A distributed CT intelligent imaging method and a distributed CT intelligent imaging system

Abstract

The application discloses a distributed CT intelligent imaging method and system, wherein the distributed CT intelligent imaging method is used for reconstructing a final CT image through four steps, specifically, original CT measurement image data of a plurality of local data centers and a CT low-dose image data-high-dose image data pair of a cloud center end are acquired, high-frequency features of the data set are extracted, then interactive training is carried out by adopting a federal learning strategy, finally, optimized local network parameters are obtained, and finally, the corresponding original CT measurement image data is reconstructed based on the local network parameters to obtain the final CT image. The application can fuse the federal learning and the unsupervised learning method, can protect the privacy of patients and can also obtain high-quality denoising reconstructed images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and computed tomography, and particularly relates to a distributed CT intelligent imaging method and a distributed CT intelligent imaging system. BACKGROUND

[0002] X-ray computed tomography (CT) technology has been widely used in the field of medical diagnosis. Excessive CT dose can induce leukemia, cancer and other genetic diseases. Clinically, the method of reducing the radiation dose by reducing the tube current is commonly used, but the reduction of the radiation dose will lead to the decline of the CT image quality. Therefore, it has important clinical significance to study the low-dose CT intelligent imaging.

[0003] With the development of deep learning, some supervised learning methods can automatically extract useful features in image data, so as to more easily adapt to new data evolution training model, and significantly improve the image quality. However, supervised learning needs to train low-dose CT images and normal dose images matched at the pixel level. However, in practical applications, it is impossible for patients to simultaneously collect low-dose and normal-dose CT images, which leads to the difficulty in obtaining paired label data sets.

[0004] Federal learning is a distributed machine learning paradigm, which allows multiple edge devices (or clients) to collaboratively train a shared global network model. However, the current mainstream work is based on an unrealistic assumption that the training data of the local client has a true label. In addition, in most cases, it is a time-consuming, expensive and complex work to combine a properly labeled data set for a given federated learning task. Therefore, it is challenging to train a high-quality global network model in the absence of labeled data in real scenarios.

[0005] Therefore, in view of the deficiencies of the prior art, it is very necessary to provide a distributed CT intelligent imaging method and a distributed CT intelligent imaging system to solve the deficiencies of the prior art. SUMMARY

[0006] One of the purposes of the present application is to provide a distributed CT intelligent imaging method to avoid the deficiencies of the prior art. The distributed CT intelligent imaging method can combine federal learning with unsupervised learning methods, protect patient privacy, and obtain high-quality denoising reconstructed images.

[0007] The above purpose of the present application is achieved by the following technical measures:

[0008] The present application provides a distributed CT intelligent imaging method, comprising the following steps:

[0009] Step (1), obtaining a local data set composed of a plurality of original CT measurement image data of local data centers, and obtaining a cloud data set composed of a plurality of pairs of standardized CT low-dose image data-high-dose image data pairs of cloud centers;

[0010] Step (2), extracting high-frequency features of the original data set to construct a CT data high-frequency information library shared by the cloud center;

[0011] Step (3), inputting the CT data high-frequency information library obtained in step (2) into the cloud basic network to train a global network model and global network parameters, distributing the global network parameters to all local data centers, inputting the global network parameters and the original CT measurement image data corresponding to step (1) into the local basic network to train a local network model and local network parameters, aggregating the local network parameters of all local data centers, and then using a federated learning strategy to interactively train to obtain optimized local network parameters;

[0012] Step (4), the local data center reconstructs the final CT image according to the optimized local network parameters obtained in step (3) and the original CT measurement image data corresponding to step (1).

[0013] In the step (2), the CT data high-frequency information library Wherein K is the number of local data centers, and the high-frequency feature information set B of the kth local data center k represented by formula (I):

[0014]

[0015] Wherein, is a high-frequency feature extraction operation, is low-dose CT image data in the original CT measurement image data, M k is the total amount of high-frequency information corresponding to the original CT measurement image data in the kth local data center, and i is the serial number of the ith original CT measurement image data.

