Image reconstruction method and device, computer device and storage medium

Abstract

The application relates to an image reconstruction method and device, computer equipment and a storage medium. The method comprises the following steps: obtaining target sparse raw data; processing the target sparse raw data by using a pre-trained neural network to obtain a target reconstructed image; wherein the neural network comprises k sub-networks, the k sub-networks are connected in sequence, and the processing of the target raw data by using the pre-trained neural network to obtain the target reconstructed image comprises the following steps: the k sub-networks iteratively process input images and the target raw data in sequence and output iterative images; the input image of the first sub-network is a preset image; the input image of the i-th sub-network is the iterative image output by the (i-1)-th sub-network, k is a positive integer, i is any one of 2 to k, and the iterative image output by the k-th sub-network is the target reconstructed image. The method can improve the image reconstruction quality and the image reconstruction efficiency.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image reconstruction method, apparatus, computer device, and storage medium. Background Technology

[0002] With the development of medical scanning equipment, medical images are playing an increasingly important role in clinical diagnosis. However, some medical scanning devices emit radiation, which can affect the health of patients. Therefore, sparse scanning can be used to reduce radiation dose. Currently, after obtaining sparse data through sparse scanning, a pre-established mathematical model is used to reconstruct the image from the sparse data to obtain a medical image. However, current reconstruction methods often introduce a large number of linear artifacts, resulting in low image quality; furthermore, adjusting the mathematical model parameters is difficult, leading to low image reconstruction efficiency. Summary of the Invention

[0003] Therefore, it is necessary to provide an image reconstruction method, apparatus, computer equipment, and storage medium that can improve the image reconstruction quality and reconstruction efficiency in response to the above-mentioned technical problems.

[0004] In a first aspect, this application provides an image reconstruction method, the method comprising:

[0005] Obtain target student data;

[0006] The target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target;

[0007] The neural network comprises k subnetworks, which are connected sequentially. The above-mentioned process of the pre-trained neural network to process the target data and obtain the target reconstructed image includes:

[0008] k sub-networks iterate sequentially based on the input image and the target generated data, and output iterated images. The input image of the first sub-network is a preset image. The input image of the i-th sub-network is the iterated image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k. The iterated image output by the k-th sub-network is the target reconstructed image.

[0009] In one embodiment, the neural network corresponds to an objective function for image reconstruction, which characterizes the determination of a target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction constraint term.

[0010] In one embodiment, the objective function includes:

[0011]

[0012] Where p represents raw data, x represents the reconstructed image, x* represents the target reconstructed image, R represents the Raydon transform, and F... j (x) represents the j-th image reconstruction constraint term, λ j Let be the weight corresponding to the j-th image reconstruction constraint term, n be the number of image reconstruction constraint terms, d be the d-norm of the image reconstruction constraint term, and d be a positive integer.

[0013] In one embodiment, each sub-network includes a gradient descent module and a constraint update module; the above k sub-networks sequentially iterate based on the input image and target data and output an iterated image, including:

[0014] For each sub-network, the input image and target raw data are input into the gradient descent module. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain, and outputs the gradient descent result by combining the step size learned by deep learning.

[0015] The gradient descent result is input into the constraint update module, which performs constraint minimization on the gradient descent result and outputs the iterative image.

[0016] In one embodiment, the constraint update module performs constraint minimization processing on the gradient descent result and outputs an iterative image, including:

[0017] The constraint update module performs forward transformation, thresholding, and backward transformation on the gradient descent results and outputs an iterative image.

[0018] In one embodiment, the training process of the neural network includes:

[0019] Obtain the training sample set; the training sample set includes multiple sparse sampling data and the gold standard corresponding to each sparse sampling data.

[0020] The neural network is trained using multiple sparse sampling data and the corresponding gold standard for each sparse sampling data.

[0021] In one embodiment, the above-mentioned network training using multiple sparse sampled data and the gold standard corresponding to each sparse sampled data to obtain a neural network includes:

[0022] The first sparse sampled data is input into the initial network to reconstruct the image and obtain the training result;

[0023] The loss value between the training result and the gold standard corresponding to the first sparse sampled data is calculated using a preset loss function.

[0024] If the loss value meets the preset convergence condition, the training ends and the neural network is obtained;

[0025] If the loss value does not meet the preset convergence condition, the network parameters in the initial network are adjusted according to the loss value, and the second sparse sampling data is input into the adjusted network to continue training until the calculated loss value meets the preset convergence condition, at which point the training ends and the neural network is obtained.

[0026] In one embodiment, obtaining the training sample set includes:

[0027] Acquire scan data and perform sparse sampling on the scan data to obtain sparse sampled data;

[0028] The gold standard corresponding to the sparse sampled data is obtained by using a preset image reconstruction algorithm to reconstruct the scanned data.

[0029] Secondly, this application also provides an image reconstruction method, the method comprising:

[0030] Obtain target student data;

[0031] The target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target;

[0032] The neural network corresponds to the objective function of image reconstruction. The objective function is used to characterize the determination of the target reconstructed image based on the difference between the target raw data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weight corresponding to the image reconstruction term.

