Sparse view angle CT image reconstruction method based on prompt learning

Through the sparse perspective CT image reconstruction method based on prompt learning, combined with the data fidelity module and the prior module, the Transformer fusion feature is used to solve the problem of insufficient noise and artifact removal capabilities in the existing technology, and the reconstruction performance and inference efficiency are improved .

CN119006629APending Publication Date: 2024-11-22SICHUAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411043705.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing CT image reconstruction methods are insufficient in noise and artifact removal capabilities and are low inference efficiency, especially under sparse perspective conditions.

Method used

Using a sparse perspective CT image reconstruction method based on prompt learning, by constructing training data sets and verification data sets, combining data fidelity modules and prior modules, Transformer is used to fuse image features and prior information to reduce model complexity and improve model Generalization ability and robustness.

Benefits of technology

The detail reconstruction quality and artifact elimination ability of CT images are significantly improved, the model complexity and calculation volume are reduced, and the reconstruction performance under low dose or sparse projection conditions is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006629A_ABST
    Figure CN119006629A_ABST
Patent Text Reader

Abstract

The invention relates to the field of medical image processing, in particular to a sparse view angle CT image reconstruction method based on prompt learning. According to the method, a training and verification data set is constructed by collecting CT images of a normal visual angle and simulating sparse visual angle projection and reconstruction. And secondly, training a model by using a training data set, evaluating the performance of the model through a verification data set, performing hyper-parameter adjustment, inputting a CT image to be reconstructed and projection data into the trained model, and outputting a final reconstructed CT image. Through experimental verification on a public data set, the method obtains an optimal result on image reconstruction evaluation indexes such as PSNR and SSIM, and is excellent in detail reconstruction and artifact elimination. Compared with other existing methods, the reconstruction method provided by the invention not only utilizes prompt learning to extract the degradation prior of the image, improves the detail reconstruction capability and the artifact elimination capability of the model, but also remarkably reduces the complexity of the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a sparse view CT image reconstruction method based on prompt learning. Background Art

[0002] In contemporary medicine, X-ray computed tomography (CT) scanning has become an indispensable imaging examination method for screening, diagnosis and treatment planning. However, the ionizing radiation dose that the human body needs to bear in a single CT scan is much higher than that of an X-ray, which often leads to an increased risk of cancer in patients. To solve this problem, the radiation dose of X-rays can be reduced by reducing the projection angle, but the CT images obtained in this way are usually accompanied by severe stripe artifacts, which seriously affects medical practice.

[0003] The reconstruction method based on CT imaging model stands out among many types of methods due to its reliable theoretical basis and advanced reconstruction performance. It defines an optimization objective function and uses classical mathematical optimization methods to obtain the optimal solution. The objective function consists of two parts: data fidelity term and regularization term. The data fidelity term makes the reconstruction result closer to the observed data. The regularization term provides prior information and restricts the solution space, making it easier to obtain the target solution. The quality of the regularization term has a very important impact on the quality of the reconstruction result. In recent years, the reconstruction methods based on imaging models and deep learning have combined to form a new type of reconstruction method based on deep models, which has played the unique advantages of imaging model-based and deep learning-based methods. However, there are still some defects in this type of method. First, the regularization term currently designed can often only extract the content prior of CT images, and is powerless for image degradation priors such as noise level and sparsity. Adding such prior information to the model can effectively improve the ability to eliminate noise and artifacts. Secondly, although the current methods based on deep models have high reconstruction performance, they are relatively backward in reasoning efficiency because they have a large number of model parameters and high computational complexity. Therefore, there is potential for further improvement in reconstruction algorithms based on deep models. Summary of the invention

[0004] The purpose of the present invention is to overcome the insufficient noise and artifact elimination capability and model reasoning efficiency in the prior art, and to provide a sparse view CT image reconstruction method based on prompt learning.