[0016] Preferably, the above step (3) specifically comprises:

[0017] Step (3.1), inputting the CT data high-frequency information library obtained in step (2) into the cloud basic network to train a global network model and global network parameters;

[0018] Step (3.2), the global network parameters are respectively distributed to all local data centers, and each local data center inputs the global network parameters and the original CT measurement image data corresponding obtained in step (1) into the local basic network to perform unsupervised learning training to obtain a local network model and local network parameters corresponding to each local data center;

[0019] Step (3.3), the local network parameters of each local data center are aggregated to obtain aggregated model parameters, and then the aggregated model parameters are uploaded to the cloud, and the cloud center inputs the aggregated model parameters into the cloud basic network to perform updating to obtain an updated global network model and updated global network parameters, and the updated global network parameters are distributed to all local data centers;

[0020] Step (3.4), each local data center respectively performs unsupervised learning training according to the updated global network parameters and the corresponding original CT measurement image data, and finally obtains an updated local imaging model and updated local network parameters by iteratively updating the gradient parameters of the local loss function.

[0021] Step (3.5), the state of the loss function of the local network model of each local data center is judged, when the loss functions of all local network models meet the end condition, the local network parameters of the current local network model are defined as the optimized local network parameters, and step (4) is entered; otherwise, step (3.3) is returned.

[0022] Firstly, the local loss function is the loss function , the loss function is represented by formula (II):

[0023]

[0024] wherein, is the loss function of the local network model of the kth local data center, is the local network model of the kth local data center, is the normal dose image generated by the local network model of the kth local data center, and the loss function The gradient parameters of the loss function

[0025]

[0026] wherein, is the local network model parameter of the kth local data center before updating, is the updated local network model parameter of the kth local data center, and λ is the learning rate, is the loss function of the local network model of the kth local data center The gradient.

[0027] In the second preferred embodiment, the aforementioned local loss function is a loss function. At that time, loss function Equation (Ⅳ) represents:

[0028]

[0029] in, Let be the loss function of the local network model for the k-th local data center. For the local network model of the k-th local data center, The loss function is used to generate the normal dose image for the k-th local data center network model. The gradient parameter is represented by equation (V):

[0030]

[0031] in, The parameters of the local network model before the update of the k-th local data center are: Let be the updated local network model parameters for the k-th local data center, and λ be the learning rate. The loss function for the local network model of the k-th local data center The gradient.

[0032] The third preferred option is that the aforementioned local loss function is a loss function. and loss function The steps include:

[0033] Step (a) Using the loss function Update the loss function. Equation (VI) represents:

[0034]

[0035] in, Let be the loss function of the local network model for the k-th local data center. For a local network model with k local data centers, The loss function is used to generate the normal dose image for the k-th local data center network model. The gradient parameter is represented by equation (Ⅶ):

[0036]

[0037] in, The parameters of the local network model before the update of the k-th local data center are: is the updated local network model parameter of the kth local data center, and λ is the learning rate, is the loss function of the local network model of the kth local data center is the gradient of the loss function

[0038] Step (b), generating new image data using the high-frequency information provided by the global shared data distribution information library, is represented by formula (VIII):

[0039]

[0040] wherein, is the inverse transform of the high-frequency feature extraction operation is the inverse transform of the high-frequency feature extraction operation is the generated new image data; i is the sequence number of the ith image data.

[0041] Step (c), updating the new image data generated in step (b) through the loss function , wherein the loss function is represented by formula (IX):

[0042]

[0043] wherein, is the loss function of the local network model of the kth local data center, and the gradient parameter of the loss function is represented by formula (X):

[0044]

[0045] wherein, is the loss function of the local network model of the kth local data center is the gradient of the loss function.

[0046] In the step (3.3), the aggregated model parameter is obtained by formula (XI):

[0047]

[0048] wherein, H agg is the aggregated model parameter, N k is the number of training data of the local network model of the kth local data center, is the local network parameter of the kth local data center.

