a third type, and first data of a fourth type. The first data of the first type is used to indicate the CT image in the iterative state.

[0023] The first back-projection network layer included in the i-th sub-network is used to determine first back-projection data based on the projection data. The orthographic projection network layer included in the i-th sub-network is used to determine orthographic projection data based on first data of the first type. The second back-projection network layer included in the i-th sub-network is used to determine second back-projection data based on the orthographic projection data. The first convolutional neural network layer included in the i-th sub-network is used to determine first sparse transformation data based on first data of the first type. The second convolutional neural network layer included in the i-th sub-network is used to determine second sparse transformation data based on the first sparse transformation data. The first back-projection network layer is any one of the plurality of back-projection network layers, and the second back-projection network layer is any one of the other back-projection network layers besides the first back-projection network layer.

[0024] The third convolutional neural network layer included in the i-th sub-network is used to determine the second data of the second type based on the first back-projection data, the second back-projection data, and the first data of the second type; the fourth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the third type based on the second sparse transform data and the first data of the third type.

[0025] The fifth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the fourth type based on the second data of the second type, the second data of the third type, and the first data of the fourth type;

[0026] The sixth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the first type based on the second data of the fourth type and the first data of the fourth type.

[0027] In the above technical solution, for each sub-network, by inputting projection data and multiple different types of data into the orthographic projection network layer, multiple back-projection network layers and multiple convolutional neural network layers included in the sub-network, the feature extraction capability of the overall CT reconstruction network can be further improved, thereby effectively completing CT image reconstruction and improving the imaging quality of the reconstructed CT image.

[0028] In one possible implementation, the system parameters include at least one of the following: detector position, detector size and sampling interval, X-ray source position, rotation center position of the CT system, number of projection data and acquisition angle, size of the reconstructed object and sampling interval.

[0029] Secondly, embodiments of this application also provide a CT image reconstruction apparatus, comprising:

[0030] The acquisition module is used to acquire the projection data of the first object;

[0031] The processing module is used to input the projection data into the CT iterative unfolding reconstruction network to obtain the target CT image corresponding to the first object; wherein, the CT iterative unfolding reconstruction network is obtained by performing neural network unfolding on the fidelity term and the compressed sensing regularization term or the total variational TV regularization term in the CT iterative reconstruction algorithm.

[0032] In one possible implementation, the projection data is obtained by the CT system scanning the first object according to system parameters, or the projection data is obtained by the computing device simulating the first object according to the system parameters.

[0033] In one possible implementation, the CT iterative unfolding reconstruction network comprises m sub-networks processed sequentially in series, wherein the m sub-networks have the same network structure, and m is an integer greater than 1.

[0034] In one possible implementation, the processing module is specifically used for:

[0035] The projection data is sequentially passed through the m sub-networks to obtain the target CT image.

[0036] In one possible implementation, the first sub-network located at the beginning of the m sub-networks is used as the input end of the CT iterative unfolding reconstruction network, and the m-th sub-network located at the end of the m sub-networks is used as the output end of the CT iterative unfolding reconstruction network. The input data acquired by the first sub-network includes the projection data, and the output data determined by the m-th sub-network includes the target CT image.

[0037] The i-th sub-network among the m sub-networks is used to determine multiple different types of second data based on the projection data and multiple different types of first data output by the (i-1)-th sub-network. The multiple different types of second data and the projection data are used as input data for the (i+1)-th sub-network. The multiple different types of second data correspond one-to-one with the multiple different types of first data. The i-th sub-network is any one of the other sub-networks among the m sub-networks, excluding the first sub-network and the m-th sub-network.

[0038] In one possible implementation, any sub-network includes an orthographic projection network layer, multiple back-projection network layers, and multiple convolutional neural network layers, wherein the multiple back-projection network layers are identical, and the multiple convolutional neural network layers are all different, and the multiple convolutional neural network layers include a first convolutional neural network layer, a second convolutional neural network layer, a third convolutional neural network layer, a fourth convolutional neural network layer, a fifth convolutional neural network layer, and a sixth convolutional neural network layer;

[0039] For the i-th sub-network, the multiple different types of first data acquired by the i-th sub-network include first data of a first type, first data of a second type, first data of a third type, and first data of a fourth type. The first data of the first type is used to indicate the CT image in the iterative state.