[0005] In a first aspect, the present invention provides a sparse view CT image reconstruction method based on prompt learning, comprising the following steps:

[0006] S1: Collect multiple groups of sparse view CT images, sparse view projection images and normal view CT images, and construct a training data set and a validation data set. The training data set includes multiple groups of sparse view CT images, sparse view projection images and normal view CT images, and the validation data set includes multiple groups of sparse view CT images and sparse view projection images;

[0007] S2: Construct an iterative reconstruction model for sparse view CT images based on cue learning;

[0008] S3: Use the image groups in the training dataset to train the sparse view CT image iterative reconstruction model and use the validation dataset to evaluate the model and adjust the hyperparameters to obtain a trained image reconstruction model;

[0009] S4: Inputting the CT image to be reconstructed and the projection image data into the trained image reconstruction model to obtain a final reconstructed CT image.

[0010] The present invention utilizes the characteristics of prompt learning to extract the degradation prior of sparse view CT images and enhance the model's artifact elimination capability. Meanwhile, the model complexity is significantly reduced through the efficient Transformer fusion of image features and prior information.

[0011] Preferably, in order to construct a training set and a validation set of sparse view CT images, each collected CT image needs to be preprocessed, and then the number of view angles of the projection operator is set to the number of sparse view angles, and projection is performed to obtain a sparse view projection image; then the FBP algorithm is applied to the sparse view projection image to obtain a sparse view CT image; finally, the corresponding images are combined to construct a training set and a validation set.

[0012] By specifically setting the number of perspectives of the projection operator to the number of sparse perspectives and applying the FBP algorithm, the sparse perspective CT images obtained can directly reflect the imaging characteristics under actual sparse perspective conditions. Such a data set is closer to the actual application scenario, making the trained model more targeted and practical. Since both the training set and the validation set are constructed based on sparse perspective projection images, the model can learn how to reconstruct high-quality CT images under sparse perspectives during the training process. This enables the model to have better generalization and robustness when processing sparse perspective CT data in the real world. At the same time, through preprocessing and projection operations, the original CT images can be quickly converted into a data set suitable for sparse perspective reconstruction. This not only improves the efficiency of data processing, but also makes the model training and optimization process more efficient and stable.

[0013] In a second aspect, the present invention provides a method for constructing an iterative reconstruction model of sparse view CT images, wherein the iterative reconstruction model of sparse view CT images based on prompt learning is composed of two iterative modules, a data fidelity module and a priori module based on prompt learning, which are cascaded. The data fidelity module ensures the consistency of the reconstruction result and the observed data through projection and back-projection operators; the priori module extracts the degradation prior in the image using prompt learning, and the iterative reconstruction model of the CT image reconstructs the details of the CT image according to the degradation prior.

[0014] Through prompt learning, the prior module can learn these degradation priors and apply them to the iterative reconstruction process of CT images, thereby improving the details and clarity of the reconstructed images. The sparse view CT image iterative reconstruction model based on prompt learning is composed of two iterative modules, the data fidelity module and the prior module based on prompt learning, which realizes high-quality CT image reconstruction under sparse view conditions. Among them, the data fidelity module ensures the consistency of the reconstruction results with the observed data, while the prior module improves the detailed reconstruction of the image through prompt learning technology. The data fidelity module and the prior module based on prompt learning work together, so that the model can still obtain high-quality CT images under the conditions of low dose or sparse projection data.

[0015] Preferably, the method proposes a formula for guiding the construction of iterative modules: According to the first half of the formula A data fidelity module can be constructed, and according to the second half of the formula Φ t (x t ) can construct a priori modules. The existence of the data fidelity module enables the reconstruction results to maintain consistency with the observed data as much as possible, making the results more realistic.