[0049] Preferably, the above local data set is represented by formula (XII):

[0050]

[0051] Preferably, the above-mentioned paired standardized low-dose CT image data-high-dose image data pair is a low-dose CT image reconstructed under low-dose scanning conditions-high-quality CT image data reconstructed under high-dose scanning conditions.

[0052] Preferably, the above-mentioned high-frequency feature extraction operation is a Fourier transform or a wavelet transform.

[0053] Preferably, the above-mentioned training end condition is convergence or reaching a maximum number of training times.

[0054] Preferably, the above-mentioned high-frequency feature is a noise distribution characteristic.

[0055] Another object of the present application is to provide a distributed CT intelligent imaging system to avoid the shortcomings of the prior art. The distributed CT intelligent imaging system can combine federated learning with unsupervised learning methods, protect patient privacy, and obtain high-quality denoised reconstructed images.

[0056] The above-mentioned objects of the present application are achieved by the following technical measures:

[0057] A distributed CT intelligent imaging system is provided, which adopts the above-mentioned distributed CT intelligent imaging method.

[0058] The distributed CT intelligent imaging method and system of the application, wherein the distributed CT intelligent imaging method comprises the following steps: step (1), obtaining a local data set composed of original CT measurement image data of a plurality of local data centers, and obtaining a cloud data set composed of a cloud center end matched standardized CT low-dose image data-high-dose image data pair; step (2), extracting high-frequency features of the original data set, and constructing a CT data high-frequency information library shared by the cloud center end; step (3), inputting the CT data high-frequency information library obtained in step (2) into a cloud basic network for training to obtain a global network model and global network parameters, distributing the global network parameters to all local data centers, inputting the global network parameters and the original CT measurement image data corresponding to step (1) into a local basic network for unsupervised training to obtain a local network model and local network parameters, aggregating the local network parameters of all local data centers, and then using a federated learning strategy for interactive training to finally obtain optimized local network parameters; and step (4), reconstructing the final CT image according to the optimized local network parameters obtained in step (3) and the original CT measurement image data corresponding to step (1) by the local data center. The specific beneficial effects of the application are as follows: first, in terms of data sets, due to patient privacy protection and data barriers of various local image centers, a large amount of modern medical CT image data cannot be fully utilized, the federated learning used in the application can achieve the purpose of "safely utilizing data of each local image center without transferring original data", and solving data security and alleviating data barriers of each local image center is one of the beneficial effects; second, in terms of model performance, under the condition of more data, federated learning can improve model performance by continuously iterating model parameters, which is superior to the model performance of most deep learning; third, in terms of medical diagnosis and treatment, the application can actively promote the "graded diagnosis and treatment" advocated by the country, and under the condition of safe sharing of data of each local image center, the diagnosis and treatment level of medical institutions at all levels is improved to a certain extent. The application can combine federated learning with unsupervised learning method, protect patient privacy, and obtain high-quality denoising reconstructed images. BRIEF DESCRIPTION OF DRAWINGS

[0059] The application is further described with reference to the accompanying drawings, but the content of the drawings does not constitute any limitation on the application.

[0060] Figure 1 The distributed CT intelligent imaging method flowchart of the application.

[0061] Figure 2 The reconstruction result of example 2. DETAILED DESCRIPTION

[0062] The technical solutions of the application are further described in combination with the following examples.

[0063] Embodiment 1

[0064] A distributed CT intelligent imaging method, as shown in the figure, comprises the following steps: Figure 1

[0065] Step (1), obtaining a local data set composed of original CT measurement image data of a plurality of local data centers, and obtaining a cloud data set composed of cloud center end pairing standardized CT low-dose image data-high-dose image data pairs; wherein the local data set represented by formula (XII):

[0066]

[0067] Step (2), extracting high-frequency features of the original data set to construct a CT data high-frequency information library shared by the cloud center end;

[0068] Step (3), inputting the CT data high-frequency information library obtained in step (2) into the cloud basic network for training to obtain a global network model and global network parameters, distributing the global network parameters to all local data centers, inputting the global network parameters and the original CT measurement image data corresponding to step (1) into the local basic network for unsupervised training to obtain a local network model and local network parameters, aggregating the local network parameters of all local data centers, and then using a federated learning strategy for interactive training to finally obtain optimized local network parameters;

[0069] Step (4), the local data center reconstructs the final CT image according to the optimized local network parameters obtained in step (3) and the original CT measurement image data corresponding to step (1).