[0033] Thirdly, this application also provides an image reconstruction apparatus, which includes:

[0034] The data acquisition module is used to acquire target biological data;

[0035] The image reconstruction module is used to process the target raw data using a pre-trained neural network to obtain the target reconstructed image;

[0036] The neural network comprises k sub-networks, which are connected sequentially. The image reconstruction module described above is used to iterate the k sub-networks sequentially based on the input image and the target image and output iterated images. The input image of the first sub-network is a preset image; the input image of the i-th sub-network is the iterated image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k; the iterated image output by the k-th sub-network is the target reconstructed image.

[0037] In one embodiment, the neural network corresponds to an objective function for image reconstruction, which characterizes the determination of a target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction constraint term.

[0038] In one embodiment, the objective function includes:

[0039]

[0040] Where p represents raw data, x represents the reconstructed image, x* represents the target reconstructed image, R represents the Raydon transform, and F... j (x) represents the j-th image reconstruction constraint term, λ j Let be the weight corresponding to the j-th image reconstruction constraint term, n be the number of image reconstruction constraint terms, d be the d-norm of the image reconstruction constraint term, and d be a positive integer.

[0041] In one embodiment, each sub-network includes a gradient descent module and a constraint update module; the image reconstruction module includes:

[0042] The gradient descent submodule is used to input the input image and target raw data into the gradient descent module for each subnetwork. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain and outputs the gradient descent result by combining the step size learned by deep learning.

[0043] The constraint update submodule is used to input the gradient descent results into the constraint update module. The constraint update module performs constraint minimization processing on the gradient descent results and outputs the iterative image.

[0044] In one embodiment, the aforementioned constraint update submodule is specifically used by the constraint update module to perform forward transformation processing, threshold processing, and backward transformation processing on the gradient descent result, and output an iterative image.

[0045] In one embodiment, the device further includes:

[0046] The sample acquisition module is used to acquire the training sample set; the training sample set includes multiple sparse sampling data and the gold standard corresponding to each sparse sampling data.

[0047] The training module is used to train the network using multiple sparse sampled data and the gold standard corresponding to each sparse sampled data to obtain the neural network.

[0048] In one embodiment, the training module is specifically used to input the first sparse sampling data into the initial network to perform image reconstruction and obtain the training result; calculate the loss value between the training result and the gold standard corresponding to the first sparse sampling data using a preset loss function; if the loss value meets the preset convergence condition, the training ends and the neural network is obtained; if the loss value does not meet the preset convergence condition, the network parameters in the initial network are adjusted according to the loss value, and the second sparse sampling data is input into the adjusted network to continue training until the calculated loss value meets the preset convergence condition and the training ends and the neural network is obtained.

[0049] In one embodiment, the sample acquisition module is specifically used to acquire scan data and perform sparse sampling on the scan data to obtain sparse sampled data; and use a preset image reconstruction algorithm to perform image reconstruction on the scan data to obtain the gold standard corresponding to the sparse sampled data.

[0050] Fourthly, this application also provides an image reconstruction apparatus, which includes:

[0051] The data acquisition module is used to acquire target biological data;

[0052] The image reconstruction module is used to process the target raw data using a pre-trained neural network to obtain the target reconstructed image;

[0053] The neural network corresponds to the objective function of image reconstruction. The objective function is used to characterize the difference between the target generated data and the reconstructed image obtained in the iteration, as well as the product of the weights corresponding to at least one image reconstruction constraint term and the image reconstruction term, to determine the target reconstructed image.

[0054] Fifthly, this application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0055] Obtain target student data;

[0056] The target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target;

[0057] The neural network comprises k subnetworks, which are connected sequentially. A pre-trained neural network processes the target data to obtain a reconstructed image of the target, including:

[0058] k sub-networks iterate sequentially based on the input image and the target generated data, and output iterated images. The input image of the first sub-network is a preset image. The input image of the i-th sub-network is the iterated image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k. The iterated image output by the k-th sub-network is the target reconstructed image.

[0059] Sixthly, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0060] Obtain target student data;

[0061] The target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target;

[0062] The neural network comprises k subnetworks, which are connected sequentially. A pre-trained neural network processes the target data to obtain a reconstructed image of the target, including:

[0063] k sub-networks iterate sequentially based on the input image and the target generated data, and output iterated images. The input image of the first sub-network is a preset image. The input image of the i-th sub-network is the iterated image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k. The iterated image output by the k-th sub-network is the target reconstructed image.

[0064] The aforementioned image reconstruction method, apparatus, computer equipment, and storage medium acquire target raw data; process the target raw data using a pre-trained neural network to obtain a reconstructed image of the target; wherein the neural network includes k sub-networks, which are sequentially connected, and the k sub-networks iterate sequentially based on the input image and the target raw data, and output an iterative image. This embodiment of the present disclosure uses target raw data for image reconstruction, which requires less computation compared to full-scan raw data, thus improving image reconstruction speed; furthermore, reconstruction using a neural network composed of k sub-networks avoids linear artifacts, thus further improving both image reconstruction efficiency and image reconstruction quality. Attached Figure Description

[0065] Figure 1 This is a diagram illustrating the application environment of an image reconstruction method in one embodiment;

[0066] Figure 2 This is a flowchart illustrating an image reconstruction method in one embodiment;

[0067] Figure 3a This is one of the schematic diagrams of a neural network structure in one embodiment;

[0068] Figure 3b This is the second schematic diagram of the neural network structure in one embodiment;

[0069] Figure 4 This is a flowchart illustrating the steps of k sub-networks sequentially iterating based on the input image and target data and outputting an iterated image in one embodiment.