[0040] The first back-projection network layer included in the i-th sub-network is used to determine first back-projection data based on the projection data. The orthographic projection network layer included in the i-th sub-network is used to determine orthographic projection data based on first data of the first type. The second back-projection network layer included in the i-th sub-network is used to determine second back-projection data based on the orthographic projection data. The first convolutional neural network layer included in the i-th sub-network is used to determine first sparse transformation data based on first data of the first type. The second convolutional neural network layer included in the i-th sub-network is used to determine second sparse transformation data based on the first sparse transformation data. The first back-projection network layer is any one of the plurality of back-projection network layers, and the second back-projection network layer is any one of the other back-projection network layers besides the first back-projection network layer.

[0041] The third convolutional neural network layer included in the i-th sub-network is used to determine the second data of the second type based on the first back-projection data, the second back-projection data, and the first data of the second type; the fourth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the third type based on the second sparse transform data and the first data of the third type.

[0042] The fifth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the fourth type based on the second data of the second type, the second data of the third type, and the first data of the fourth type;

[0043] The sixth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the first type based on the second data of the fourth type and the first data of the fourth type.

[0044] In one possible implementation, the system parameters include at least one of the following: detector position, detector size and sampling interval, X-ray source position, rotation center position of the CT system, number of projection data and acquisition angle, size of the reconstructed object and sampling interval.

[0045] Thirdly, embodiments of this application provide a computing device, including:

[0046] Memory, used to store computer programs;

[0047] The processor is used to call the computer program stored in the memory and execute the steps of the CT image reconstruction method according to the obtained program.

[0048] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer-executable program for causing a computer to perform the steps of a CT image reconstruction method.

[0049] Fifthly, embodiments of this application provide a computer program product, which includes a computer program or instructions that, when executed on a computer, cause the computer to perform the steps of a CT image reconstruction method. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A schematic flowchart of a CT image reconstruction method provided in an embodiment of this application;

[0052] Figure 2 A schematic diagram of a CT iterative unfolding reconstruction network provided in an embodiment of this application;

[0053] Figure 3 This is a schematic diagram of the structure of a sub-network provided in an embodiment of this application;

[0054] Figure 4 A schematic diagram of the structure of multiple network layers included in a sub-network provided in an embodiment of this application;

[0055] Figure 5 This is a schematic diagram of the structure of a CT image reconstruction device provided in an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The illustrative embodiments and descriptions provided in this application are for explaining the application, but are not intended to limit the application. Furthermore, elements / components using the same or similar reference numerals in the drawings and embodiments are used to represent the same or similar parts.

[0058] It should be understood that the terms "first," "second," etc., used in this application do not specifically refer to the order or sequence of events, nor are they used to limit this application. They can be used interchangeably where appropriate, and are only used to distinguish elements or operations described with the same technical terms.

[0059] Furthermore, the terms "comprising," "including," "having," and "containing" as used in this application are open-ended, meaning they include but are not limited to. Additionally, the term "and / or" as used in this application includes any or all of the stated items.

[0060] Figure 1 An exemplary flowchart illustrating a CT image reconstruction method provided in this application embodiment is shown. This method can be executed by a CT image reconstruction device or a component (such as a chip or circuit) capable of supporting the CT image reconstruction device in implementing the method. Exemplarily, the CT image reconstruction device can be a computing device with CT image reconstruction capabilities, or other devices with CT image reconstruction capabilities, or a functional component (such as a chip) with CT image reconstruction capabilities incorporated into the computing device or other devices. To facilitate the introduction of the CT image reconstruction scheme provided in this application embodiment, the following description uses the execution of the CT image reconstruction method by a CT image reconstruction device as an example.