[0070] It should be noted that the cloud basic network and the local basic network of the present application both use Resnet network.

[0071] In the step (2), the CT data high-frequency information library wherein K is the number of local data centers, and the high-frequency feature information set B of the kth local data center k represented by formula (I):

[0072]

[0073] wherein, is a high-frequency feature extraction operation, is low-dose CT image data in the original CT measurement image data, M k ​The total amount of high-frequency information contributed to the corresponding original CT measurement image data in the kth local data center is i, and the i th original CT measurement image data sequence number. The high-frequency feature of the application is the noise distribution characteristic.

[0074] It should be noted that the application extracts the high-frequency features in the original CT measurement image data, and then shares these high-frequency features into the CT data high-frequency information library, thereby realizing the encryption of the original CT measurement image data. Moreover It can be understood that when the original CT measurement image data only includes low-dose CT image data, i.e. The low-dose CT image data; when the original CT measurement image data includes low-dose CT image data and normal-dose data, Then only low-dose CT image data.

[0075] The step (3) of the application specifically comprises:

[0076] Step (3.1), the CT data high-frequency information library obtained according to step (2) is input into the cloud basic network to obtain a global network model and global network parameters;

[0077] Step (3.2), the global network parameters are respectively distributed to all local data centers, and each local data center inputs the global network parameters and the corresponding original CT measurement image data obtained in step (1) into the local basic network to perform unsupervised learning training to obtain a local network model and local network parameters corresponding to each local data center;

[0078] Step (3.3), the local network parameters of each local data center are aggregated to obtain aggregated model parameters, and then the aggregated model parameters are uploaded to the cloud, and the cloud center inputs the aggregated model parameters into the cloud basic network for updating to obtain an updated global network model and updated global network parameters, and the updated global network parameters are distributed to all local data centers;

[0079] Step (3.4), each local data center respectively performs unsupervised learning training according to the updated global network parameters and the corresponding original CT measurement image data, and finally obtains updated local imaging models and updated local network parameters by iteratively updating the gradient parameters of the local loss function;

[0080] Step (3.5), the state of the loss function of the local network model of each local data center is judged, and when the loss functions of all local network models meet the end condition, the local network parameters of the current local network model are defined as the optimized local network parameters, and step (4) is entered; otherwise, return to step (3.3).

[0081] In the step (3.3), the polymerization model parameter is obtained by formula (XI):

[0082]

[0083] wherein, H agg is a polymerization model parameter, N k is the number of training data of the local network model of the kth local data center, is the local network parameter of the kth local data center.

[0084] It should be noted that the local loss function in step (3.4) has three forms, 1, using the loss function 2, using the loss function 3, using the loss function and the loss function

[0085] The first form:

[0086] When the local loss function is the loss function , the loss function is represented by formula (II):

[0087]

[0088] wherein, is the loss function of the local network model of the kth local data center, is the local network model of the kth local data center, is the normal dose image generated by the local network model of the kth local data center, and the gradient parameter of the loss function is represented by formula (III):

[0089]

[0090] wherein, is the local network model parameter of the kth local data center before updating, is the updated local network model parameter of the kth local data center, and λ is the learning rate, is the gradient of the loss function of the local network model of the kth local data center.