[0016] Furthermore, in the regularization module, the intrinsic prior of the CT image and the degradation prior of the sparse view CT image are extracted according to the following formula and used to guide the reconstruction process:

[0017] z t =Φ t (x t )

[0018] β t =Conv 5×5 (Conv 5×5 (x t ))

[0019]

[0020] γ t =PIB(β t ,p t)

[0021] z t =Conv 5×5 (γ t )

[0022] In the above formula, the input image x t After two convolutional layers, the intrinsic prior of the CT image is extracted. Then, the PGB (prompt generation module) is used to extract the degradation prior from the image and generate the prompt feature p t , then prompt feature p t With the intermediate feature β t Entering PIB prompts the fusion module to fully interact, and finally obtains the CT image after this iteration through convolution. t It is the prompt context information generated by the previous iterative module, which is used to gradually provide the model with degraded prior information. Specifically, it is generated by this formula:

[0023] c t =γ t-1 =PIB(β t-1 ,p t-1 )

[0024] In the above formula, PGB follows the following formula:

[0025] w i =Softmax(Linear(GAP(β t )))

[0026]

[0027] In the above formula, PIB uses an efficient fusion Transformer, which uses channel cross attention to fuse the prompt feature p t and image features β t , the cross attention can effectively integrate the prior information into the image features, while the channel attention significantly reduces the model complexity while enabling the model to have global modeling capabilities. The following is the formula followed by PIB:

[0028]

[0029] Where LN represents layer normalization, GDFN represents a special forward network, and CCA represents channel cross attention. Channel cross attention is the core part of the efficient fusion Transformer, which receives the normalized prompt feature f p and image features f β As input, f β and f p After Conv 1×1 and DConv3×3 Get the image Q, K, V and the prompt Q', K', V', and then merge Q, K, V and Q', K', V' to generate Finally, channel attention is performed, because The prompts Q', K', V' are integrated in the channel attention, so the channel attention integrates the prompt information into the image features in a cross-form. The above overall CCA process satisfies the following formula:

[0030] Q,K,V=DConv 3×3 (Conv 1×1 (f β ))

[0031] Q',K',V'=DConv 3×3 (Conv 1×1 (f p ))

[0032]

[0033] A further technical solution is that in S3, the model is trained in a supervised manner, and the loss function used is the L2 loss function. The L2 loss function is usually more robust to noise and small disturbances. During the generation process, if the generated output changes slightly, the L2 loss will not cause a large gradient change, which helps to make the training more stable. In image generation, it can help the generated image to be closer to the target image.

[0034] The channel attention mechanism emphasizes important features by calculating the correlation between channels, thereby enhancing the model's ability to capture key information such as details and structures in the image and improving the quality of reconstruction. At the same time, it can help reduce the amount of data that the model needs to process, thereby reducing the complexity of the model. When the prompt information is integrated into the features in a cross-form, the model can use this information more efficiently, further enhancing the accuracy of reconstruction and reducing the model's computational and storage requirements.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention provides a sparse view CT image reconstruction method based on prompt learning. Through the prompt learning technology, trainable parameters are used to interact with the input sparse view CT image features to extract the noise distribution of the sparse view CT image, which can effectively limit the reconstructed image solution space in a regularized manner under the model-based method using the degradation prior information, so that the reconstructed image retains more details while removing more artifacts;

[0037] 2. The present invention provides a progressive prompting strategy, which realizes feature interaction enhancement between adjacent iterative blocks by passing prior information layer by layer, and also significantly enhances the prompting effect. The prior information of noise distribution is more effectively utilized, thereby enhancing the overall performance of the model;

[0038] 3. The present invention provides a Transformer module that is different from the one used in the existing methods. By operating in a cross-attention manner and capturing dependencies at the feature dimension level, it is possible to reduce the number of module parameters and the amount of computation. In the method of the present invention, prior information can be more effectively integrated into image features, thereby enhancing the model's ability to remove noise artifacts and significantly improving model efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a principle framework diagram in Example 1 of the present invention.

[0040] Figure 2 It is a structural diagram of the prompt generation module PGB and the prompt fusion module PIB proposed in Example 1 of the present invention.

[0041] Figure 3 Comparison of evaluation indicators of reconstruction results in the test case conducted by the present invention and other methods.

[0042] Figure 4 These are the 32-view sparse-view CT images and the reconstruction results of each method in the test example conducted by the present invention.

[0043] Figure 5 These are the 64-view sparse-view CT images and the reconstruction results of each method in the test example conducted by the present invention.