[0061] like Figure 1 As shown, the method flow may include:

[0062] Step 101: The CT image reconstruction device acquires the projection data of the first object.

[0063] Step 102: The CT image reconstruction device inputs the projection data into the CT iterative unfolding reconstruction network to obtain the target CT image corresponding to the first object.

[0064] In step 101 above, the first object can be a human body or an object (such as an animal or other items), and this embodiment of the application does not limit this. In one example, the projection data of the first object can be acquired by the CT hardware system through scanning the first object according to system parameters. Then, the CT hardware system can transmit the projection data of the first object to the CT image reconstruction device. In another example, the projection data of the first object can also be obtained by a computing device through simulating the first object according to system parameters. Then, the computing device can transmit the projection data of the first object to the CT image reconstruction device.

[0065] The system parameters are the parameters required for the CT hardware system to acquire projection data or for the computing device to simulate projection data. For example, the system parameters may include at least one of the following: detector position, detector size and sampling interval, X-ray source position, rotation center position of the CT system, number of projection data and acquisition angle, size of the reconstructed object and sampling interval.

[0066] For example, taking the acquisition of projection data by a CT hardware system as an example, the CT hardware system uses cone-beam X-rays to irradiate a first object, uses a flat panel detector to perform imaging, and extracts projection data at the center position on the flat panel detector under different irradiation angles to form projection data g. For example, in this embodiment of the application, only the tomographic plane where the rotation center is located is reconstructed. As an example, the number of projections is 360, and the length of the projection data is 768 pixels.

[0067] In step 102 above, the CT iterative unfolding reconstruction network is obtained by the CT image reconstruction device through neural network unfolding of the fidelity term and the compressed sensing regularization term or the total variational TV regularization term in the CT iterative reconstruction algorithm. In one example, when the CT iterative unfolding reconstruction network is obtained by the CT image reconstruction device through neural network unfolding of the fidelity term and the compressed sensing regularization term in the CT iterative reconstruction algorithm, the CT image reconstruction device can use the primal-dual algorithm to solve the CT reconstruction problem based on the compressed sensing term, resulting in an iterative reconstruction algorithm that is easy to unfold into a neural network. Then, the CT image reconstruction device performs neural network unfolding on the CT iterative reconstruction algorithm, especially on the compressed sensing regularization term, and simultaneously performs neural network unfolding on the fidelity term, thus obtaining a complete CT iterative unfolding reconstruction network. In another example, when the CT iterative unfolding reconstruction network is obtained by the CT image reconstruction device through neural network unfolding of the fidelity term and the total variational TV regularization term in the CT iterative reconstruction algorithm, the CT image reconstruction device can use the primal-dual algorithm to solve the CT reconstruction problem based on the TV regularization term, resulting in an iterative reconstruction algorithm that is easy to unfold into a neural network. Subsequently, the CT image reconstruction device expands the CT iterative reconstruction algorithm into a neural network, especially the TV regularization term in the CT iterative reconstruction algorithm. At the same time, it also expands the fidelity term in the CT iterative reconstruction algorithm into a neural network, thus obtaining a complete CT iterative expansion reconstruction network.

[0068] The CT iterative unfolding reconstruction network can include m sub-networks processed sequentially in series. The m sub-networks have the same network structure, and m is an integer greater than 1. After acquiring the projection data of the first object, the CT image reconstruction device can sequentially pass the projection data of the first object through the m sub-networks to obtain the target CT image corresponding to the first object (i.e., the three-dimensional information of the first object, or, as can be understood, obtaining the internal information of the first object without damaging it). For example, in this embodiment, when the CT iterative unfolding reconstruction network is obtained by the CT image reconstruction device through simultaneous neural network unfolding of the fidelity term and the compressed sensing regularization term in the CT iterative reconstruction algorithm, the value of m can be 8; when the CT iterative unfolding reconstruction network is obtained by the CT image reconstruction device through simultaneous neural network unfolding of the fidelity term and the total variational TV regularization term in the CT iterative reconstruction algorithm, the value of m can be 10.