[0091] The second form:

[0092] When the local loss function is the loss function , the loss function is represented by formula (IV):

[0093]

[0094] wherein, is a loss function of the local network model of the kth local data center, is the local network model of the kth local data center, is a normal dose image generated by the local network model of the kth local data center, and the loss function is a gradient parameter of the loss function

[0095]

[0096] wherein, is a local network model parameter of the kth local data center before updating, is a local network model parameter of the kth local data center after updating, and λ is a learning rate, is a loss function of the local network model of the kth local data center, is a gradient of the loss function

[0097] The third form is:

[0098] The local loss function is a loss function and a loss function When the step includes:

[0099] Step (a) is updated by a loss function , wherein the loss function is represented by formula (VI):

[0100]

[0101] wherein, is a loss function of the local network model of the kth local data center, is the local network model of the kth local data center, is a normal dose image generated by the local network model of the kth local data center, and the loss function is a gradient parameter of the loss function

[0102]

[0103] wherein, is a local network model parameter of the kth local data center before updating, is a local network model parameter of the kth local data center after updating, and λ is a learning rate, is a loss function of the local network model of the kth local data center, is a gradient of the loss function

[0104] Step (b), generating new image data by using high frequency information provided by the global shared data distribution information base, is represented by formula (VIII):

[0105]

[0106] wherein, is the inverse transform of high frequency feature extraction operation , is the generated new image data; i is the i-th image data sequence number;

[0107] Step (c), updating the new image data generated in step (b) through a loss function , wherein the loss function is represented by formula (IX):

[0108]

[0109] wherein, is the gradient of the loss function of the local network model of the k-th local data center. The gradient parameter of the loss function is represented by formula (X):

[0110]

[0111] wherein is the gradient of the loss function of the local network model of the k-th local data center.

[0112] It should be noted that the three forms of the local loss function of the present application each has its own advantages, the first one uses the loss function alone, which is commonly used for feature selection, the main reason being that the loss function regularization will make more parameters 0, thus producing a sparse solution, and the present application can discard the features corresponding to 0, and thus use them for feature selection, so the loss function regularization can also prevent model overfitting. The second one uses the loss function alone, which can be used to prevent model overfitting, and intuitively, the loss function regularization is a severe punishment for weight vectors with large values. The third one uses the loss function in combination with the loss function , which can utilize the advantages of the above two norms to a certain extent.

[0113] The paired standardized CT low-dose image data-high-dose image data pair of the present application is a low-dose CT image reconstructed under a low-dose scanning condition-a high-quality CT image data pair reconstructed under a high-dose scanning condition. The specific low-dose CT image reconstructed under a low-dose scanning condition is specifically a CT image data reconstructed using an iterative reconstruction algorithm under a lower-than-normal scanning protocol. The high-quality CT image reconstructed under a high-dose scanning condition is specifically a CT image data reconstructed using an iterative reconstruction algorithm under a normal scanning protocol.

[0114] The high-frequency feature extraction operation of the present application is a Fourier transform or a wavelet transform, and in this embodiment, it is a Fourier transform.

[0115] The training end condition is convergence or reaching the maximum number of training. When the training end condition is convergence, the loss function plays a role in judging the convergence of the network model in the technical field, and the loss value of different loss functions is different based on the different loss functions, and there is no fixed threshold value, which can only be observed, mainly to judge the trend of the loss value, if the trend no longer changes, the model can be judged to be converged. For how to judge the loss function of the local network model, the person skilled in the art can decide according to the actual operation, which will not be repeated here.

[0116] It should be noted that unsupervised learning refers to a class of learning algorithms for unlabeled data. Because there is no label information, it is necessary to discover and summarize patterns or structures from the data set. Common unsupervised algorithms include principal component analysis (PCA), anomaly detection, auto-encoding algorithm, deep belief network, etc.

