Method and system for ct image reconstruction based on x-ray film sampling position prediction

Abstract

This invention discloses a CT image reconstruction method and system based on X-ray image sampling location prediction. The method includes acquiring X-ray images I. X According to X-ray image I X Calculate the optimal sampling position distribution for subsequent sampling of individual CT images; perform sampling based on the optimal sampling position distribution to obtain the sparse projection y. u ; for sparse projection y u Denoising and completion are performed to obtain the projection signal y. e According to the projection signal y e and sparse projection y u Reconstructing CT images u. This invention performs personalized sampling and reconstruction based on the individual characteristics of patients, saving imaging time and improving image reconstruction performance at the same dose. It also utilizes dual domain constraints, resulting in better robustness.

Description

Technical Field

[0001] This invention relates to the field of CT image reconstruction technology, and in particular to a CT image reconstruction method and system based on X-ray film sampling location prediction. Background Technology

[0002] Computed tomography (CT) is a technique that uses the attenuation of X-rays in different tissues as they pass through the human body to create differential images. Due to its high imaging speed and quality, this technology is widely used in clinical diagnosis and treatment. However, X-ray radiation poses certain risks to the human body; excessive radiation can lead to cell mutation and cancer, which limits its clinical application. Therefore, developing high-quality CT imaging algorithms for low-dose conditions has become an important research area in this field.

[0003] Currently, mainstream CT imaging algorithms first acquire the projection information of the human body by uniformly distributing X-ray sources around it. Then, they introduce expert and data prior information as regularization constraints on the reconstructed image to design reconstruction algorithms suitable for high-quality CT imaging in low-dose scenarios. Generally, low-dose scenarios mainly involve 1) reducing the radiation dose from X-ray sources; 2) reducing the number of X-ray projections; and 3) limiting the X-ray projection angle. These three methods aim to reduce radiation, but all three cause additional image artifacts. Therefore, the aforementioned expert and data prior-based methods can utilize prior information to make the reconstructed image more consistent with the image features of the prior information, thereby suppressing image artifacts caused by low-dose scenarios to some extent. However, these methods also suffer from problems such as high computational cost and susceptibility to overfitting to training data, making them difficult to apply to actual CT imaging equipment.

[0004] Meanwhile, as a framework dedicated to finding high-quality imaging in low-dose scenarios, the aforementioned traditional CT imaging algorithms, whether utilizing prior knowledge or data to build reconstruction models, fail to consider the clinical needs of physicians. Specifically, traditional CT equipment acquires observational data using a fixed uniform sampling pattern, and some of the data acquired in this way is not adaptively designed according to the clinical task. This makes it impossible to adaptively cover the most advantageous position for current imaging with the applied dose. Noise based on this uniform sampling pattern may be amplified, thus wasting some dose, reducing the performance of the reconstruction algorithm, and hindering the efficiency of physician diagnosis and treatment. This necessitates the design of a better sampling pattern to improve the performance of the imaging algorithm.

[0005] Recent technologies have proposed active imaging for dynamic optimization and design of sampling locations. However, existing methods for optimizing sampling locations often lack the utilization of clinical information, creating a disconnect between imaging algorithms and clinical scenarios. In orthopedic examinations (such as pelvic and leg bone examinations), patients undergo pre-examination X-rays before CT scans. Similarly, during CT imaging, radiologists use low-dose CT images for localization before performing standard-dose CT scans. This clinical information provides highly specific guidance, offering information on the patient's posture, position, and contours for optimizing sampling locations, thus providing significant guidance.

[0006] It is evident that both traditional low-dose CT reconstruction algorithms, which rely on fixed uniform sampling patterns for imaging, and recent active imaging methods that dynamically optimize and design sampling locations, suffer from the following drawbacks:

[0007] First, traditional knowledge- and data-driven reconstruction algorithms suffer from problems such as excessive computational cost and overfitting.

[0008] Second, online active CT imaging algorithms suffer from problems such as excessively long online search times and overly complex hardware designs.

[0009] Third, offline sampling location recommendation and imaging algorithms lack the use of clinical information. Summary of the Invention

[0010] To overcome the above-mentioned shortcomings, the present invention aims to provide a CT image reconstruction method and system based on X-ray sampling location prediction, which can perform personalized sampling and reconstruction according to the different individual characteristics of patients. Compared with online sampling location prediction, it saves more imaging time and can improve the image reconstruction performance under the same dose.