[0069] Figure 2 An exemplary schematic diagram of a CT iterative unfolding reconstruction network provided in an embodiment of this application is shown. Figure 2As shown, the CT iterative unfolding reconstruction network includes m subnetworks, namely subnetwork 1 (or the first subnetwork), subnetwork 2 (or the second subnetwork), subnetwork 3 (or the third subnetwork)... subnetwork m (or the m-th subnetwork). The first subnetwork (subnetwork 1), located at the beginning of the m subnetworks, serves as the input to the CT iterative unfolding reconstruction network, acquiring input data. The m-th subnetwork (subnetwork m), located at the end of the m subnetworks, serves as the output to the CT iterative unfolding reconstruction network, outputting target data. Specifically, for the first subnetwork, the CT image reconstruction device can input projection data and multiple different types of initial data into the first subnetwork to obtain multiple different types of data, and can use the projection data and multiple different types of data as input data for the next subnetwork adjacent to the first subnetwork. For the m-th subnetwork, the CT image reconstruction device can input projection data and multiple different types of data output from the previous subnetwork adjacent to the m-th subnetwork into the m-th subnetwork to obtain multiple different types of target data, including the target CT image corresponding to the first object.

[0070] Furthermore, for any subnetwork (e.g., the i-th subnetwork) among the m subnetworks excluding the first and m-th subnetworks, the CT image reconstruction device can input the projection data and multiple different types of first data output from the (i-1)-th subnetwork into the i-th subnetwork to obtain multiple different types of second data. Each of the multiple different types of second data corresponds one-to-one with a different type of first data. Then, the CT image reconstruction device can input the projection data and these multiple different types of second data into the (i+1)-th subnetwork, and so on, up to the m-th subnetwork, thereby obtaining the target CT image (i.e., the reconstructed CT image) corresponding to the first object.

[0071] In the embodiments of this application, by iteratively expanding according to the CT iterative reconstruction algorithm based on compressed sensing terms (or the CT iterative reconstruction algorithm based on TV regularization terms), and replacing the calculation steps of each iteration in the iterative algorithm with a neural network, the sub-network structure of each iteration can be obtained.

[0072] Among them, the primal-dual algorithm is used to solve the CT reconstruction problem based on compressed sensing terms (or the CT reconstruction problem based on TV regularization terms), resulting in an iterative CT reconstruction algorithm that is easy to expand into a neural network. The calculation formulas for each iteration in this iterative reconstruction algorithm are as follows:

[0073]

[0074]

[0075] u n+1 =u n -τp n+1 -τq n+1

[0076]

[0077] Where g represents the projection data. The matrix is ​​used to represent the sparse transformation matrix, A represents the system matrix, T represents the transpose operation, and σ, λ, τ, and θ represent the weight parameters, respectively. n q n u n These are used to represent intermediate variables, both with an initial value of 0. This is used to represent the input CT image for any subnetwork, with an initial value of 0. This is used to represent the CT image output by this subnetwork.

[0078] Figure 3 An exemplary schematic diagram of a sub-network structure provided in an embodiment of this application is shown. Figure 3 As shown, any subnetwork includes orthographic projection network layers (such as CNN). A ), multiple back-projection network layers (e.g., two) And multiple convolutional neural network layers. For example, an orthographic projection network layer (CNN) A This can be achieved by using a ray-driven method to construct the network layer, thus realizing the orthographic projection operation and backprojection operation in the iterative algorithm. The backprojection operation in the iterative algorithm can be implemented by constructing the network using a pixel-driven method. Both the forward and backprojection network layers are neural networks with no adjustable parameters.

[0079] Among them, the multiple convolutional neural network layers are all different, and the multiple convolutional neural network layers include the first convolutional neural network layer (e.g. ), the second convolutional neural network layer (e.g. ), the third convolutional neural network layer (such as CNN) p ), the fourth convolutional neural network layer (such as CNN) q ), the fifth convolutional neural network layer (such as CNN) u ) and the sixth convolutional neural network layer (e.g. ).

