Sparse spiral CT image reconstruction method based on micro-spiral reconstruction operator

Abstract

A sparse spiral CT image reconstruction method based on a micro-spiral reconstruction operator comprises the steps that firstly, real spiral scanning geometric parameters of an object and corresponding full-angle spiral projection data are obtained, and a final reconstruction image is obtained through seven steps; according to the method, front projection is performed on the reconstructed sparse angle image by using real scanning geometry, and geometric prior guidance is provided for missing projection; meanwhile, according to similarity and redundancy characteristics of spiral scanning adjacent projections, a projection completion network is constructed, and middle missing projection data is jointly synthesized by learning a bidirectional motion field of adjacent angles and combining geometric prior projection; the global strip artifacts of the image are recovered; finally, joint training of the projection domain and the image domain is achieved, overall recovery can be carried out by combining projection-image double-domain information, recovery is carried out by adopting data collected in two screw pitches, and the overlarge calculation amount is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and computer tomography, and in particular to a sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator. Background Art

[0002] As a high-contrast, non-invasive, fast imaging technology, computed tomography (CT) has been widely used in medical imaging diagnosis, surgical positioning navigation, radiotherapy planning and other fields. However, studies have shown that when exposed to a large amount of X-ray radiation, systemic acute and chronic radiation damage, DNA and cell damage, etc. will occur, greatly increasing the risk of future cancer. However, reducing the X-ray radiation dose will inevitably lead to the degradation of CT image quality and affect the diagnostic and therapeutic performance. Therefore, achieving high-quality imaging of low-dose CT and ensuring that clinical diagnosis and treatment needs are met are cutting-edge scientific issues in the field of high-end medical CT research. In order to alleviate the radiation dose hazards of medical CT to the human body, a variety of low-dose CT scanning protocols have been proposed. At present, low milliampere-second scanning protocols combined with advanced reconstruction algorithms can realize conventional low-dose clinical task scanning, but such methods are limited by hardware. When the X-ray irradiation is further reduced, the measured projection data may experience a "photon starvation" effect, that is, the amount of X-rays generated by ultra-low milliampere-seconds can no longer penetrate the object or the amount of X-rays that pass through is extremely low, resulting in the detector being unable to collect effective and sufficient X-ray signals, making the reconstructed image tissue structure overwhelmed by noise artifacts and difficult to recover.

[0003] In terms of achieving ultra-low-dose CT imaging, the sparse angle scanning protocol is one of the most promising technologies. It reduces the radiation dose by reducing the projection exposure angle, and can keep the tube voltage and tube current of each projection exposure at the normal dose level. Compared with the low milliampere-second scanning protocol, this protocol has the following advantages: 1. It can effectively avoid the photon starvation effect and avoid modeling complex noise; 2. It can avoid the problem of model performance degradation caused by the domain difference between simulated low-dose data and real low-dose data; 3. It can effectively shorten the scanning time. However, the existing sparse angle scanning still cannot meet the current clinical diagnosis and treatment needs. When the scanning projection angle is reduced, the sampling rate does not meet the classic Shannon sampling theorem, which makes the image reconstruction degenerate into an ill-posed inverse problem. When using classical filtered back projection for reconstruction, it is easy to introduce artifacts in the image.

[0004] With the rapid development of deep learning, a large number of deep learning-based methods have been proposed to effectively improve image quality. According to the different methods and modeling objects, these methods can be divided into four categories: (1) Sparse angle CT imaging based on projection interpolation, which relies on the correlation between projection angles, learns the mapping between sparse angle projections and full angle projections, and then uses the filtered back projection algorithm to reconstruct the target image. However, due to the characteristics of the FBP reconstruction algorithm, a data point in the projection data corresponds to all pixel values ​​in an entire ray path passing through the CT image. Therefore, the error caused by projection interpolation can easily introduce new artifacts that are difficult to handle in the image domain, which has become the main reason why this type of method is difficult to break through; 2. Sparse angle CT imaging based on image domain post-processing. This type of method directly learns the mapping between sparse angle images and full-angle images. Although this type of method can effectively remove most of the noise and artifacts in the image, the convolution in the restoration process is equivalent to a filtering process, which inevitably causes the loss of structural detail information in the image, and the improvement of image quality is limited; 3. Sparse angle CT imaging based on projection-image cross-domain restoration. This type of method learns the mapping between sparse angle projection data and full-angle image data, effectively combining the projection domain and image domain data information, and can obtain imaging performance that is better than the projection domain or image domain alone.