[0070] Figure 5 This is a schematic diagram of the constraint update module in one embodiment;

[0071] Figure 6 This is a flowchart illustrating the training process of a neural network in one embodiment;

[0072] Figure 7 This is a flowchart illustrating the image reconstruction method in another embodiment;

[0073] Figure 8 This is a structural block diagram of an image reconstruction apparatus in one embodiment;

[0074] Figure 9 This is a structural block diagram of an image reconstruction apparatus in one embodiment;

[0075] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0077] The image reconstruction method provided in this application can be applied to, for example... Figure 1 The application environment shown may include a terminal 101 and a medical scanning device 102. The terminal 101 can communicate with the medical scanning device 102 via a network. The terminal 101 may be, but is not limited to, various personal computers, laptops, and tablets, and the medical scanning device 102 may be, but is not limited to, a CT (Computed Tomography) device, a PET (Positron Emission Computed Tomography)-CT device, and an MR (Magnetic Resonance) device.

[0078] Taking the medical scanning device 102 as a CT device as an example, the CT device can be any of the parallel beam, fan beam, and cone beam, and the CT scanning mode includes, but is not limited to, axial scanning and spiral scanning.

[0079] This application environment may also include a PACS (Picture Archiving and Communication Systems) server 103, with terminals 101 and medical scanning devices 102 all able to communicate with the PACS server 103 via a network. The aforementioned PACS server 103 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0080] In one embodiment, such as Figure 2 As shown, an image reconstruction method is provided, which is applied to... Figure 1 Taking the terminal in the example, the explanation includes the following steps:

[0081] Step 201: Obtain target student data.

[0082] The target raw data can be sparse raw data obtained by sampling the raw data of a full scan at a preset sampling rate, or it can be sparse raw data obtained by sparse scanning with a medical scanning device. Since sparse scanning can reduce radiation dose, it is preferable to obtain sparse raw data by sparse scanning with a medical scanning device. This disclosure does not limit the sparsity of the sparse raw data.

[0083] The terminal can acquire full-scan biodata from a medical scanning device or a PACS server, and then sample the full-scan biodata according to a preset sampling rate to obtain target biodata; alternatively, it can directly acquire target biodata obtained by sparse scanning of a medical scanning device or a PACS server. This disclosure does not limit the data acquisition method.

[0084] The preset sampling rate can be 5%. For example, 2400 samples per lap can yield 120 sparse data points. This embodiment does not limit the preset sampling rate.

[0085] Step 202: Process the target raw data using a pre-trained neural network to obtain the target reconstructed image.

[0086] like Figure 3a As shown, the neural network includes k sub-networks, which are connected sequentially. The process of using a pre-trained neural network to process the target raw data and obtain the target reconstructed image may include: the k sub-networks iterating sequentially based on the input image and the target raw data and outputting an iterated image.

[0087] The input image of the first sub-network is a preset image; the input image of the i-th sub-network is the iterative image output by the (i-1)-th sub-network, k is a positive integer, and i is any one from 2 to k; the iterative image output by the k-th sub-network is the target reconstructed image.

[0088] like Figure 3a As shown, the terminal inputs a preset image into the first sub-network. The first sub-network iterates based on the preset image and the target generated data, and outputs the first iterated image. Then, the first iterated image is input into the second sub-network, which iterates based on the first iterated image and the target generated data, and outputs the second iterated image. This process continues, with the (i-1)th iterated image input into the ith sub-network, which iterates based on the (i-1)th iterated image and the target generated data, and outputs the ith iterated image. Finally, the (k-1)th iterated image is input into the kth sub-network, which iterates based on the (k-1)th iterated image and the target generated data, and outputs the kth iterated image, which is then identified as the target reconstructed image.

[0089] The preset image can be any one of an all-zero image, an all-one image, and an FBP image. This embodiment of the disclosure does not limit the preset image.

[0090] In one embodiment, the neural network described above is a deep unfolded neural network, and the structure of the neural network is not limited in this embodiment.

[0091] In the above image reconstruction method, target raw data is acquired; the target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target; wherein the neural network includes k sub-networks, which are connected sequentially, and the k sub-networks iterate sequentially based on the input image and the target raw data and output an iterative image. This embodiment of the present disclosure uses target raw data for image reconstruction, which requires less computation compared to full-scan raw data, thus improving the image reconstruction speed; furthermore, the reconstruction using a neural network composed of k sub-networks avoids linear artifacts, thus not only further improving image reconstruction efficiency but also improving image reconstruction quality.

[0092] In one embodiment, the terminal is equipped with a pre-trained neural network, which corresponds to an image reconstruction objective function. The objective function characterizes the determination of the target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction term. The image reconstruction constraint term may include at least one of gradient constraint terms and optical difference constraint terms. This disclosure does not limit the image reconstruction constraint term.