[0080] In the embodiments of this application, shallow convolutional neural networks (CNNs) are used respectively. p CNN q CNN u , Replace the variable calculations in the above iterations, and use shallow neural networks respectively. and Implementing sparse transformations and their transposes in compressed sensing (or TV regularization) can improve the feature extraction capabilities of the overall CT reconstruction network.

[0081] When sparse transformation Gradient operator for images When this happens, the original CT reconstruction problem degenerates into a CT reconstruction problem based on TV regularization terms. Therefore, the CT reconstruction algorithm based on TV regularization terms is a specific implementation provided in this application. In this specific implementation, a convolutional neural network with fixed parameters is used. and Implement operation and operation Among them, neural networks The convolution template is:

[0082] and

[0083] Neural Networks The convolution template is:

[0084] and

[0085] For example, taking one of the m sub-networks (e.g., the j-th sub-network) as an example, and assuming that the j-th sub-network acquires multiple different types of first data, including the first data of the first type, as follows: The first data of the second type is p. n The first data of the third type is q n The first data of the fourth type is u n Among them, the first data of the first type. It can be used to indicate CT images that are in an iterative state.

[0086] Specifically, the CT image reconstruction device can input projection data into the first back-projection network layer in the j-th sub-network. Obtain the first back projection data (e.g., A) T g) , the first data of the first type Input to the orthographic projection network layer CNN in the j-th sub-network A To obtain orthographic projection data (e.g.) ), and the first data of the first type Input to the first convolutional neural network layer in the j-th subnetwork Obtain the first sparse transformation data (e.g.) Afterwards, the CT image reconstruction device can reconstruct the orthographic data (e.g., The input is fed into the second back-projection network layer in the j-th sub-network. Obtain the second back projection data (e.g.) ), transform the first sparse data (e.g. The input is fed into the second convolutional neural network layer in the j-th subnetwork. Obtain the second sparse transform data (e.g.) ).

[0087] Subsequently, the CT image reconstruction device can reconstruct the first backprojection data, the second backprojection data, and the second type of first data p. n The input is fed into the third convolutional neural network layer (CNN) in the j-th sub-network. p To obtain the second type of second data p n+1 And the second sparse transformation data and the first data q of the third type, and n The input is fed into the fourth convolutional neural network layer (CNN) in the j-th sub-network. q To obtain the second data q of the third type n+1 Then, the CT image reconstruction device can convert the second type of second data p n+1 The second data of the third type q n+1 and the first data of the fourth type u n The input is fed into the fifth convolutional neural network layer (CNN) in the j-th sub-network. u To obtain the second data u of the fourth type n+1 Finally, the CT image reconstruction device can convert the first data of the fourth type into... n and the second data of the fourth type u n+1 Input to the sixth convolutional neural network layer in the j-th subnetwork Get the second data of the first type

[0088] Furthermore, in the embodiments of this application, the structure of multiple network layers included in any sub-network (such as the j-th sub-network) can be as follows: Figure 4 As shown. Figure 4 As shown, the number of channels for each variable in the j-th subnetwork can be found in [reference needed]. Figure 4 The initial values ​​of each variable are u0 = 0, p0 = 0, and q0 = 0. The output of the j-th subnetwork This is the reconstructed CT image.

[0089] It should be noted that during the testing of the CT iterative unfolding reconstruction network, the target CT image corresponding to the first object calculated by the CT iterative unfolding reconstruction network is directly output. During the training of the CT iterative unfolding reconstruction network, a loss function is constructed to calculate the difference between the output value (i.e., the predicted value) and the true value of the CT iterative unfolding reconstruction network. This difference is then used as gradient information to feed back to the CT iterative unfolding reconstruction network based on compressed sensing (or based on TV regularization) to update the model parameter values ​​in the CT iterative unfolding reconstruction network. In one example, the loss function can be constructed using point-to-point pixel differences or image structural similarity, such as the L1 norm loss function, L2 norm loss function, or SSIM (structural similarity index) loss function. In another example, a loss function can also be constructed using a convolutional neural network that extracts image structural features, such as the VGG loss function. In this embodiment, the L2 norm loss function is used as the loss function for the entire network model.