[0005] At present, the research on sparse angles is mostly focused on fan beams under circular scanning trajectories. The sparse angle simulation process mainly includes two types: 1. Starting from the acquired image data, the image is projected into a fan beam to obtain its simulated fan beam projection data, and the projection data is uniformly sampled at sparse angles to obtain its corresponding sparse angle projection. This method lacks analysis of the real scanning geometry and spiral projection data and is difficult to be directly applied in clinical practice; 2. Starting from the real spiral projection data, the spiral projection data is rearranged into fan beam projections, and the fan beam projections are uniformly sampled to obtain their corresponding sparse angle projections. Although this method starts from the real projection data, its recovery performance will be greatly reduced in the case of large pitch. At present, there is almost no detailed work on the sparse angle CT reconstruction algorithm and system directly starting from spiral projection data, so it cannot be directly applied to clinical diagnosis and treatment.

[0006] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide a sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator to solve the shortcomings of the existing technology. Summary of the invention

[0007] The purpose of the present invention is to avoid the shortcomings of the prior art and provide a sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator constructs a differentiable spiral reconstruction operator with gradient backpropagation, builds a cross-domain deep learning framework for sparse spiral projection to high-quality reconstructed images, and realizes diagnostic-level sparse spiral CT reconstruction under a low-dose scanning protocol.

[0008] The above-mentioned purpose of the present invention is achieved by the following technical measures:

[0009] A sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator is provided, comprising the following steps:

[0010] S1, obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data;

[0011] S2, performing uniform sparse sampling on the full-angle spiral projection data obtained in S1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters;

[0012] S3, reconstructing the full-angle spiral CT image by a weighted filter back-projection algorithm based on the real spiral scanning geometric parameters and the full-angle spiral projection data of S1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by S2 by a weighted filter back-projection algorithm to obtain a sparse angle spiral CT image;

[0013] S4, performing 3D forward projection on the sparse angle spiral CT image obtained in S3 through the real spiral scanning geometric parameters of S1 to obtain the corresponding geometric prior projection;

[0014] S5, restore the projection data of the missing angles in the geometric prior projection obtained by S4 through the trained projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of S3 to obtain the final completed projection data;

[0015] S6, using a differentiable spiral reconstruction operator to perform a cross-domain connection from the projection domain to the image domain by combining the real spiral scanning geometric parameters in S1 with the final completed projection data obtained in S5, to obtain the initial restored image data after projection completion;

[0016] S7. Use the trained image domain deep learning model to perform detail restoration and residual artifact restoration on the sparse angle spiral CT image obtained in S3 and the initial restored image data obtained in S6 to obtain the final reconstructed image.

[0017] In the sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator of the present invention, the trained projection completion deep learning model and the trained image domain deep learning model are obtained by joint training, wherein each joint training is performed according to the following steps:

[0018] A1. Obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data;

[0019] A2, performing uniform sparse sampling on the full-angle spiral projection data obtained in A1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters;

[0020] A3, reconstructing the full-angle spiral CT image by weighted filtering back-projection algorithm based on the real spiral scanning geometric parameters and full-angle spiral projection data of A1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by A2 by weighted filtering back-projection algorithm to obtain a sparse angle spiral CT image;

[0021] A4, using the real spiral scanning geometric parameters of A1, 3D forward projection is performed on the sparse angle spiral CT image obtained in A3 to obtain the corresponding geometric prior projection;

[0022] A5, restore the projection data of the missing angles in the geometric prior projection obtained by A4 through the projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of A3 to obtain the final completed projection data;

[0023] A6, through the differentiable spiral reconstruction operator, the real spiral scanning geometric parameters in A1 and the final completed projection data obtained in A5 are connected from the projection domain to the image domain to obtain the initial restored image data after projection completion;

[0024] A7, performing detail restoration and residual artifact restoration on the sparse angle spiral CT image obtained in A3 and the initial restored image data obtained in A6 through an image domain deep learning model to obtain a final reconstructed image;

[0025] A8, calculate the loss through the loss function, where the final completed projection data of A5 is compared with the full-angle spiral projection data of A1 to obtain the loss The gradient of the final completed projection data of A5 is compared with the gradient of the full-angle spiral projection data of A1 to obtain the loss The initial restored image data of A6 is compared with the full-angle spiral CT image of A3 to obtain the loss The final reconstructed image of A7 is compared with the full-angle spiral CT image of A3 to obtain the loss

[0026] A9. The update of the image domain deep learning model is to convert the loss Directly pass the gradient back to the image domain deep learning model to update the parameters of the image domain deep learning model; the update of the projection completion deep learning model is to transfer the loss and loss Directly pass the gradient back to the projection completion deep learning model, and at the same time convert the loss and loss The gradient is returned to the projection completion deep learning model through the 3D differentiable gradient return operator, and finally the parameters of the projection completion deep learning model are updated.