[0093] After the terminal obtains the target data, it inputs the target data into the neural network corresponding to the target function of image reconstruction. The neural network performs image reconstruction based on the target data and outputs the target reconstructed image.

[0094] Understandably, the objective function is related to the target raw data, image reconstruction constraints, and the reconstructed image; therefore, establishing the objective function is relatively easy. Furthermore, the process of image reconstruction using the target raw data by the neural network is essentially the process of solving the objective function. Since the neural network is a pre-trained model, solving it is also relatively easy and fast.

[0095] In the above embodiments, the establishment and solution of the objective function are relatively easy. Using the neural network corresponding to the objective function for reconstruction can avoid the generation of linear artifacts. Therefore, it can not only further improve the efficiency of image reconstruction, but also improve the quality of image reconstruction.

[0096] In one embodiment, the objective function for the above image reconstruction includes:

[0097]

[0098] Where p is the target generated data, x is the reconstructed image, x* is the target reconstructed image, R is the Raydon transform, and F... j (x) represents the j-th image reconstruction constraint term, λ j Let be the weight corresponding to the j-th image reconstruction constraint term, n be the number of image reconstruction constraint terms, d be the d-norm of the image reconstruction constraint term, and d be a positive integer.

[0099] For the objective function described above, iterative gradient descent can theoretically be used to update x:

[0100]

[0101] Where, τ k Let x be the step size. k Let x be the reconstructed image after k iterations. k+1 This is the reconstructed image after the (k+1)th iteration.

[0102] In practice, due to If the solution is difficult, a proximal gradient descent method can be used. The iterative process can include the following steps:

[0103] Step 1:

[0104] Step 2:

[0105] The neural network is built according to the above two steps, and the objective function can be easily solved through the neural network.

[0106] In the above embodiments, k sub-networks in the neural network iterate sequentially based on the input image and target data, and output iterative images. The iterative image output by the kth sub-network is the reconstructed target image. Through this embodiment, a solution to the objective function can be obtained using a finite number of sub-networks for iteration. Compared with other iterative methods, this method requires fewer iterations and has a faster solution speed.

[0107] In one embodiment, such as Figure 3b As shown, each sub-network includes a gradient descent block and an optimization block; as... Figure 4 As shown, the process by which the above k sub-networks iterate sequentially based on the input image and target data and output an iterated image can include the following steps:

[0108] Step 301: For each sub-network, the input image and target raw data are input into the gradient descent module. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain, and outputs the gradient descent result by combining the step size learned by deep learning.

[0109] Taking the first sub-network as an example, the preset image and target raw data are input into the gradient descent module of the first sub-network. This gradient descent module forward projects the preset image to obtain the first raw data and performs deep learning on the difference between the first raw data and the target raw data. If the difference between the first raw data and the target raw data is greater than a preset difference value, the difference is transformed from the Raeden space to the image domain, and the difference in the image domain is multiplied by the step size learned by deep learning. Then, the preset image is added to the above product, and deep learning is performed on the data obtained after addition and the target raw data until the learned difference is less than the preset difference value, at which point the gradient descent result is output.

[0110] The gradient descent module can employ any network architecture from UNet, RsNet, and CNN (Convolutional Neural Networks). The gradient descent result can be a scalar, such as the stride; or it can be a matrix of the same size as the image. This embodiment of the disclosure does not limit the structure of the gradient descent module or the gradient descent result.

[0111] Understandably, the structures of the subnetworks can be the same or different.

[0112] Step 302: Input the gradient descent result into the constraint update module. The constraint update module performs constraint minimization processing on the gradient descent result and outputs the iterative image.

[0113] like Figure 5 As shown, the constraint update module may include a forward transformation processing submodule, a thresholding processing submodule, and a backward transformation processing submodule connected in sequence. The process of the constraint update module performing constraint minimization processing on the gradient descent result may include: the constraint update module performing forward transformation processing, thresholding processing, and backward transformation processing on the gradient descent result, and outputting an iterative image.

[0114] The aforementioned forward transformation and backward transformation submodules can employ convolutional neural networks of arbitrary structures to implement linear or nonlinear transformations. Furthermore, the forward and backward transformations can be identity transformations; for example, if the forward transformation transforms the gradient descent result from the time domain to the frequency domain, then the backward transformation transforms the thresholding result from the frequency domain to the time domain. The thresholding process can employ functions such as soft thresholding or the hard tanh function. This disclosure does not limit the forward transformation, backward transformation, and thresholding processes; they can be configured according to actual conditions.

[0115] In the above embodiments, for each sub-network, the input image and target raw data are input into the gradient descent module. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain, and outputs the gradient descent result by combining the step size learned by deep learning. The gradient descent result is input into the constraint update module, which performs constraint minimization processing on the gradient descent result and outputs the iterative image. In this embodiment, each sub-network uses the gradient descent module for deep learning and the constraint update module for constraint minimization processing. Since the modules in the sub-network are pre-defined, that is, the role of the sub-network is fixed, it can control the implementation effect of each step of the iterative reconstruction, so the interpretability of the neural network is strong. At the same time, errors can be corrected by comparing the output result after each module with the expectation. Furthermore, the sub-network with fixed role is not sensitive to data, which can make the image reconstruction effect more stable.