[0011] To achieve the above objectives, the technical solution adopted by this invention is: a CT image reconstruction method based on X-ray film sampling position prediction, comprising the following steps:

[0012] Obtain at least one X-ray image I X ;

[0013] According to the X-ray image I X Calculate the optimal sampling location distribution for subsequent sampling of individual CT images;

[0014] Sampling is performed based on the optimal sampling position distribution to obtain the sparse projection y. u ;

[0015] Load the pre-trained first deep learning network model onto the sparse projection y u Denoising and completion are performed to obtain the projection signal y.e ;

[0016] The sparse projection y u and projection signal y e FBP images were reconstructed using the analytical FBP algorithm respectively. and FBP image

[0017] Load the pre-trained second deep learning network model onto the FBP image and FBP image The images are fused together to form a reconstructed CT image.

[0018] The beneficial effects of this invention are as follows:

[0019] First, based on X-ray images, individual factors are learned in advance, and optimal sampling locations are predicted according to these individual factors, demonstrating the ability for individual adaptive sampling. This reduces the limitations of fixed sampling patterns on the performance of CT reconstruction models, achieving optimal sampling and high-quality reconstruction at equivalent doses.

[0020] Second, the sparse projection information is enhanced by the deep learning algorithm of the first deep learning network model, and the image domain is reconstructed by the deep learning algorithm of the second deep learning network to obtain a higher quality reconstructed image.

[0021] Furthermore, based on the X-ray image I X The calculation of the optimal sampling location distribution for subsequent sampling specifically includes:

[0022] A sampling location prediction model is constructed, which includes an optimizable sampling operator P based on a Gaussian mixture distribution and a prior R2(P) for the sampling operator P.

[0023] The sampling location prediction model is based on the currently acquired X-ray image I. X Calculate parameterized distribution Where θ X X-ray image I X The encoding, θ g These are trainable parameters based on prior R2(P);

[0024] Using Monte Carlo methods in parameterized distributions Sampling is performed on the image to obtain the image I from the current X-ray. X The optimal sampling location distribution for pairing.

[0025] Based on prior information from X-ray images, a parameterized Gaussian mixture distribution is constructed. Due to the introduction of the parameterized Gaussian mixture distribution, individual factors can be learned in advance based on X-ray images, and the optimal sampling location can be predicted based on individual factors, thus having the ability to perform adaptive sampling based on individual factors.

[0026] Furthermore, the sampling location prediction model is as follows:

[0027] min u, D(PAu,y u )+λR1(u)+βR2(P),

[0028] Where λ and β are adjustable hyperparameters, u is the CT image to be reconstructed, A is the full sampling operator, R1(u) is the image prior of u, and y u It is the sparse projection obtained by sampling, and D(*) is the distance function.

[0029] The sampling location prediction model uses R1(u) and R2(P) as constraints. The model is then parameterized and improved, and the current optimal sampling location distribution is obtained by sampling from this distribution. Based on a parameterized Gaussian mixture distribution of prior information from X-ray images, the optimal sampling location for the patient corresponding to the X-ray image is predicted.

[0030] Furthermore, the image prior R1 includes the data prior learned through the second deep learning network model and the pixel gradient smoothing prior TV(u) of the image, and the image prior R1 constrains the test image u.

[0031] The prior R2 includes X-ray image I X The sampling quantity, sampling position, and distance function d between adjacent sampling positions are given by the prior R2, which constrains the sampling operator P.

[0032] Furthermore, the training process of the first deep learning network model includes:

[0033] The first deep network structure is constructed based on the Swin-Transformer network structure with time preservation characteristics;

[0034] Collect a first training dataset, which includes data pairs formed by collected CT images and corresponding X-ray images;

[0035] The first deep network structure is trained using the first training dataset, and the trained data is saved as the first deep learning network model.

[0036] By leveraging the Swing Transformer architecture to extract projection features and constructing correlations between multiple projections, highly sparse projections can be interpolated and completed to obtain the predicted projection signal y.e .

[0037] Furthermore, the training process of the first deep learning network model employs joint supervision using L1 and SSIM loss functions to improve the ability to learn edge features of the projection map.