[0044] Reference numerals:

[0045] (a)-,Detailed structure diagram of the prompt generation module PGB;

[0046] (b)-,Detailed structural diagram of the prompt fusion module. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0048] Example 1

[0049] A sparse view CT image reconstruction method based on hint learning, the flowchart is as follows Figure 1 As shown, the following steps are included:

[0050] S1: Collect and construct training data sets and validation data sets. The training data sets are multiple sets of sparse view CT images, sparse view projection images, and normal view CT images. The validation data sets are multiple sets of sparse view CT images and sparse view projection images.

[0051] S2: Construct an iterative reconstruction model of sparse view CT images based on prompt learning. The model is based on prompt learning and iterative reconstruction, and is composed of a cascade of multiple iterative modules; the iterative module is composed of a data fidelity module and a priori module based on prompt learning; the data fidelity module uses projection and back-projection operators to ensure the consistency of the reconstruction result and the observed data, and the priori module uses the characteristics of prompt learning to learn the degradation features of sparse view CT images to assist the reconstruction process. S2 includes the following steps:

[0052] S21, collect and construct a training data set and a verification data set, the training data set is multiple groups of sparse view CT images, sparse view projection images and normal view CT images, and the verification data set is multiple groups of sparse view CT images and sparse view projection images.

[0053] S22, constructing an iterative reconstruction model for sparse view CT images based on prompt learning, wherein the model is based on prompt learning and iterative reconstruction, and is composed of a plurality of cascaded iterative modules; the iterative module is composed of a data fidelity module and a priori module based on prompt learning; the data fidelity module ensures the consistency between the reconstruction result and the observed data through projection and back-projection operators, and the priori module uses the characteristics of prompt learning to learn the degradation features of the sparse view CT images to assist the reconstruction process;

[0054] S3: The model is trained in a supervised manner, and the loss function used is the L2 loss function to ensure that the error between the model output and the real CT image is minimized;

[0055] S4: Input the CT image to be reconstructed and the projection image data to obtain the final reconstructed CT image.

[0056] In an optional implementation, step S1 may further optimize the selection and processing of the data set to improve the generalization ability and robustness of the model. For example, the data set may be expanded by increasing the diversity and quantity of the data set, or by using data enhancement techniques;

[0057] In an optional implementation, step S2 can further adjust the model structure and parameters. For example, the number of iterative modules can be increased to improve the reconstruction accuracy, or the number of convolutional layers, activation functions and other parameters can be adjusted to optimize the performance of the model. In addition, other types of prior modules or regularization methods can be explored to further improve the reconstruction effect. According to the above description, the number of iterative modules is set to 16, and the network structure of each iterative module is as follows: Figure 1As shown, and conforms to the following formula:

[0058]

[0059] Among them, x t+1 is the result after the t+1th iteration, x t , y are the CT image and sparse view projection image input by the t-th iteration module, Φ t represents the prior module of the tth iteration; according to the first half of the formula A data fidelity module can be constructed, and according to the second half of the formula Φ t (x t ) can construct a priori module. In the data fidelity module, α t represents the learnable step size parameter in the tth iteration, Represents the back-projection operator, where the FBP operator with gradient back-propagation is used, and A represents the projection operator.

[0060] exist Figure 1 In the proposed method, the prompt module consists of two important submodules: the prompt generation module (PGB) and the prompt fusion module (PIB). The sparse view CT image iterative reconstruction model expands the gradient descent algorithm into multiple iterative modules. Each iterative module converts the sparse view CTx t , sparse view projection y and hint context c t As input, to gradually generate the reconstruction result X * For the first iteration module, the input is only the initial FBP result x 0 and sparse view projection y. Inside the iterative module, x t and y pass through the data fidelity module, replacing A with the inverse transform of A (e.g., FBP) T , better performance can be achieved in iterative reconstruction. t The image content and noise distribution information are extracted through the prior module, which is achieved by injecting the prompt block into the shallow CNN. t It is the prompt context generated by the previous iterative module, which is used to progressively provide the model with degradation prior information. Specifically, it is the fusion feature γ output by PIB in the previous iteration (t-1 iteration). t-1 :

[0061] c t =γ t-1 =PIB(β t-1 ,p t-1 )