[0090] The above embodiments demonstrate that, compared to existing CT iterative unfolding reconstruction networks, the expansion mainly targets the fidelity terms in the CT iterative reconstruction algorithm, directly replacing the regularization terms with neural networks, thus limiting the interpretability of the overall neural network. The technical solution provided in this application, by employing a primal-dual algorithm to solve the CT reconstruction problem based on compressed sensing, can obtain an iterative reconstruction algorithm that is easily unfolded by neural networks. Furthermore, it performs neural network unfolding on the CT iterative reconstruction algorithm, particularly directly unfolding the compressed sensing regularization terms (or total variational TV regularization terms) in the CT iterative reconstruction algorithm. Simultaneously, it also unfolds the fidelity terms in the CT iterative reconstruction algorithm, thus obtaining a complete CT iterative unfolding reconstruction network. Therefore, this method, by performing complete neural network unfolding on the CT iterative reconstruction algorithm, can improve the interpretability of the overall CT reconstruction network. Moreover, by replacing the sparse transformation and its transpose in the compressed sensing regularization terms with shallow neural networks, it can further improve the visual quality of the CT reconstructed images.

[0091] Based on the same technological concept Figure 5 An exemplary embodiment of this application provides a CT image reconstruction apparatus that can execute the flow of a CT image reconstruction method. Exemplarily, the CT image reconstruction apparatus may be a computing device with CT image reconstruction functionality, or other devices with CT image reconstruction functionality, or it may be a functional component (such as a chip) with CT image reconstruction functionality incorporated into a computing device or other device.

[0092] like Figure 5 As shown, the device includes:

[0093] Module 501 is used to acquire the projection data of the first object;

[0094] The processing module 502 is used to input the projection data into the CT iterative unfolding reconstruction network to obtain the target CT image corresponding to the first object; wherein, the CT iterative unfolding reconstruction network is obtained by performing neural network unfolding on the fidelity term and the compressed sensing regularization term or the total variational TV regularization term in the CT iterative reconstruction algorithm.

[0095] In one possible implementation, the projection data is obtained by the CT system scanning the first object according to system parameters, or the projection data is obtained by the computing device simulating the first object according to the system parameters.

[0096] In one possible implementation, the CT iterative unfolding reconstruction network comprises m sub-networks processed sequentially in series, wherein the m sub-networks have the same network structure, and m is an integer greater than 1.

[0097] In one possible implementation, the processing module 502 is specifically used for:

[0098] The projection data is sequentially passed through the m sub-networks to obtain the target CT image.

[0099] In one possible implementation, the first sub-network located at the beginning of the m sub-networks is used as the input end of the CT iterative unfolding reconstruction network, and the m-th sub-network located at the end of the m sub-networks is used as the output end of the CT iterative unfolding reconstruction network. The input data acquired by the first sub-network includes the projection data, and the output data determined by the m-th sub-network includes the target CT image.

[0100] The i-th sub-network among the m sub-networks is used to determine multiple different types of second data based on the projection data and multiple different types of first data output by the (i-1)-th sub-network. The multiple different types of second data and the projection data are used as input data for the (i+1)-th sub-network. The multiple different types of second data correspond one-to-one with the multiple different types of first data. The i-th sub-network is any one of the other sub-networks among the m sub-networks, excluding the first sub-network and the m-th sub-network.

[0101] In one possible implementation, any sub-network includes an orthographic projection network layer, multiple back-projection network layers, and multiple convolutional neural network layers, wherein the multiple back-projection network layers are identical, and the multiple convolutional neural network layers are all different, and the multiple convolutional neural network layers include a first convolutional neural network layer, a second convolutional neural network layer, a third convolutional neural network layer, a fourth convolutional neural network layer, a fifth convolutional neural network layer, and a sixth convolutional neural network layer;

[0102] For the i-th sub-network, the multiple different types of first data acquired by the i-th sub-network include first data of a first type, first data of a second type, first data of a third type, and first data of a fourth type. The first data of the first type is used to indicate the CT image in the iterative state.