[0027] In S2 or A2, uniform sparse sampling specifically involves uniform sampling of the full-angle projection data in the angle direction, as expressed by formula (1):

[0028]

[0029] Among them, P SV To simulate sparse angular spiral projection data, P FV is the full-angle spiral projection data, is a uniform sampling process in the angular direction, and N is a sparse sampling interval.

[0030] The reconstruction process of the weighted filter back-projection algorithm in S3 or A3 is expressed by equations (2) and (3):

[0031]

[0032] Among them, I FV It is a full-angle spiral CT image. SV is a sparse angle spiral CT image, α is the true spiral scanning geometric parameter, α ′ To simulate sparse sweep geometry, For full-angle reconstruction using true spiral scanning geometry, Sparse angular reconstruction using simulated sparse scan geometry.

[0033] In S4 or A4, the geometric prior projection uses the prior information of the real spiral scanning geometric parameters and the sparse angle CT image to obtain the initial projection value of the missing angle in the sparse angle scan, which is expressed by formula (4):

[0034]

[0035] Among them, P GPP is the geometric prior projection, For 3D front projection in true helical scanning geometry.

[0036] In S5 or A5, the projection completion deep learning model Φ is used. The projection completion deep learning model synthesizes the missing projection by learning the motion field between adjacent angles. The projection completion deep learning model is expressed by formula (5):

[0037] P Φ =Φ(P SV ,P GPP ,θ)……Formula (5);

[0038] Among them, P Φ is the final completed projection data, and θ is the learnable parameter of the projection completion deep learning model.

[0039] In S6, the differentiable spiral reconstruction operator consists of a 3D spiral reconstruction operator, and in A6, the differentiable spiral reconstruction operator consists of a 3D spiral reconstruction operator and a 3D front projection operator.

[0040] Preferably, the 3D spiral reconstruction operator transforms the complementary projection data from the projection domain to the image domain to obtain the initial restored image data. The 3D spiral reconstruction operator is represented by formula (6);

[0041]

[0042] Among them, I Φ To initially restore the image data, 3D spiral reconstruction operator.

[0043] Preferably, the above 3D front projection operator transforms the initial restored image data from the image domain to the projection domain, which is expressed by equation (7);

[0044]

[0045] in, is the 3D front projection operator, P ′ Φ The projection data is obtained by performing 3D front projection on the initial restored image data, and I is the data that needs to be gradient-transmitted from the image to the projection.

[0046] In S7 or A7, the image domain deep learning model uses sparse angle CT images and projection-completed image data to restore details and residual artifacts, uses a multi-scale algorithm to extract global image information, and uses a channel attention module to extract information between different layers to obtain the final reconstructed image; the image domain deep learning model is expressed by formula (8):

[0047]

[0048] in, is the image domain deep learning model, ω is the learnable parameter of the image domain deep learning model, for the final reconstructed image.

[0049] In A8, the loss function is expressed by equations (9) to (13):

[0050]

[0051]

[0052] in, is the total loss function, β1 is the loss The corresponding weight, β2 is the loss The corresponding weight, β3 is the loss The corresponding weight, β4 is the loss The corresponding weight, N D is the total number of angles corresponding to the full-angle spiral projection data, M D is the number of images corresponding to the full-angle spiral CT image, is the reconstruction operator.

[0053] In A9, gradient backpropagation is to perform gradient backpropagation on the learnable parameters, and the projection domain Φ is combined with the loss The relevant learnable parameter is defined as θ, which is the projection domain and loss loss and loss The relevant learnable parameter is defined as ω;

[0054] In A9, the update of the image domain deep learning model is expressed by equation (14):

[0055]

[0056] In A9, the update of the projection completion deep learning model is expressed by equations (15) and (16):

[0057]

[0058] Among them, I loss is the loss between the current image and the full-angle spiral CT image, and Middle I loss for exist Middle I loss for

[0059] The actual spiral scanning geometric parameters in S1 or A1 are the distance from the X-ray source to the detector and the rotation center, the detector size, the detector unit size, the number of exposure angles in one circle, the collimator width, the pitch, and the starting scanning angle.