[0116] In one embodiment, such as Figure 6 As shown, the training process of a neural network can include the following steps:

[0117] Step 401: Obtain the training sample set.

[0118] The training sample set includes multiple sparse sampling data and the gold standard corresponding to each sparse sampling data.

[0119] Multiple sparse sampling data and the corresponding gold standard for each sparse sampling data are obtained separately, and a training sample set is composed of the multiple sparse sampling data and the corresponding gold standard for each sparse sampling data.

[0120] In one embodiment, the process of obtaining sparse sampling data and a gold standard may include: obtaining scan data and performing sparse sampling on the scan data to obtain sparse sampling data; and using a preset image reconstruction algorithm to perform image reconstruction on the scan data to obtain the gold standard corresponding to the sparse sampling data.

[0121] The aforementioned preset image reconstruction algorithm may include the FBP algorithm or other CT image reconstruction algorithms. This disclosure does not limit the preset image reconstruction algorithm.

[0122] Step 402: Train the network using multiple sparse sampled data and the gold standard corresponding to each sparse sampled data to obtain the neural network.

[0123] After obtaining the training sample set, the network can be trained as follows: input the first sparse sampling data into the initial network to reconstruct the image and obtain the training result; calculate the loss value between the training result and the gold standard corresponding to the first sparse sampling data using a preset loss function; if the loss value meets the preset convergence condition, the training ends and the neural network is obtained; if the loss value does not meet the preset convergence condition, adjust the network parameters in the initial network according to the loss value, and input the second sparse sampling data into the adjusted network to continue training until the calculated loss value meets the preset convergence condition, at which point the training ends and the neural network is obtained.

[0124] In one embodiment, the preset loss function may include two parts: one part calculates the loss value between the training result and the gold standard, and the other part calculates the loss value between the input image and the output iterative image of each sub-network. This disclosure does not limit the preset loss function.

[0125] In the above embodiments, a training sample set is obtained; the network is trained using multiple sparse sampling data and the gold standard corresponding to each sparse sampling data to obtain a neural network. Through the embodiments of this disclosure, the neural network is pre-trained so that the target image can be obtained by solving the objective function of the reconstructed image using the neural network, thereby improving the image reconstruction quality and image reconstruction efficiency.

[0126] In one embodiment, such as Figure 7 As shown, an image reconstruction method is provided, which is applied to... Figure 1 Taking the terminal in the example, the explanation includes the following steps:

[0127] Step 501: Obtain target student data.

[0128] The target raw data can be sparse raw data obtained by sampling the raw data of a full scan at a preset sampling rate, or it can be sparse raw data obtained by sparse scanning with a medical scanning device. Since sparse scanning can reduce radiation dose, it is preferable to obtain sparse raw data by sparse scanning with a medical scanning device. This disclosure does not limit the sparsity of the sparse raw data.

[0129] The terminal can acquire full-scan biodata from a medical scanning device or a PACS server, and then sample the full-scan biodata according to a preset sampling rate to obtain target biodata; alternatively, it can directly acquire target biodata obtained by sparse scanning of a medical scanning device or a PACS server. This disclosure does not limit the data acquisition method.

[0130] The preset sampling rate can be 5%. For example, 2400 samples per lap can yield 120 sparse data points. This embodiment does not limit the preset sampling rate.

[0131] Step 502: Process the target raw data using a pre-trained neural network to obtain the target reconstructed image.

[0132] The terminal is equipped with a pre-trained neural network, which corresponds to an image reconstruction objective function. The objective function characterizes the determination of the target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction term. The image reconstruction constraint term may include at least one of gradient constraint terms and optical difference constraint terms. This disclosure does not limit the image reconstruction constraint term.

[0133] After the terminal obtains the target data, it inputs the target data into the neural network corresponding to the target function of image reconstruction. The neural network performs image reconstruction based on the target data and outputs the target reconstructed image.

[0134] Understandably, the objective function is related to the target raw data, image reconstruction constraints, and the reconstructed image; therefore, establishing the objective function is relatively easy. Furthermore, the process of image reconstruction using the target raw data by the neural network is essentially the process of solving the objective function. Since the neural network is a pre-trained model, solving it is also relatively easy and fast.

[0135] In the above image reconstruction method, target raw data is acquired; a pre-trained neural network is used to process the sparse target raw data to obtain the target reconstructed image. This embodiment uses target raw data for image reconstruction, which requires less computation compared to full-scan raw data, thus improving image reconstruction speed. Furthermore, the establishment and solution of the objective function are relatively easy, and using the neural network corresponding to the objective function for reconstruction can avoid linear artifacts. Therefore, it can not only further improve image reconstruction efficiency but also improve image reconstruction quality.