[0103] The first back-projection network layer included in the i-th sub-network is used to determine first back-projection data based on the projection data. The orthographic projection network layer included in the i-th sub-network is used to determine orthographic projection data based on first data of the first type. The second back-projection network layer included in the i-th sub-network is used to determine second back-projection data based on the orthographic projection data. The first convolutional neural network layer included in the i-th sub-network is used to determine first sparse transformation data based on first data of the first type. The second convolutional neural network layer included in the i-th sub-network is used to determine second sparse transformation data based on the first sparse transformation data. The first back-projection network layer is any one of the plurality of back-projection network layers, and the second back-projection network layer is any one of the other back-projection network layers besides the first back-projection network layer.

[0104] The third convolutional neural network layer included in the i-th sub-network is used to determine the second data of the second type based on the first back-projection data, the second back-projection data, and the first data of the second type; the fourth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the third type based on the second sparse transform data and the first data of the third type.

[0105] The fifth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the fourth type based on the second data of the second type, the second data of the third type, and the first data of the fourth type;

[0106] The sixth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the first type based on the second data of the fourth type and the first data of the fourth type.

[0107] In one possible implementation, the system parameters include at least one of the following: detector position, detector size and sampling interval, X-ray source position, rotation center position of the CT system, number of projection data and acquisition angle, size of the reconstructed object and sampling interval.

[0108] Based on the same technical concept, embodiments of this application provide a computing device, including:

[0109] Memory, used to store computer programs;

[0110] The processor is used to call the computer program stored in the memory and execute the steps of the CT image reconstruction method according to the obtained program.

[0111] Based on the same technical concept, embodiments of this application provide a computer-readable storage medium storing a computer-executable program for causing a computer to perform the steps of a CT image reconstruction method.

[0112] Based on the same concept, this application also provides a computer program product, which includes a computer program or instructions that, when run on a computer, cause the computer to perform the steps of a CT image reconstruction method.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, or may be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0116] Although this application has been described in conjunction with specific features and embodiments, it is obvious that those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, this application is also intended to include such modifications and variations if they fall within the scope of the claims of this application and their equivalents.

Claims

1. A method for reconstructing computed tomography (CT) images, characterized in that, include: Obtain the projection data of the first object; The projection data is input into the CT iterative unfolding reconstruction network to obtain the target CT image corresponding to the first object; The CT iterative unfolding reconstruction network is obtained by unfolding the fidelity term and compressed sensing regularization term or total variational TV regularization term in the CT iterative reconstruction algorithm into a neural network. The CT iterative unfolding reconstruction network includes m sub-networks processed in series. The network structure of the m sub-networks is the same, and m is an integer greater than 1. The i-th sub-network among the m sub-networks is used to determine multiple different types of second data based on the projection data and multiple different types of first data output by the (i-1)-th sub-network. The multiple different types of second data and the projection data are used as input data for the (i+1)-th sub-network. The multiple different types of second data correspond one-to-one with the multiple different types of first data. The i-th sub-network is any one of the other sub-networks among the m sub-networks except for the first sub-network and the m-th sub-network. Each sub-network includes an orthographic projection network layer, multiple back-projection network layers, and multiple convolutional neural network layers. The multiple back-projection network layers are identical, while the multiple convolutional neural network layers are all different. The multiple convolutional neural network layers include a first convolutional neural network layer, a second convolutional neural network layer, a third convolutional neural network layer, a fourth convolutional neural network layer, a fifth convolutional neural network layer, and a sixth convolutional neural network layer.

2. The method as described in claim 1, characterized in that, The projection data is obtained by the CT system scanning the first object according to system parameters, or the projection data is obtained by the computing device simulating the first object according to the system parameters.