[0062] like Figure 1As shown in the lower half of the middle branch structure, in the regularization module, the intrinsic prior of the CT image and the degradation prior of the sparse view CT image are extracted according to the following formula and used to guide the reconstruction process:

[0063] z t =Φ t (x t )

[0064] β t =Conv 5×5 (Conv 5×5 (x t ))

[0065]

[0066] γ t =PIB(β t ,p t )

[0067] z t =Conv 5×5 (γ t )

[0068] In the above formula, Φ t represents the prior module of the tth iteration, x t represents the input of the prior module, z t Represents the output of the prior module. In the prior module, the image x is first input t After two convolutional layers Conv 5×5 Extract the intrinsic prior of the CT image and obtain the intermediate feature β containing the intrinsic prior information t , followed by β t The degradation prior of the CT image is extracted through PGB, and the feature p containing the degradation prior information is generated t , then p t With the intermediate feature β t Enter PIB for full fusion to generate features that are rich in both intrinsic priors and degraded priors γ t , and finally γ t After another convolution, the output image z after this iteration is obtained t .

[0069] In the above formula, Represents a set of prompt components, which is a set of learnable parameters, N c Indicates the number of prompt members. The size of each prompt member is consistent with the input feature size of PGB (H×W×C). The input feature β of PGB t 、Prompt context ct Interact together to generate hint features p that can provide degradation priors t .

[0070] PGB focuses on perceiving image content and noise priors, which can be used to pursue better artifact removal effects. Figure 2 As shown in (a), given the input feature β of PGB t , first use To initialize the prompt feature, and then let With input feature β t Interact to generate adaptive hint features p t .

[0071] Specifically, first, β t After the GAP (Global Average Pooling) operation, a vector is obtained, and the dimension of the vector is β t The dimension C of the feature is then transformed into The number N c , and then use Softmax to get the weight w of each prompt member i , The weights are summed to obtain a feature with a dimension of H×W×C. Finally, the feature is combined with the prompt context c t Concat is connected according to the channel dimension and then passes through a convolution Conv 3×3 Get the output p of PGB t .

[0072] The process described above follows the following formula:

[0073] w i =Softmax(Linear(GAP(β t )))

[0074]

[0075] Among them, w i =Softmax(w) represents the i-th weight component of the softmax function output, GAP is the global average pooling operation, Represents the i-th member of the prompt member.

[0076] The main purpose of PIB is to enrich spatial features with cues learned from local and global perspectives to achieve guided reconstruction. The detailed process of PIB is as follows Figure 2 In the above formula, the main function of PIB is to transform the enhanced features p generated by PGB, which are rich in degradation prior information, into tWith image feature β t Fully fused to guide the artifact removal of CT images. PIB uses an efficient fusion Transformer, which uses channel cross attention CCA (Channel Cross Attention) to fuse the prompt feature p t and image features β t , the cross attention can effectively integrate the prior information into the image features, while the channel attention can significantly reduce the model complexity while enabling the model to have global modeling capabilities. Specifically, first prompt feature p t and image features β t All of them are normalized by layer normalization LN to obtain the normalized prompt feature f p and image features f β , then f p and f β Enter the channel cross attention CCA for full feature fusion, and then add residual connection to obtain intermediate features Then the output γ of PIB is obtained through layer normalization LN, forward network GDFN and residual connection t , the cross attention can effectively integrate the prior information into the image features, while the channel attention significantly reduces the model complexity while enabling the model to have global modeling capabilities. The following is the formula followed by PIB:

[0077]

[0078]

[0079] Among them, LN represents layer normalization, GDFN represents a special forward network, and CCA represents channel cross attention. Channel cross attention is the core part of the efficient fusion Transformer, which receives the normalized prompt feature f p and image features f β As input, f β and f p After Conv 1×1 and DConv 3×3 Get the image Q, K, V and the prompt Q', K', V' (where DConv 3×3 Denotes a depthwise convolution with a kernel size of 3×3, and then Q, K, V and Q', K', V' are fused to generate Finally, perform two channel attention and And the results are combined to get the output f' of CCA β .because The prompts Q', K', V' are integrated in the channel attention, so the channel attention integrates the prompt information into the image features in a cross-form. The above overall CCA process satisfies the following formula:

[0080] Q,K,V=DConv 3×3 (Conv 1×1 (f β ))

[0081] Q',K',V'=DConv 3×3 (Conv 1×1 (f p ))

[0082]

[0083] Among them, CCA represents channel cross attention, LN represents layer normalization, and GDFN represents a special forward network.