[0038] Furthermore, the training process of the second deep learning network model includes:

[0039] A second deep network structure is constructed based on a hybrid architecture of convolutional neural network (CNN) and visual Transformer structure;

[0040] Construct a second training dataset, which includes data pairs consisting of low-quality CT images and high-quality CT images. The low-quality CT images are constructed from M collected X-ray images using the FBP parsing algorithm, and the high-quality CT images are constructed from N collected X-ray images using the FBP parsing algorithm, where N is greater than M.

[0041] The second deep network structure is trained using the second training dataset, and the trained data is saved as the second deep learning network model.

[0042] By leveraging the powerful local feature representation capabilities of CNNs and the global feature learning capabilities of the visual Transformer architecture, high-quality CT image reconstruction is performed, resulting in the final reconstructed CT image u.

[0043] Furthermore, during the training process of the second deep learning network model, three loss functions—L1, texture-aware loss, and SSIM—are used for joint supervision to improve the final image reconstruction performance.

[0044] This invention also discloses a CT image reconstruction system based on X-ray film sampling location prediction, comprising:

[0045] The collection module is used to acquire X-ray images. X ;

[0046] The sampling position prediction module, based on the X-ray image I X The optimal sampling position distribution for subsequent individual CT images is calculated, and sampling is performed based on the optimal sampling position distribution to obtain the sparse projection y. u ;

[0047] A projection domain completion module, which stores a trained first deep learning network model, is used to complete the sparse projection y. u Denoising and completion are performed to obtain the projection signal y. e ;

[0048] An image domain reconstruction module, which stores a second deep learning network, is used to reconstruct the sparse projection y. u and projection signal y e FBP images were reconstructed using the analytical FBP algorithm respectively. and FBP image Fusing the FBP images and FBP image And form a reconstructed CT image u.

[0049] The present invention also discloses a storage medium, characterized in that: the computer-readable storage medium stores instructions, which, when executed by a processor, are used to implement the above-mentioned CT image reconstruction method based on X-ray film sampling position prediction. Attached Figure Description

[0050] Figure 1 The process of Embodiment 1 of the present invention Figure 1 ;

[0051] Figure 2 The process of Embodiment 1 of the present invention Figure 2 ;

[0052] Figure 3 The process of Embodiment 1 of the present invention Figure 3 ;

[0053] Figure 4 The process of Embodiment 1 of the present invention Figure 4 ;

[0054] Figure 5 This is a system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0055] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0056] Example 1

[0057] The CT image reconstruction method based on X-ray film sampling location prediction of the present invention is described in the appendix. Figure 1 As shown, the method includes the following steps:

[0058] S100. Obtain at least one X-ray image. X .

[0059] X-ray image I X The data was obtained when the sampling locations were evenly distributed.

[0060] S200, Based on X-ray image I X Calculate the optimal sampling location distribution for subsequent sampling of individual CT images.

[0061] S300. Sampling is performed according to the optimal sampling position distribution to obtain the sparse projection y. u .

[0062] S400, Load the pre-trained first deep learning network model onto the sparse projection y u Denoising and completion are performed to obtain the projection signal y. e .

[0063] S500, sparse projection y u and projection signal y e FBP images were reconstructed using the analytical FBP algorithm respectively. and FBP image

[0064] S600, Load the trained second deep learning network model onto the FBP image. and FBP image The images are fused together to form a reconstructed CT image.

[0065] Compared to existing CT image reconstruction methods, this embodiment can learn individual patient factors in advance based on X-ray images and predict the optimal sampling location based on these factors, demonstrating patient-adaptive sampling capabilities. This reduces the limitations imposed by fixed sampling patterns on CT reconstruction model performance, achieving optimal sampling and high-quality reconstruction at equivalent doses. Simultaneously, it utilizes a first deep learning network model's deep learning algorithm to enhance sparse projection information and a second deep learning network's deep learning algorithm for image domain reconstruction, obtaining higher-quality reconstructed images. This embodiment's method, by combining X-ray image information and deep learning technology, achieves personalized sampling location prediction and reconstruction, bringing greater personalization and quality to CT image reconstruction. In practice, this method is expected to improve the efficiency and accuracy of CT image reconstruction, while overcoming the limitations of fixed sampling patterns and reconstruction time in online algorithms of traditional methods, bringing new possibilities to the field of medical imaging.

[0066] In one embodiment, 10 X-ray images are required. X To reconstruct the CT images, in step S100, three X-ray images are acquired. X That is, through these 3 X-ray images I X To compare with the subsequent 7 X-ray images I X The sampling location is predicted to achieve accurate prediction of the sampling location.