[0060] The present invention discloses a sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator, comprising the following steps: S1, obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data; S2, performing uniform sparse sampling on the full-angle spiral projection data obtained by S1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters; S3, reconstructing the full-angle spiral CT image by a weighted filtering back-projection algorithm based on the real spiral scanning geometric parameters and the full-angle spiral projection data of S1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by S2 by a weighted filtering back-projection algorithm to obtain a sparse angle spiral CT image; S4, performing 3D front projection on the sparse angle spiral CT image obtained by S3 by the real spiral scanning geometric parameters of S1 to obtain the corresponding geometric prior projection;

[0061] S5. Restore the projection data of the missing angles in the geometric prior projection obtained by S4 through the trained projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of S3 to obtain the final completed projection data; S6. Use the differentiable spiral reconstruction operator to connect the real spiral scanning geometric parameters in S1 and the final completed projection data obtained by S5 from the projection domain to the image domain to obtain the initial restored image data after projection completion; S7. Use the trained image domain deep learning model to restore details and residual artifacts on the sparse angle spiral CT image obtained by S3 and the initial restored image data obtained by S6 to obtain the final reconstructed image. Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a geometric prior forward projection, the real scanning geometry is used to perform forward projection on the reconstructed sparse angle image, providing geometric prior guidance for the missing projection; 2. According to the similarity and redundancy characteristics of adjacent projections of spiral scanning, a projection completion network is constructed, and the intermediate missing projection data is synthesized by learning the bidirectional motion field of adjacent angles and combining the geometric prior projection; 3. An image domain fine-tuning module is constructed to realize the restoration of global strip artifacts in the image; 4. In view of the problems that the loss gradient cannot be returned and the amount of calculation is large during dual-domain training, on the one hand, a differentiable spiral reconstruction operator with gradient return is constructed to realize joint training of the projection domain and the image domain, which is conducive to combining the projection-image dual-domain information for overall restoration; on the other hand, during the training process, only the data collected within two pitches are used for restoration, which effectively avoids excessive calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0063] Figure 1This is a flowchart of a sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator.

[0064] Figure 2 Flowchart of joint training.

[0065] Figure 3 This is a comparison chart of the results before and after reconstruction using the algorithm of the present invention. DETAILED DESCRIPTION

[0066] The technical solution of the present invention is further described in conjunction with the following embodiments.

[0067] Example 1

[0068] A sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator, such as Figure 1 As shown, the following steps are included:

[0069] S1, obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data;

[0070] S2, performing uniform sparse sampling on the full-angle spiral projection data obtained in S1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters;

[0071] S3, reconstructing the full-angle spiral CT image by a weighted filter back-projection algorithm based on the real spiral scanning geometric parameters and the full-angle spiral projection data of S1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by S2 by a weighted filter back-projection algorithm to obtain a sparse angle spiral CT image;

[0072] S4, performing 3D forward projection on the sparse angle spiral CT image obtained in S3 through the real spiral scanning geometric parameters of S1 to obtain the corresponding geometric prior projection;

[0073] S5, restore the projection data of the missing angles in the geometric prior projection obtained by S4 through the trained projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of S3 to obtain the final completed projection data, thereby maintaining the consistency of the measurement data before and after recovery, wherein the projection completion deep learning model is constructed based on the similarity and redundancy characteristics between adjacent projections of spiral scanning;

[0074] S6, using a differentiable spiral reconstruction operator to perform a cross-domain connection from the projection domain to the image domain by combining the real spiral scanning geometric parameters in S1 with the final completed projection data obtained in S5, to obtain the initial restored image data after projection completion;

[0075] S7. Use the trained image domain deep learning model to perform detail restoration and residual artifact restoration on the sparse angle spiral CT image obtained in S3 and the initial restored image data obtained in S6 to obtain the final reconstructed image.

[0076] The trained projection completion deep learning model and the trained image domain deep learning model of the present invention are obtained by joint training, wherein each joint training is performed according to the following steps, such as Figure 2 As shown:

[0077] A1. Obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data;

[0078] A2, performing uniform sparse sampling on the full-angle spiral projection data obtained in A1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters;

[0079] A3, reconstructing the full-angle spiral CT image by weighted filtering back-projection algorithm based on the real spiral scanning geometric parameters and full-angle spiral projection data of A1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by A2 by weighted filtering back-projection algorithm to obtain a sparse angle spiral CT image;

[0080] A4, using the real spiral scanning geometric parameters of A1, 3D forward projection is performed on the sparse angle spiral CT image obtained in A3 to obtain the corresponding geometric prior projection;