[0136] It should be understood that, although Figures 2 to 7 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figures 2 to 7 At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0137] In one embodiment, such as Figure 8 As shown, an image reconstruction apparatus is provided, comprising:

[0138] Data acquisition module 601 is used to acquire target biological data;

[0139] Image reconstruction module 602 is used to process target raw data using a pre-trained neural network to obtain a reconstructed image of the target;

[0140] The neural network comprises k sub-networks, which are connected sequentially. The image reconstruction module 502 is used to iterate the k sub-networks sequentially based on the input image and the target image and output iterated images. The input image of the first sub-network is a preset image. The input image of the i-th sub-network is the iterated image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k. The iterated image output by the k-th sub-network is the target reconstructed image.

[0141] In one embodiment, the neural network corresponds to an objective function for image reconstruction, which characterizes the determination of a target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction constraint term.

[0142] In one embodiment, the objective function includes:

[0143]

[0144] Where p is the target generated data, x is the reconstructed image, x* is the target reconstructed image, R is the Raydon transform, and F... j (x) represents the j-th image reconstruction constraint term, λ j Let be the weight corresponding to the j-th image reconstruction constraint term, n be the number of image reconstruction constraint terms, d be the d-norm of the image reconstruction constraint term, and d be a positive integer.

[0145] In one embodiment, each sub-network includes a gradient descent module and a constraint update module; the image reconstruction module 602 includes:

[0146] The gradient descent submodule is used to input the input image and target raw data into the gradient descent module for each subnetwork. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain and outputs the gradient descent result by combining the step size learned by deep learning.

[0147] The constraint update submodule is used to input the gradient descent results into the constraint update module. The constraint update module performs constraint minimization processing on the gradient descent results and outputs the iterative image.

[0148] In one embodiment, the aforementioned constraint update submodule is specifically used by the constraint update module to perform forward transformation processing, threshold processing, and backward transformation processing on the gradient descent result, and output an iterative image.

[0149] In one embodiment, the device further includes:

[0150] The sample acquisition module is used to acquire the training sample set; the training sample set includes multiple sparse sampling data and the gold standard corresponding to each sparse sampling data.

[0151] The training module is used to train the network using multiple sparse sampled data and the gold standard corresponding to each sparse sampled data to obtain the neural network.

[0152] In one embodiment, the training module is specifically used to input the first sparse sampling data into the initial network to perform image reconstruction and obtain the training result; calculate the loss value between the training result and the gold standard corresponding to the first sparse sampling data using a preset loss function; if the loss value meets the preset convergence condition, the training ends and the neural network is obtained; if the loss value does not meet the preset convergence condition, the network parameters in the initial network are adjusted according to the loss value, and the second sparse sampling data is input into the adjusted network to continue training until the calculated loss value meets the preset convergence condition and the training ends and the neural network is obtained.

[0153] In one embodiment, the sample acquisition module is specifically used to acquire scan data and perform sparse sampling on the scan data to obtain sparse sampled data; and use a preset image reconstruction algorithm to perform image reconstruction on the scan data to obtain the gold standard corresponding to the sparse sampled data.

[0154] In one embodiment, such as Figure 9 As shown, an image reconstruction apparatus is provided, the apparatus comprising:

[0155] Data acquisition module 701 is used to acquire target biological data;

[0156] Image reconstruction module 702 is used to process target raw data using a pre-trained neural network to obtain a reconstructed image of the target;

[0157] The neural network corresponds to the objective function of image reconstruction. The objective function is used to characterize the difference between the target generated data and the reconstructed image obtained in the iteration, as well as the product of the weights corresponding to at least one image reconstruction constraint term and the image reconstruction term, to determine the target reconstructed image.

[0158] For specific limitations regarding the image reconstruction apparatus, please refer to the limitations on the image reconstruction method above, which will not be repeated here. Each module in the aforementioned image reconstruction apparatus can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0159] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an image reconstruction method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0160] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0161] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0162] Obtain target student data;

[0163] The target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target;

[0164] The neural network comprises k sub-networks, which are connected sequentially. Each sub-network iterates based on the input image and the target data and outputs an iterative image. The input image of the first sub-network is a preset image. The input image of the i-th sub-network is the iterative image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k. The iterative image output by the k-th sub-network is the target reconstructed image.

[0165] In one embodiment, the neural network corresponds to an objective function for image reconstruction, which characterizes the determination of a target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction constraint term.

[0166] In one embodiment, the objective function includes:

[0167]

[0168] Where p is the target generated data, x is the reconstructed image, x* is the target reconstructed image, R is the Raydon transform, and F... j (x) represents the j-th image reconstruction constraint term, λ j Let be the weight corresponding to the j-th image reconstruction constraint term, n be the number of image reconstruction constraint terms, d be the d-norm of the image reconstruction constraint term, and d be a positive integer.

[0169] In one embodiment, each sub-network includes a gradient descent module and a constraint update module; when the processor executes the computer program, it also implements the following steps:

[0170] For each sub-network, the input image and target raw data are input into the gradient descent module. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain, and outputs the gradient descent result by combining the step size learned by deep learning.

[0171] The gradient descent result is input into the constraint update module, which performs constraint minimization on the gradient descent result and outputs the iterative image.

[0172] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0173] The constraint update module performs forward transformation, thresholding, and backward transformation on the gradient descent results and outputs an iterative image.