[0067] In one embodiment, see Appendix Figure 2As shown, based on the X-ray image I X The calculation of the subsequent optimal sampling location distribution specifically includes:

[0068] S21. Construct a sampling location prediction model.

[0069] The sampling location prediction model includes an optimizable sampling operator P based on a Gaussian mixture distribution and a prior R2(P) for the sampling operator P.

[0070] S22, The sampling location prediction model is based on the currently acquired X-ray image I X Calculate parameterized distribution

[0071]

[0072] Where θ X X-ray image I X The encoding, θ g represents the trainable parameters based on prior R2(P).

[0073] S23. Using the Monte Carlo method in parameterized distributions Sampling is performed on the image to obtain the image I from the current X-ray. X The optimal sampling location distribution for pairing.

[0074] In this embodiment, a parameterized Gaussian mixture distribution is constructed based on prior information from X-ray images. Due to the introduction of this parameterized Gaussian mixture distribution, this embodiment can learn individual patient factors in advance based on X-ray images and predict the optimal sampling location based on these factors, demonstrating the ability to adaptively sample based on individual patient factors. Compared to online active CT reconstruction algorithms, the sampling location prediction module can predict the sampling location in advance based on individual patient information due to the inclusion of X-ray image information, thus reducing reconstruction time.

[0075] The sampling location prediction model is as follows:

[0076] min u, D(PAu,y u )+λR1(u)+βR2(P),

[0077] Where λ and β are adjustable hyperparameters, u is the CT image to be reconstructed, A is the full sampling operator, R1(u) is the image prior of u, and y u It is the sparse projection obtained by sampling, and D(*) is the distance function.

[0078] In this embodiment, the sampling location prediction model uses R1(u) and R2(P) as constraints, and the model is parameterized and improved. The current optimal sampling location distribution is obtained by sampling from this distribution. Based on the parameterized Gaussian mixture distribution of prior information from X-ray images, the optimal sampling location of the patient corresponding to the X-ray image is predicted.

[0079] In one embodiment, the image prior R1 constrains the test image u, including the data prior learned by the second deep learning network model and the pixel gradient smoothing prior TV(u) of the image.

[0080] The data prior learned through the second deep learning network model is the mapping function f from low-quality CT images to high-quality CT images. I The data prior learned using the constructed second training dataset. TV(u)=∫|Δu|dx, where |Δu| is the absolute value of the difference between adjacent pixels in the image.

[0081] The prior R2 is a constraint on the sampling operator P, and the prior R2 includes the X-ray image I. X The number of samples, the sampling location, and the distance function d between adjacent sampling locations, where d = |p i -p i+1 |,p i p i+1 These represent the discrete geometric positions of the two samples.

[0082] The sampling location prediction model in this embodiment, by combining optimization techniques and statistical prior information, can more effectively predict the sampling location by optimizing the downsampling operator P and designing appropriate prior information.

[0083] In this embodiment, θ X It is formed by encoding pre-trained autoencoders, which use autoencoders to encode X-ray images I X Information is encoded into the latent space to obtain the corresponding code θ X This is used as a conditional encoding. A prior R²(P) is designed based on a Gaussian mixture model, introducing an additional trainable parameter θ. g and with θ X By merging, a parameterized distribution is obtained.

[0084] The sampling location prediction model also incorporates a gradient backpropagation mechanism for parameter training within the PyTorch framework. In deep learning, gradient backpropagation is a commonly used optimization algorithm for updating parameters in a neural network to minimize the loss function. PyTorch, a popular deep learning framework, provides convenient tools and interfaces for implementing gradient backpropagation. By utilizing the PyTorch framework, image reconstruction is performed using the optimal sampling location obtained from sampling, and gradient backpropagation is used to optimize the sampling location prediction model, thereby improving the quality and accuracy of reconstructed CT images.

[0085] The sampling location prediction model in this embodiment is trained using pre-collected CT and X-ray images, which cover different anatomical locations and clinical tasks. The sampling location prediction model uses the collected data to predict the sampling location.

[0086] In one embodiment, CT images and X-ray images can be collected from NIH-AAPM, LIDC, DeepLesion, and collaborating hospitals, and grouped according to anatomical location and clinical task. Within the same group, CT images and X-ray images are paired to form a CT image and its corresponding X-ray image data pair. For cases where only CT images are available, the physical simulation method of DeepDRR is used to generate the corresponding X-ray image for the CT image. The generated X-ray image and CT image are then paired to form a data pair for training the sampling location prediction model.