[0081] A5, restore the projection data of the missing angles in the geometric prior projection obtained by A4 through the projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of A3 to obtain the final completed projection data, thereby maintaining the consistency of the measurement data before and after recovery, wherein the projection completion deep learning model is constructed based on the similarity and redundancy characteristics between adjacent projections of spiral scanning;

[0082] A6, through the differentiable spiral reconstruction operator, the real spiral scanning geometric parameters in A1 and the final completed projection data obtained in A5 are connected from the projection domain to the image domain to obtain the initial restored image data after projection completion;

[0083] A7, performing detail restoration and residual artifact restoration on the sparse angle spiral CT image obtained in A3 and the initial restored image data obtained in A6 through an image domain deep learning model to obtain a final reconstructed image;

[0084] A8, calculate the loss through the loss function, where the final completed projection data of A5 is compared with the full-angle spiral projection data of A1 to obtain the loss The gradient of the final completed projection data of A5 is compared with the gradient of the full-angle spiral projection data of A1 to obtain the loss The initial restored image data of A6 is compared with the full-angle spiral CT image of A3 to obtain the loss The final reconstructed image of A7 is compared with the full-angle spiral CT image of A3 to obtain the loss

[0085] A9. The update of the image domain deep learning model is to convert the loss Directly pass the gradient back to the image domain deep learning model to update the parameters of the image domain deep learning model; the update of the projection completion deep learning model is to transfer the loss and loss Directly pass the gradient back to the projection completion deep learning model, and at the same time convert the loss and loss The gradient is returned to the projection completion deep learning model through the 3D differentiable gradient return operator, and finally the parameters of the projection completion deep learning model are updated.

[0086] It should be noted that, in the present invention, the training of the projection completion deep learning model and the image domain deep learning model is completed when they converge after multiple trainings, and the condition for convergence may be that the number of trainings reaches a predetermined number of times, and the predetermined number of times is determined according to the amount of data and the size of the network.

[0087] In S2 or A2, uniform sparse sampling specifically involves uniform sampling of the full-angle projection data in the angle direction, as expressed by formula (1):

[0088]

[0089] Among them, P SV To simulate sparse angular spiral projection data, P FV is the full-angle spiral projection data, is a uniform sampling process in the angular direction, and N is a sparse sampling interval.

[0090] The reconstruction process of the weighted filter back-projection algorithm in S3 or A3 is expressed by equations (2) and (3):

[0091]

[0092] Among them, I FV It is a full-angle spiral CT image. SV is a sparse angle spiral CT image, α is the true spiral scanning geometric parameter, α ′ To simulate sparse sweep geometry parameters, For full-angle reconstruction using true spiral scanning geometry, Sparse angular reconstruction using simulated sparse scan geometry.

[0093] It should be noted that since the sampling rate of sparse angular projection data does not satisfy Shannon's sampling theorem, directly using the traditional filtered back projection algorithm for reconstruction will result in a large number of stripe artifacts and noise in the image, affecting clinical diagnosis. Therefore, the present invention uses a weighted filtered back projection algorithm for reconstruction.

[0094] In S4 or A4, the geometric prior projection uses the prior information of the real spiral scanning geometric parameters and the sparse angle CT image to obtain the initial projection value of the missing angle in the sparse angle scan, which is expressed by formula (4):

[0095]

[0096] Among them, P GPP is the geometric prior projection, For 3D front projection in true helical scanning geometry.

[0097] It should be noted that the geometric prior projection is helpful to guide the recovery of the missing angle

[0098] In S5 or A5, the projection completion deep learning model Φ is used. The projection completion deep learning model synthesizes the missing projection by learning the motion field between adjacent angles. The projection completion deep learning model is expressed by formula (5):

[0099] P Φ =Φ(P SV ,P GPP ,θ)……Formula (5);

[0100] Among them, P Φ is the final completed projection data, and θ is the learnable parameter of the projection completion deep learning model. The final completed projection data still retains the original measurement data at the true sparse scanning angle position, which can maintain the consistency of the data before and after projection restoration.

[0101] In S6, the differentiable spiral reconstruction operator consists of a 3D spiral reconstruction operator, and in A6, the differentiable spiral reconstruction operator consists of a 3D spiral reconstruction operator and a 3D front projection operator;

[0102] The 3D spiral reconstruction operator transforms the completed projection data from the projection domain to the image domain to obtain the initial restored image data. The 3D spiral reconstruction operator is expressed by formula (6);

[0103]

[0104] Among them, I Φ To initially restore the image data, is the 3D spiral reconstruction operator,

[0105] The 3D front projection operator transforms the initial restored image data from the image domain to the projection domain, which is expressed by equation (7);

[0106]

[0107] in, is the 3D front projection operator, P ′ Φ The projection data obtained by performing 3D front projection on the initial restored image data, I Φ The data that needs to be returned from the image domain to the projection domain for gradient.