[0084] Test Example 1

[0085] In order to verify the effectiveness and superiority of the reconstruction method proposed in Example 1, this test example selected five other existing reconstruction methods for comparative experiments. Each method was trained according to the experimental settings of the present invention to ensure the fairness of the comparison of each method. The five comparative methods are:

[0086] Method 1, FBPConvNet proposed by Jin et al., Reference Jin, KH, McCann, MT, Froustey, E., Unser, M.: Deep convolutional neural network for inverse problems in imaging. IEEE transactions on image processing 26(9), 4509–4522 (2017)

[0087] Method 2, DuDoTrans proposed by Wang et al., Reference Wang, C., Shang, K., Zhang, H., Li, Q., Zhou, SK: Dudotrans: Dual-domain transformer for sparse-view count construction. In: International Workshop on Machine Learning for Medical Image Reconstruction. pp. 84–94. Springer (2022)

[0088] Method 3, LEARN proposed by Chen et al., Reference Chen, H., Zhang, Y., Chen, Y., Zhang, J., Zhang, W., Sun, H., Lv, Y., Liao, P., Zhou, J., Wang, G.: Learn: Learned experts' assessment-based reconstruction network for sparse-data ct. IEEE transactions on medical imaging 37(6), 1333–1347 (2018)

[0089] Method 4, LEARN++ proposed by Zhang et al., Reference Zhang, Y., Chen, H., Xia, W., Chen, Y., Liu, B., Liu, Y., Sun, H., Zhou, J.: Learn++: recurrent dual-domain reconstruction network for compressed sensing ct. IEEE Transactions on Radiation and Plasma Medical Sciences 7(2), 132–142(2022)

[0090] Method 5, RegFormer proposed by Xia et al., Reference Xia, W., Yang, Z., Lu, Z., Wang, Z., Zhang, Y.: Regformer: A local-nonlocal regularization-based model for sparse-view ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences (2023)

[0091] The training and testing data sets used in this invention are all Mayo data sets which are widely used in the field of CT image reconstruction. Figure 3As shown. SVCT is a sparse view CT image with 64 views, method 1 is FBPConvNet, method 2 is DuDoTrans, method 3 is LEARN, method 4 is LEARN++, and method 5 is RegFormer and LEARN. By analyzing the results in Table 1, it can be seen that the method proposed in the present invention has achieved the best in both PSNR and SSIM image reconstruction evaluation indicators, and also has the least model parameters and calculation amount, which shows that the method proposed in the present invention not only improves the model performance but also significantly reduces the model complexity.

[0092] This test case randomly selects two representative sparse view CT images from the data set for instance testing. Figure 4 This is the test result comparison of 32-view sparse view CT images. Figure 5 This is the test result comparison of 64-view sparse view CT images. Figure 4 and Figure 5 The yellow arrow in the figure indicates that the method proposed in the present invention has better detail reconstruction and artifact elimination capabilities, which is due to the prior module based on prompt learning and the progressive prompt strategy proposed in the present invention. In general, by comparison with the recent sparse perspective CT image reconstruction method, the method proposed in the present invention not only achieves the best results in quantitative evaluation indicators (PSNR and SSIM), but also outperforms other methods in visual perception. In addition, the model complexity of the method proposed in the present invention is lower than that of other methods based on deep models, and it has obvious advantages in reasoning efficiency, which can be better applied to medical practice. Therefore, the present invention is an effective and efficient sparse perspective CT image reconstruction method.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A sparse view CT image reconstruction method based on hint learning, It is characterized in that The following steps are involved: S1: Construct a training dataset and a validation dataset. The training dataset includes multiple sets of sparse view CT images, sparse view projection images, and normal view CT images. The validation dataset includes multiple sets of sparse view CT images and sparse view projection images. S2: Construct an iterative reconstruction model for sparse view CT images based on cue learning; S3: Use the training data set to train the sparse view CT image iterative reconstruction model, use the validation data set to evaluate the model, and adjust the hyperparameters to obtain the trained sparse view CT image iterative reconstruction model; S4: Inputting the CT image to be reconstructed and the projection image data into the trained sparse view CT image iterative reconstruction model to obtain a final reconstructed CT image.