[0087] In one embodiment, the training process of the first deep learning network model includes:

[0088] S101. Construct the first deep network structure based on the Swin-Transformer network structure with time-preserving characteristics.

[0089] S102. Collect a first training dataset, which includes data pairs formed by collected CT images and corresponding X-ray images.

[0090] S103. Train the first deep network structure using the first training dataset, and save it as the first deep learning network model after training.

[0091] In the projection domain, for undersampled projection data - sparse projection y for different patients u First, completion is performed. Based on the temporal characteristics of the projection data, the Swing Transformer architecture is used to extract projection features and construct the correlation between multiple projections. This allows for interpolation and completion of the highly sparse projections, resulting in the predicted projection signal y. e .

[0092] The first deep learning network model can learn from sparse projection y during training. u The projection signal y to the full projection e The projection completion mapping function f s sparse projection y u After the mapping function f s Calculations can yield the predicted full projection data - the projection signal y. e .

[0093] See appendix Figure 3 As shown, the first deep learning network model includes two Swin Transformer modules with sparse projection y u The projected features are input into Swing Transformer module 1 and processed by the mapping function f between Swing Transformer module 1 and Swing Transformer module 2. s The full projection data - projection signal y is output by the Swing Transformer module 2. e .

[0094] In one embodiment, the training process of the first deep learning network model employs joint supervision using L1 and SSIM loss functions to improve its ability to learn edge features of the projection image. The L1 loss function is typically used to minimize pixel-level differences between the reconstructed image and the real image, while SSIM (Structural Similarity Metric) measures the structural similarity of images, helping to preserve structural information. By jointly using these two loss functions, the details and structural information of the reconstructed image can be comprehensively considered, thereby improving the reconstruction quality.

[0095] In one embodiment, the training process of the second deep learning network model includes:

[0096] S201. A second deep network structure is constructed based on a hybrid architecture of convolutional neural network (CNN) and visual Transformer.

[0097] By combining CNN and visual Transformer architectures for image domain reconstruction, we can achieve the extraction of global and local information of image features and high-quality imaging.

[0098] S202. Construct a second training dataset, which includes data pairs formed by low-quality CT images and high-quality CT images.

[0099] The low-quality CT image is constructed from M collected X-ray images using the FBP (Framework Backpropagation) algorithm, while the high-quality CT image is constructed from N collected X-ray images using the FBP algorithm, where N is greater than M. The high-quality CT image has higher image clarity than the low-quality CT image, both being constructed from collected X-ray images using the FBP algorithm.

[0100] S203. Train the second deep network structure using the second training dataset, and save it as the second deep learning network model after training.

[0101] In one embodiment, the enhanced projection signal y e There will be some error, and the sparse projection y of the original sample... u Having accurate, realistic, and sparse information, the analytical reconstruction method FBP is used to first reconstruct the two separately to obtain... and Then, an image fusion algorithm is used to reconstruct the two images. and By fusing the images, features that closely resemble those of the real CT image can be preserved. Finally, a hybrid network architecture combining CNN and Visual Transformer is used to reconstruct the fused image. By leveraging the powerful local feature representation capabilities of CNN and the global feature learning capabilities of the Visual Transformer architecture, a high-quality CT image reconstruction is achieved, resulting in the final reconstructed CT image u.

[0102] The trained second deep learning network model has learned the mapping function f from low-quality CT images to high-quality CT images. I The merged graphic is processed by the mapping function f I Then, a CT image u is generated. See appendix. Figure 4 As shown, the pre-trained second deep learning network model is loaded onto the FBP image. and FBP image The process of fusing and reconstructing CT images specifically includes:

[0103] S61, Image Fusion: Combine FBP images and FBP image The images are then fused together to form a fused image.

[0104] S62 and CNN networks perform shallow feature extraction on the fused images.

[0105] S63. The Visual Transformer network performs deep feature extraction on the fused image.

[0106] S64. The CNN network reconstructs the image based on the extracted shallow and deep features, forming a reconstructed CT image u.

[0107] In one embodiment, the training process of the second deep learning network model utilizes three loss functions—L1, texture-aware loss, and SSIM—for joint supervision to improve the final image reconstruction performance.