[0108] It should be noted that the projection data P ′ Φ Since the image has data truncation on the z-axis, it is different from P Φ In comparison, there is data truncation in the detector channel direction, but this part of the data truncation has no effect on the image to be reconstructed.

[0109] In S7 or A7, the image domain deep learning model uses sparse angle CT images and projection-completed image data to restore details and residual artifacts, uses a multi-scale algorithm to extract global image information, and uses a channel attention module to extract information between different layers to obtain the final reconstructed image; the image domain deep learning model is expressed by formula (8):

[0110]

[0111] in, is the image domain deep learning model, ω is the learnable parameter of the image domain deep learning model, for the final reconstructed image.

[0112] In A8, the loss function is expressed by equations (9) to (13):

[0113]

[0114] in, is the total loss function, β1 is the loss The corresponding weight, β2 is the loss The corresponding weight, β3 is the loss The corresponding weight, β4 is the loss The corresponding weight, N D is the total number of angles corresponding to the full-angle spiral projection data, M D is the number of images corresponding to the full-angle spiral CT image, is the reconstruction operator.

[0115] It should be noted that in A9, the gradient backpropagation is to backpropagate the gradient of the learnable parameters, and the projection domain Φ can be used to learn the parameter θ loss. Related, the projection domain The learnable parameters ω and loss loss and loss Since θ and ω are learnable parameters, we need to calculate and

[0116] In A9, the update of the image domain deep learning model is expressed by equation (14):

[0117]

[0118] In A9, the update of the projection completion deep learning model is expressed by equations (15) and (16):

[0119]

[0120] Among them, I loss is the loss between the current image and the full-angle spiral CT image, and Middle I loss for exist Middle I loss for

[0121] The actual spiral scanning geometric parameters in S1 or A1 are the distance from the X-ray source to the detector and the rotation center, the detector size, the detector unit size, the number of exposure angles in one circle, the collimator width, the pitch, and the starting scanning angle; wherein the spiral projection data of the present invention, such as full-angle spiral projection data, sparse-angle spiral projection data, etc., refers to the projection data after logarithmic transformation.

[0122] Both the projection completion deep learning model and the image domain deep learning model are U-net models.

[0123] The sparse spiral CT image reconstruction method based on the differentiable spiral reconstruction operator realizes the joint training of real spiral projection data to 3D images by constructing a differentiable spiral reconstruction operator, and fully utilizes the information of the dual domains of the projection image to perform diagnostic-level image reconstruction; at the same time, the redundant information between adjacent spiral projection data is used to restore the missing angle, which fully utilizes the information of the projection data and greatly improves the projection restoration performance. Moreover, the present invention can be directly applied to clinical scanning. During the clinical spiral scanning process, the final CT image can be obtained by inputting the scanning parameters and the scanned sparse spiral projection data of each case into the present invention. The present invention is directly applied to clinical CT scanning and can be applicable to different scanning conditions and scanning models.

[0124] The sparse spiral CT image reconstruction method based on the differentiable spiral reconstruction operator has the following beneficial effects: 1. By constructing a geometric prior forward projection, the real scanning geometry is used to perform forward projection on the reconstructed sparse angle image, providing geometric prior guidance for the missing projection; 2. According to the similarity and redundancy characteristics of adjacent projections of spiral scanning, a projection completion network is constructed, and the intermediate missing projection data is synthesized by learning the bidirectional motion field of adjacent angles and combining the geometric prior projection; 3. An image domain fine-tuning module is constructed to realize the restoration of global strip artifacts of the image; 4. In view of the problems that the loss gradient cannot be returned during dual-domain training and the amount of calculation is large, on the one hand, a differentiable spiral reconstruction operator with gradient return is constructed to realize joint training in the projection domain and the image domain, which is conducive to combining the projection-image dual-domain information for overall restoration; on the other hand, during the training process, only the data collected within two pitches are used for restoration, which effectively avoids excessive calculation.

[0125] Example 2

[0126] A sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator, the other features of which are the same as those of Example 1. In this embodiment, the actual spiral scanning geometric parameters of S1 and A1 are that the distances from the X-ray source to the rotation center and the detector are 1085.6 mm and 595 mm respectively, the exposure angle of one circle is 1152, the detector size is 736×64, and the detector unit size is 1.2858×1.0947.