2. The method for sparse view CT image reconstruction based on prompt learning according to claim 1, It is characterized in that In S3, the model is trained by supervised learning, and the loss function used is L2 loss function.

3. The method for sparse view CT image reconstruction based on prompt learning according to claim 1, It is characterized in that The S1 also includes preprocessing the CT images collected by the training data set and the verification data set, which specifically includes the following steps: S31: setting the number of viewing angles of the projection operator to the number of sparse viewing angles, and performing projection to obtain a sparse viewing angle projection image; S32: applying the FBP algorithm to the sparse view projection image to obtain a sparse view CT image; S33: Combine sparse view CT images to construct training sets and validation sets.

4. The method for sparse view CT image reconstruction based on prompt learning according to claim 1, It is characterized in that The sparse view CT image iterative reconstruction model of prompt learning described in S2 is composed of a cascade of multiple iterative modules, and the iterative modules include a data fidelity module and a priori module based on prompt learning; The data fidelity module ensures the consistency of the reconstruction result and the observed data through projection and back-projection operators; The prior module uses prompt learning to extract the degradation prior information in the image, and the CT image iterative reconstruction model reconstructs the CT image details according to the degradation prior information.

5. The method for sparse view CT image reconstruction based on prompt learning according to claim 4, It is characterized in that The prior module includes a prompt generation module and a prompt fusion module.

6. The method for sparse view CT image reconstruction based on prompt learning according to claim 5, It is characterized in that The process of extracting the degradation prior information in the image specifically comprises the following steps: S61: The input image is passed through two convolutional layers to extract the intrinsic prior information of the CT image to obtain intermediate features, which contain intrinsic prior information. S62: The intermediate features are extracted from the CT image by the prompt generation module to obtain a degradation prior and generate a prompt feature that can provide a degradation prior. The prompt feature and the intermediate features are then fused in the prompt fusion module to generate a feature including an intrinsic prior and a degradation prior. S63: The features are convolved to obtain the CT image after this iteration.

7. The method for sparse view CT image reconstruction based on prompt learning according to claim 6, It is characterized in that The prompt feature is generated by a group of prompt members, and the prompt members include a group of learnable parameters. The learnable parameters interact with the input features of the prompt generation module and the prompt context. The prompt context is iteratively generated by the previous layer of the iterative module, that is, the fused features output by the prompt fusion module, which are used to progressively provide degradation prior information for the model.

8. The method for sparse view CT image reconstruction based on prompt learning according to claim 6, It is characterized in that The prompt fusion module adopts a fusion Transformer, which includes a channel cross-attention mechanism module. The Transformer uses the channel cross-attention mechanism to fuse prompt features and image features, and fuses prior information into image features.

9. The method for sparse view CT image reconstruction based on prompt learning according to claim 8, It is characterized in that The channel cross attention mechanism module receives the normalized prompt features and image features as input, and the prompt features and image features are processed by Conv 1×1 The image Q, K, V and the prompt Q', K', V' are obtained by depthwise separable convolution with a convolution kernel size of 3×3, and then the image Q, K, V and the prompt Q', K', V' are fused to generate because The prompts Q', K', V' are integrated in the channel, so the channel attention integrates the prompt information into the image features in a cross-form.

10. An apparatus for the method of claim 1, comprising a CT device, at least one processor, and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, the instructions are executed by the at least one processor so that the at least one processor can execute any of the methods of claims 1-9 and output a sparse view CT reconstruction scheme based on prompt learning to the CT device.