[0108] In one implementation, a total of 100 rounds of learning and gradient updates were performed on the data, with the learning rate starting at 0.001 and decreasing by 10% every 20 rounds to gradually stabilize the learning process of the sampling location prediction model. The trained sampling location prediction model can then perform personalized sampling and CT image reconstruction for patients in any scenario encompassed by the constructed data.

[0109] Compared with traditional CT image reconstruction methods, this application's method firstly constructs a parameterized Gaussian mixture distribution based on prior information from X-ray images. Compared to traditional CT reconstruction algorithms, this allows for personalized sampling design for each patient. Compared to online active CT reconstruction algorithms, pre-sampling in the parameterized Gaussian mixture distribution allows for pre-designed sampling patterns, reducing reconstruction time. Secondly, for the sparse sampling process, this application designs a projection domain completion module based on the Swin Transformer architecture, achieving dual-domain regularization in both the image and projection domains, which helps improve the performance of the reconstruction algorithm. Finally, this application combines CNN and visual Transformer architectures to design an image domain reconstruction module. Compared with the network framework of traditional CT reconstruction methods, this module achieves global feature extraction from the initial FBP image, improving the model's expressive power.

[0110] Example 2

[0111] This invention also discloses a CT image reconstruction system based on X-ray film sampling location prediction, which employs the method described in Example 1. (See appendix) Figure 5 As shown, the system includes a collection module, a sampling location prediction module, a projection domain completion module, and an image domain reconstruction module.

[0112] The collection module is used to acquire X-ray images. X .

[0113] The sampling location prediction module is based on the X-ray image I X The optimal sampling position distribution for subsequent individual CT images is calculated, and sampling is performed based on the optimal sampling position distribution to obtain the sparse projection y. u The sampling location prediction module stores the sampling location prediction model.

[0114] The projection domain completion module can store a trained first deep learning network model, and the projection domain completion module is used to complete the sparse projection y. u Denoising and completion are performed to obtain the projection signal y. eThe first deep learning network model is built upon the Swin-Transformer network structure, which has temporal preservation properties. This model constructs the correlation between multiple projections, thereby addressing the highly sparse projection y. u By performing interpolation and completion, the predicted projection signal y can be obtained. e This module utilizes joint supervision of L1 and SSIM loss functions throughout the model training process to improve the learning ability of edge features in the projection map.

[0115] The image domain reconstruction module stores a second deep learning network, and the image domain reconstruction module is used to convert the sparse projection y u and projection signal y e FBP images were reconstructed using the analytical FBP algorithm respectively. and FBP image Fusing the FBP images and FBP image The reconstructed CT image u is then generated. The second deep learning network is built on a hybrid architecture of Convolutional Neural Network (CNN) and Visual Transformer (STM). It leverages the powerful local feature representation capabilities of CNN and the global feature learning capabilities of the STM architecture to perform high-quality CT image reconstruction, resulting in the final reconstructed CT image u. Throughout the model's training, this module utilizes joint supervision of three loss functions—L1, texture-aware loss, and SSIM—to improve the final image reconstruction performance.

[0116] In one embodiment, the reconstruction system further includes a storage module for storing pre-collected CT and X-ray images. The CT and X-ray images are grouped according to anatomical location and clinical task and stored in the same folder. For CT and X-ray images within the same group, data pairs of CT images and their corresponding X-ray images are created. For cases where only CT images are available, the DeepDRR physical simulation method is used to generate corresponding X-ray images, and then the generated X-ray images are paired with the CT images. The sampling location prediction module, projection domain completion module, and image domain reconstruction module can call the data pairs in the storage module for model training.

[0117] Example 3

[0118] The embodiments of this specification also provide a computer-readable storage medium storing a computer program thereon. When the computer program is executed, it implements any of the CT image reconstruction methods based on X-ray film sampling position prediction provided in Embodiment 1 of this application.