[0127] In S2 and A2, uniform sparse sampling is performed using sparse 1 / 6 multiples (192 angles, N=6). In S4 and A4, in order to balance video memory usage and network performance, two circles of projection data are used for recovery, that is, sparse projection data of 384*736*64, which is completed into full-angle projection data of 2304*736*64. During model training, the strategy of projecting two adjacent frames and interpolating the missing frames in the middle is adopted. Due to the large amount of data, the data is adjusted to 384*1*736*64 during training, and the data of different adjacent frames are restored in parallel to reduce the amount of calculation, realize spiral projection completion, and lay the foundation for subsequent reconstruction.

[0128] In S6 and A6, the size of the final completed projection data is 32*512*512. In S7 and A7, in order to reduce the number of image domain deep learning model parameters during image domain training, this embodiment selects a 2.5D training strategy, that is, placing data of different slices in the channel direction.

[0129] exist Figure 3The first column is the full-angle spiral CT abdominal transverse and coronal images in S3, the second column is the sparse angle (N=6) spiral CT abdominal transverse and coronal images of S3, and the third column is the abdominal transverse and coronal images reconstructed by the present invention at the sparse angle (N=6). Figure 3 It can be seen that the cross-section and coronal planes of the sparse angle spiral CT images have more noise and artifacts, resulting in image distortion. These artifacts overlap and cover up the real tissue structure, affecting the doctor's diagnosis. After being processed by the method of the present invention, Figure 3 The images in the image have changed significantly. The corrected images can effectively remove a lot of noise and strip artifacts in sparse angle spiral CT images, making the detailed structure of the tissue clearly visible.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator, characterized in that: The steps include: S1, obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data; S2, performing uniform sparse sampling on the full-angle spiral projection data obtained in S1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters; S3, reconstructing the full-angle spiral CT image by a weighted filter back-projection algorithm based on the real spiral scanning geometric parameters and the full-angle spiral projection data of S1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by S2 by a weighted filter back-projection algorithm to obtain a sparse angle spiral CT image; S4, performing 3D forward projection on the sparse angle spiral CT image obtained in S3 through the real spiral scanning geometric parameters of S1 to obtain the corresponding geometric prior projection; S5, restore the projection data of the missing angles in the geometric prior projection obtained by S4 through the trained projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of S3 to obtain the final completed projection data; S6, using a differentiable spiral reconstruction operator to perform a cross-domain connection from the projection domain to the image domain by combining the real spiral scanning geometric parameters in S1 with the final completed projection data obtained in S5, to obtain the initial restored image data after projection completion; S7. Use the trained image domain deep learning model to perform detail restoration and residual artifact restoration on the sparse angle spiral CT image obtained in S3 and the initial restored image data obtained in S6 to obtain the final reconstructed image.

2. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 1, characterized in that: The trained projection completion deep learning model and the trained image domain deep learning model are obtained by joint training, wherein each joint training is performed according to the following steps: A1. Obtaining the real spiral scanning geometric parameters of the object and the corresponding full-angle spiral projection data; A2, performing uniform sparse sampling on the full-angle spiral projection data obtained in A1 to obtain simulated sparse angle spiral projection data and simulated sparse scanning geometric parameters; A3, reconstructing the full-angle spiral CT image by weighted filtering back-projection algorithm based on the real spiral scanning geometric parameters and full-angle spiral projection data of A1; and reconstructing the sparse angle spiral projection data and the simulated sparse scanning geometric parameters obtained by A2 by weighted filtering back-projection algorithm to obtain a sparse angle spiral CT image; A4, using the real spiral scanning geometric parameters of A1, 3D forward projection is performed on the sparse angle spiral CT image obtained in A3 to obtain the corresponding geometric prior projection; A5, restore the projection data of the missing angles in the geometric prior projection obtained by A4 through the projection completion deep learning model to obtain the initial completed projection data, and then replace the corresponding position in the completed projection data with the simulated sparse angle spiral projection data of A3 to obtain the final completed projection data; A6, through the differentiable spiral reconstruction operator, the real spiral scanning geometric parameters in A1 and the final completed projection data obtained in A5 are connected from the projection domain to the image domain to obtain the initial restored image data after projection completion; A7, performing detail restoration and residual artifact restoration on the sparse angle spiral CT image obtained in A3 and the initial restored image data obtained in A6 through an image domain deep learning model to obtain a final reconstructed image; A8, calculate the loss through the loss function, where the final completed projection data of A5 is compared with the full-angle spiral projection data of A1 to obtain the loss The gradient of the final completed projection data obtained by A5 is compared with the gradient of the full-angle spiral projection data of A1 to obtain the loss The initial restored image data of A6 is compared with the full-angle spiral CT image of A3 to obtain the loss The final reconstructed image of A7 is compared with the full-angle spiral CT image of A3 to obtain the loss A9. The update of the image domain deep learning model is to change the loss Directly pass the gradient back to the image domain deep learning model to update the parameters of the image domain deep learning model; the update of the projection completion deep learning model is to transfer the loss and loss Directly pass the gradient back to the projection completion deep learning model, and at the same time convert the loss and loss The gradient is returned to the projection completion deep learning model through the 3D differentiable gradient return operator, and finally the parameters of the projection completion deep learning model are updated.

3. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 2, characterized in that: In S2 or A2, uniform sparse sampling specifically involves uniform sampling of the full-angle projection data in the angle direction, as expressed by formula (1): Among them, P SV To simulate sparse angular spiral projection data, P FV is the full-angle spiral projection data, is a uniform sampling process in the angular direction, and N is a sparse sampling interval.

4. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 3, characterized in that: The reconstruction process of the weighted filter back-projection algorithm in S3 or A3 is expressed by equations (2) and (3): Among them, I FV It is a full-angle spiral CT image. SV is a sparse angle spiral CT image, α is the true spiral scanning geometric parameter, α ′ To simulate sparse sweep geometry, For full-angle reconstruction using true spiral scanning geometry, Sparse angular reconstruction using simulated sparse scan geometry.

5. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 4, characterized in that: In S4 or A4, the geometric prior projection uses the prior information of the real spiral scanning geometric parameters and the sparse angle CT image to obtain the initial projection value of the missing angle in the sparse angle scan, which is expressed by formula (4): Among them, P GPP is the geometric prior projection, For 3D front projection in true helical scanning geometry.

6. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 5, characterized in that: In S5 or A5, the projection completion deep learning model Φ is used. The projection completion deep learning model synthesizes the missing projection by learning the motion field between adjacent angles. The projection completion deep learning model is expressed by formula (5): P Φ = Φ(P SV , P GPP , θ) …… Equation (5); Among them, P Φ is the final completed projection data, and θ is the learnable parameter of the projection completion deep learning model.

7. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 6, characterized in that: In S6, the differentiable spiral reconstruction operator consists of a 3D spiral reconstruction operator, and in A6, the differentiable spiral reconstruction operator consists of a 3D spiral reconstruction operator and a 3D front projection operator; The 3D spiral reconstruction operator transforms the complementary projection data from the projection domain to the image domain to obtain the initial restored image data. The 3D spiral reconstruction operator is expressed by formula (6); Among them, I Φ To initially restore the image data, is the 3D spiral reconstruction operator, The 3D front projection operator transforms the initial restored image data from the image domain to the projection domain, which is expressed by equation (7); in, is the 3D front projection operator, P ′ Φ The projection data is obtained by performing 3D front projection on the initial restored image data, and I is the data that needs to be gradient-transmitted from the image domain to the projection domain.

8. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 7, characterized in that: In S7 or A7, the image domain deep learning model uses sparse angle CT images and projection-completed image data to restore details and residual artifacts, uses a multi-scale algorithm to extract global image information, and uses a channel attention module to extract information between different layers to obtain the final reconstructed image; the image domain deep learning model is expressed by formula (8): in, is the image domain deep learning model, ω is the learnable parameter of the image domain deep learning model, for the final reconstructed image.

9. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 8, characterized in that: In A8, the loss function is expressed by equations (9) to (13): in, is the total loss function, β1 is the loss The corresponding weight, β2 is the loss The corresponding weight, β3 is the loss The corresponding weight, β4 is the loss The corresponding weight, N D is the total number of angles corresponding to the full-angle spiral projection data, m D is the number of images corresponding to the full-angle spiral CT image, is the reconstruction operator.

10. The sparse spiral CT image reconstruction method based on a differentiable spiral reconstruction operator according to claim 9, characterized in that: In A9, the update of the image domain deep learning model is expressed by equation (14): In A9, the update of the projection completion deep learning model is expressed by equations (15) and (16): Among them, I loss is the loss between the current image and the full-angle spiral CT image, and Middle I loss for exist Middle I loss for The actual spiral scanning geometric parameters in S1 or A1 are the distance from the X-ray source to the detector and the rotation center, the detector size, the detector unit size, the number of exposure angles in one circle, the collimator width, the pitch, and the starting scanning angle.

Images