[0174] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0175] Obtain the training sample set; the training sample set includes multiple sparse sampling data and the gold standard corresponding to each sparse sampling data.

[0176] The neural network is trained using multiple sparse sampling data and the corresponding gold standard for each sparse sampling data.

[0177] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0178] The first sparse sampled data is input into the initial network to reconstruct the image and obtain the training result;

[0179] The loss value between the training result and the gold standard corresponding to the first sparse sampled data is calculated using a preset loss function.

[0180] If the loss value meets the preset convergence condition, the training ends and the neural network is obtained;

[0181] If the loss value does not meet the preset convergence condition, the network parameters in the initial network are adjusted according to the loss value, and the second sparse sampling data is input into the adjusted network to continue training until the calculated loss value meets the preset convergence condition, at which point the training ends and the neural network is obtained.

[0182] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0183] Acquire scan data and perform sparse sampling on the scan data to obtain sparse sampled data;

[0184] The gold standard corresponding to the sparse sampled data is obtained by using a preset image reconstruction algorithm to reconstruct the scanned data.

[0185] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0186] Obtain target student data;

[0187] The target raw data is processed using a pre-trained neural network to obtain a reconstructed image of the target;

[0188] The neural network comprises k sub-networks, which are connected sequentially. Each sub-network iterates based on the input image and the target data and outputs an iterative image. The input image of the first sub-network is a preset image. The input image of the i-th sub-network is the iterative image output by the (i-1)-th sub-network, where k is a positive integer and i is any one from 2 to k. The iterative image output by the k-th sub-network is the target reconstructed image.

[0189] In one embodiment, the neural network corresponds to an objective function for image reconstruction, which characterizes the determination of a target reconstructed image based on the difference between the target generated data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weights corresponding to the image reconstruction constraint term.

[0190] In one embodiment, the objective function includes:

[0191]

[0192] Where p is the target generated data, x is the reconstructed image, x* is the target reconstructed image, R is the Raydon transform, and F... j (x) represents the j-th image reconstruction constraint term, λ j Let be the weight corresponding to the j-th image reconstruction constraint term, n be the number of image reconstruction constraint terms, d be the d-norm of the image reconstruction constraint term, and d be a positive integer.

[0193] In one embodiment, each sub-network includes a gradient descent module and a constraint update module; when the computer program is executed by a processor, it also implements the following steps:

[0194] For each sub-network, the input image and target raw data are input into the gradient descent module. The gradient descent module performs deep learning on the difference between the raw data obtained by forward projection of the input image and the target raw data, transforms the difference in the Raeden space back to the image domain, and outputs the gradient descent result by combining the step size learned by deep learning.

[0195] The gradient descent result is input into the constraint update module, which performs constraint minimization on the gradient descent result and outputs the iterative image.

[0196] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0197] The constraint update module performs forward transformation, thresholding, and backward transformation on the gradient descent results and outputs an iterative image.

[0198] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0199] Obtain the training sample set; the training sample set includes multiple sparse sampling data and the gold standard corresponding to each sparse sampling data.

[0200] The neural network is trained using multiple sparse sampling data and the corresponding gold standard for each sparse sampling data.

[0201] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0202] The first sparse sampled data is input into the initial network to reconstruct the image and obtain the training result;

[0203] The loss value between the training result and the gold standard corresponding to the first sparse sampled data is calculated using a preset loss function.

[0204] If the loss value meets the preset convergence condition, the training ends and the neural network is obtained;

[0205] If the loss value does not meet the preset convergence condition, the network parameters in the initial network are adjusted according to the loss value, and the second sparse sampling data is input into the adjusted network to continue training until the calculated loss value meets the preset convergence condition, at which point the training ends and the neural network is obtained.

[0206] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0207] Acquire scan data and perform sparse sampling on the scan data to obtain sparse sampled data;

[0208] The gold standard corresponding to the sparse sampled data is obtained by using a preset image reconstruction algorithm to reconstruct the scanned data.

[0209] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. 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.

[0210] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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, they should be considered to be within the scope of this specification.

[0211] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method of image reconstruction, characterized by, The method comprises: acquiring target raw data; wherein the target raw data comprises sparse raw data obtained by sampling full-scan raw data at a preset sampling rate, or sparse raw data obtained by sparse scanning by a medical scanning device; processing the target raw data by using a pre-trained neural network to obtain a target reconstructed image; wherein the neural network comprises k sub-networks, the k sub-networks are connected in sequence, and the processing of the target raw data by using the pre-trained neural network to obtain the target reconstructed image comprises: the k sub-networks iteratively process input images and the target raw data in sequence and output iterative images; the input image of the first sub-network is a preset image; the input image of the i-th sub-network is the iterative image output by the (i-1)-th sub-network, k is a positive integer, and i is any one of 2 to k; the iterative image output by the k-th sub-network is the target reconstructed image; the k sub-networks iteratively process input images and the target raw data in sequence and output iterative images, comprising: inputting the preset image into the first sub-network, and the first sub-network iteratively processes the preset image and the target raw data and outputs a first iterative image; inputting the (i-1)-th iterative image into the i-th sub-network, and the i-th sub-network iteratively processes the (i-1)-th iterative image and the target raw data and outputs an i-th iterative image.