[0119] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0120] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0121] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable those skilled in the art to understand the content of the present invention and implement it. They should not be used to limit the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for CT image reconstruction based on X-ray projection sampling position prediction, characterized in that: Includes the following steps: Acquiring at least one x-ray image ; According to the X-ray film image calculating an optimal sampling position distribution for subsequent sampling of the individual CT images; Sampling is performed based on the optimal sampling location distribution to obtain the sparse projection. ; Load the pre-trained first deep learning network model onto the sparse projection Denoising and completion are performed to obtain the projection signal. ; The sparse projection and projection signal They were reconstructed separately using the analytical algorithm FBP; Load the pre-trained second deep learning network model onto the FBP image and FBP image The images are fused together to form a reconstructed CT image. ; Based on the X-ray image The calculation of the optimal sampling location distribution for subsequent sampling specifically includes: A sampling location prediction model is constructed, which includes an optimizable sampling operator based on a Gaussian mixture distribution. and the sampling operator a priori ; The sampling location prediction model is based on the currently acquired X-ray image. Calculate parameterized distribution ,in X-ray image The encoding, For prior knowledge Trainable parameters; Using Monte Carlo methods in parameterized distributions Sampling is performed on the image to obtain the image from the current X-ray film. The optimal sampling location distribution for pairing.

2. The CT image reconstruction method based on X-ray film sampling position prediction according to claim 1, characterized in that: The sampling location prediction model is as follows: , in and It is an adjustable hyperparameter. These are CT images to be reconstructed. It is a full sampling operator. It is the image prior of u. It is the sparse projection obtained by sampling, and D(*) is the distance function.

3. The CT image reconstruction method based on X-ray film sampling position prediction according to claim 2, characterized in that: The image prior R1 includes the data prior learned through the second deep learning network model and the pixel gradient smoothing prior of the image. ; The prior R2 includes X-ray images. The number of samples, the sampling location, and the distance function d between adjacent sampling locations.

4. The CT image reconstruction method based on X-ray film sampling position prediction according to claim 1, characterized in that: The training process of the first deep learning network model includes: The first deep network structure is constructed based on the Swin-Transformer network structure with time preservation characteristics; Collect a first training dataset, which includes data pairs formed by collected CT images and corresponding X-ray images; The first deep network structure is trained using the first training dataset, and the trained data is saved as the first deep learning network model.

5. The CT image reconstruction method based on X-ray film sampling position prediction according to claim 4, characterized in that: The training process of the first deep learning network model is jointly supervised by L1 and SSIM loss functions.

6. The CT image reconstruction method based on X-ray film sampling position prediction according to claim 1, characterized in that: The training process of the second deep learning network model includes: A second deep network structure is constructed based on a hybrid architecture of convolutional neural network (CNN) and visual Transformer structure; Construct a second training dataset, which includes data pairs consisting of low-quality CT images and high-quality CT images. The low-quality CT images are constructed from M collected X-ray images using the FBP parsing algorithm, and the high-quality CT images are constructed from N collected X-ray images using the FBP parsing algorithm, where N is greater than M. The second deep network structure is trained using the second training dataset, and the trained data is saved as the second deep learning network model.

7. The CT image reconstruction method based on X-ray film sampling position prediction according to claim 6, characterized in that: The training process of the second deep learning network model utilizes three loss functions—L1, texture-aware loss, and SSIM—for joint supervision.

8. A CT image reconstruction system based on X-ray film sampling location prediction, characterized in that: include: The collection module is used to acquire X-ray images. ; The sampling position prediction module, based on the X-ray image... The optimal sampling position distribution for subsequent individual CT images is calculated, and sampling is performed based on the optimal sampling position distribution to obtain sparse projection. ; A projection domain completion module, which stores a trained first deep learning network model, is used to complete the sparse projection. Denoising and completion are performed to obtain the projection signal. ; An image domain reconstruction module, which stores a second deep learning network, is used to reconstruct the sparse projection. and projection signal FBP images were reconstructed using the analytical FBP algorithm respectively. and FBP image fused FBP images and FBP image And form reconstructed CT images. ; Based on the X-ray image The calculation of the optimal sampling location distribution for subsequent sampling specifically includes: A sampling location prediction model is constructed, which includes an optimizable sampling operator based on a Gaussian mixture distribution. and the sampling operator a priori ; The sampling location prediction model is based on the currently acquired X-ray image. Calculate parameterized distribution ,in X-ray image The encoding, For prior knowledge Trainable parameters; Using Monte Carlo methods in parameterized distributions Sampling is performed on the image to obtain the image from the current X-ray film. The optimal sampling location distribution for pairing.

9. A storage medium, characterized in that: The computer-readable storage medium stores instructions that, when executed by a processor, are used to implement the CT image reconstruction method based on X-ray film sampling position prediction as described in any one of claims 1-7.

Images