3. The method as described in claim 1, characterized in that, The projection data is input into a CT iterative unfolding reconstruction network to obtain the target CT image corresponding to the first object, including: The projection data is sequentially passed through the m sub-networks to obtain the target CT image.

4. The method as described in claim 3, characterized in that, The first sub-network at the beginning of the m sub-networks is used as the input of the CT iterative unfolding reconstruction network, and the m-th sub-network at the end of the m sub-networks is used as the output of the CT iterative unfolding reconstruction network. The input data acquired by the first sub-network includes the projection data, and the output data determined by the m-th sub-network includes the target CT image.

5. The method as described in claim 4, characterized in that, For the i-th sub-network, the multiple different types of first data acquired by the i-th sub-network include first data of a first type, first data of a second type, first data of a third type, and first data of a fourth type. The first data of the first type is used to indicate the CT image in the iterative state. The first back-projection network layer included in the i-th sub-network is used to determine first back-projection data based on the projection data. The orthographic projection network layer included in the i-th sub-network is used to determine orthographic projection data based on first data of the first type. The second back-projection network layer included in the i-th sub-network is used to determine second back-projection data based on the orthographic projection data. The first convolutional neural network layer included in the i-th sub-network is used to determine first sparse transformation data based on first data of the first type. The second convolutional neural network layer included in the i-th sub-network is used to determine second sparse transformation data based on the first sparse transformation data. The first back-projection network layer is any one of the plurality of back-projection network layers, and the second back-projection network layer is any one of the other back-projection network layers besides the first back-projection network layer. The third convolutional neural network layer included in the i-th sub-network is used to determine the second data of the second type based on the first back-projection data, the second back-projection data, and the first data of the second type; the fourth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the third type based on the second sparse transform data and the first data of the third type. The fifth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the fourth type based on the second data of the second type, the second data of the third type, and the first data of the fourth type; The sixth convolutional neural network layer included in the i-th sub-network is used to determine the second data of the first type based on the second data of the fourth type and the first data of the fourth type.

6. The method according to any one of claims 1-5, characterized in that, The system parameters include at least one of the following: detector position, detector size and sampling interval, X-ray source position, rotation center position of the CT system, number of projection data and acquisition angle, size of the reconstructed object and sampling interval.

7. A CT image reconstruction device, characterized in that, include: The acquisition module is used to acquire the projection data of the first object; A processing module is used to input the projection data into a CT iterative unfolding reconstruction network to obtain a target CT image corresponding to the first object; wherein, the CT iterative unfolding reconstruction network is obtained by performing neural network unfolding on the fidelity term and the compressed sensing regularization term or the total variational TV regularization term in the CT iterative reconstruction algorithm; the CT iterative unfolding reconstruction network includes m sub-networks processed in series, the network structure of the m sub-networks is the same, and m is an integer greater than 1; the i-th sub-network among the m sub-networks is used to determine multiple different types of second data based on the projection data and multiple different types of first data output by the (i-1)-th sub-network, the multiple different types of second data and the projection data are used as input data for the (i+1)-th sub-network, the multiple different types of second data correspond one-to-one with the multiple different types of first data, and the i-th sub-network is any one of the other sub-networks among the m sub-networks except for the first sub-network and the m-th sub-network; Each sub-network includes an orthographic projection network layer, multiple back-projection network layers, and multiple convolutional neural network layers. The multiple back-projection network layers are identical, while the multiple convolutional neural network layers are all different. The multiple convolutional neural network layers include a first convolutional neural network layer, a second convolutional neural network layer, a third convolutional neural network layer, a fourth convolutional neural network layer, a fifth convolutional neural network layer, and a sixth convolutional neural network layer.

8. A computing device, characterized in that, include: Memory, used to store computer programs; A processor is configured to invoke a computer program stored in the memory and execute the method according to any one of claims 1 to 6 in accordance with the obtained program.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer-executable program for causing a computer to perform the method according to any one of claims 1 to 6.

Images