2. The method of claim 1, wherein, The neural network corresponds to an objective function of image reconstruction, and the objective function is used to represent the difference between the target raw data and the reconstructed image obtained in the iteration, and the product of at least one image reconstruction constraint term and the weight corresponding to the image reconstruction constraint term, to determine the target reconstructed image.

3. The method of claim 2, wherein, The objective function comprises: wherein p is the target sinogram, x is the reconstructed image, x* is the target reconstructed image, R is the Radon transform, F j (x) is the jth image reconstruction constraint term, λ j is the weight corresponding to the jth image reconstruction constraint term, n is the number of image reconstruction constraint terms, d is the d-norm of the image reconstruction constraint term, and d is a positive integer.

4. The method of claim 1, wherein, Each of the sub-networks comprises a gradient descent module and a constraint update module; the k sub-networks iteratively process input images and the target raw data in sequence and output iterative images, comprising: for each of the sub-networks, inputting the input image and the target raw data into the gradient descent module, the gradient descent module performing deep learning on the difference between the raw data obtained by the positive projection of the input image and the target raw data, transforming the difference in the Ledge space back to the image domain, and outputting a gradient descent result in combination with the deep learned step length; inputting the gradient descent result into the constraint update module, and the constraint update module performing constraint minimization processing on the gradient descent result and outputting the iterative image.

5. The method of claim 4, wherein, The constraint update module performs constraint minimization processing on the gradient descent result and outputs the iterative image, comprising: The constraint update module performs forward transformation processing, threshold processing, and backward transformation processing on the gradient descent result and outputs the iterative image.

6. The method of claim 1, wherein, The training process of the neural network comprises: acquiring a training sample set; the training sample set comprises a plurality of sparse sampling data and a gold standard corresponding to each of the sparse sampling data; training the network by using the plurality of sparse sampling data and the gold standard corresponding to each of the sparse sampling data to obtain the neural network.

7. The method of claim 6, wherein, The network training is performed by using the plurality of sparse sampling data and the corresponding gold standard, and the neural network is obtained, and the network training comprises: inputting the first sparse sampling data into an initial network to perform image reconstruction to obtain a training result; calculating a loss value between the training result and the gold standard corresponding to the first sparse sampling data by using a preset loss function; if the loss value meets a preset convergence condition, the training is ended to obtain the neural network; if the loss value does not meet the preset convergence condition, the network parameters in the initial network are adjusted according to the loss value, the second sparse sampling data is input into the adjusted network to continue the training, and the training is ended to obtain the neural network when the calculated loss value meets the preset convergence condition.

8. The method of claim 6, wherein, The training sample set is obtained, and the method comprises: obtaining scanning data, and performing sparse sampling on the scanning data to obtain the sparse sampling data; performing image reconstruction on the scanning data by using a preset image reconstruction algorithm to obtain the gold standard corresponding to the sparse sampling data.

9. An image reconstruction method, characterized by, The method comprises: obtaining target raw data; wherein the target raw data comprises sparse raw data obtained by sampling full scanning raw data according to a preset sampling rate, or sparse raw data obtained by sparse scanning of a medical scanning device; processing the target raw data by using a pre-trained neural network to obtain a target reconstruction image; wherein the neural network corresponds to a target function of image reconstruction, and the target function is used to represent a product of a difference between the target raw data and a reconstruction image obtained in iteration, and a weight corresponding to at least one image reconstruction constraint term, to determine the target reconstruction image; processing the target raw data by using the pre-trained neural network to obtain the target reconstruction image comprises: inputting a preset image into a first sub-network, and the first sub-network iterates according to the preset image and the target raw data, and outputs a first iteration image; inputting an i-1th iteration image into an i-th sub-network, and the i-th sub-network iterates according to the i-1th iteration image and the target raw data, and outputs an i-th iteration image; wherein k is a positive integer, and i is any one of 2 to k.

10. An image reconstruction apparatus, characterized by comprising: The device comprises: a data acquisition module configured to obtain target raw data; wherein the target raw data comprises sparse raw data obtained by sampling full scanning raw data according to a preset sampling rate, or sparse raw data obtained by sparse scanning of a medical scanning device; an image reconstruction module configured to process the target raw data by using a pre-trained neural network to obtain a target reconstruction image. The neural network comprises k sub-networks, the k sub-networks are connected in sequence, and the image reconstruction module is configured to perform iteration on the k sub-networks in sequence according to an input image and the target data and output an iterative image; an input image of a first sub-network is a preset image; an input image of an i-th sub-network is an iterative image output by an (i-1)-th sub-network, k is a positive integer, and i is any one of 2 to k; and the iterative image output by the k-th sub-network is the target reconstructed image. The image reconstruction module is specifically configured to input the preset image into the first sub-network, perform iteration on the first sub-network according to the preset image and the target data, and output a first iterative image; input an (i-1)-th iterative image into an i-th sub-network, perform iteration on the i-th sub-network according to the (i-1)-th iterative image and the target data, and output an i-th iterative image. 11.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-10. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 8.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 8.

Images