[0117] The distributed CT intelligent imaging method has the following beneficial effects: first, in terms of data sets, due to patient privacy protection and data barriers of various local image centers, a large amount of modern medical CT image big data cannot be fully utilized. The federated learning used in the present application can achieve the purpose of "safely utilizing data of each local image center without transferring original data", which is one of the beneficial effects of solving data security and alleviating data barriers of each local image center; second, in terms of model performance, under the condition of having more data, federated learning can improve model performance by continuously iterating model parameters, which is superior to the model performance of most deep learning; third, in terms of medical diagnosis and treatment, the present application can actively promote the "graded diagnosis and treatment" advocated by the state, and under the condition of safely sharing data of each local image center, the diagnosis and treatment level of medical institutions at all levels is improved to a certain extent. The present application can combine federated learning with unsupervised learning method, which can protect patient privacy while obtaining high-quality denoising reconstructed images.

[0118] Example 2

[0119] A distributed CT intelligent imaging method, other features are the same as embodiment 1, the difference is that:

[0120] In the CT original measurement data collection process, a standard paired CT data set is collected as a cloud data set, and a local data set composed of three hospital (local data center) original CT measurement image data.

[0121] The cloud data set and the local data set are from three local medical institutions of different CT manufacturers and different scanning protocols. In the model training test process, 300 pairs of standardized CT low dose image data-high dose image data are used to train the global network model of the cloud center, of which 200 are used to construct the training set and 100 are used to construct the test set.

[0122] The local data set includes 700 cases of original CT measurement image data to train the local network model, of which 600 are used to construct the training set and 100 are used to construct the test set.

[0123] The overall training parameter settings of the local data set and the cloud data set are as follows: (1) Set the image block in the CT data set: the block size is set to 64x64, and the step size is set to 64; (2) The learning rate and batch size of the local network model and the global network model are set to 0.0001 and 64 respectively; (3) The number of rounds of the local network model is set to 1 epoch. This embodiment is carried out under the pytorch1 version of the Pytorch framework equipped with two NVIDIA Tesla P40 GPUs.

[0124] Figure 2 The CT image denoising effect of the specific embodiment and the conventional CT imaging method (FBP) in the CT test data of three patients in different local data sets is shown, and the corresponding region of interest (ROI) is shown in the figure. Enlarged diagram, as shown in the box. Figure 2 Each row in the table represents different patient data, and the first column is the label CT image of normal dose, the second column is the conventional CT imaging method, and the third column is the CT image obtained by using the application. In Figure 2 Each row in the table represents the same anatomical position of different patients, and the first column indicates the true value of the image, that is, the CT image at the normal dose level, the second column indicates the image reconstructed by FBP, and the third column indicates the reconstruction result of the application method.

[0125] From Figure 2It can be seen that by evaluating the whole image and comparing the enlarged region of interest, the reconstruction result of the method of the present application can greatly suppress the noise of the low-dose CT image; when observing the enlarged region of interest, it can be seen that the reconstruction result of the method of the present application is superior to the traditional FBP reconstruction algorithm in terms of noise reduction and image detail preservation. The reason why the present application can produce such an effect is that the present application fully exploits the information of big data and uses the federated learning method to improve the performance of the low-dose CT image reconstruction model. It can be seen that the present application can significantly denoise and is superior to the traditional FBP method. Moreover, the CT image results of the present embodiment can further indicate that the distributed CT intelligent imaging system of the present application can solve the problem that the existing multi-center data training federated learning framework may not be suitable for new center CT data, and can perform high-quality CT image reconstruction according to unknown / new center low-dose CT data.

[0126] Embodiment 3

[0127] A distributed CT intelligent imaging system adopts the distributed CT intelligent imaging method of embodiment 1.

[0128] The distributed CT intelligent imaging system can fuse the federated learning and the unsupervised learning method, can protect the privacy of patients and also obtain high-quality denoising reconstruction images.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit the protection scope of the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present application.

Claims

1. A distributed intelligent CT imaging method, characterized in that, The steps include: Step (1) Obtaining a local dataset consisting of raw CT measurement image data from multiple local data centers, and obtaining a cloud dataset consisting of paired and standardized CT low-dose image data and high-dose image data pairs from the cloud center. Step (2): Extract high-frequency features from the original dataset and construct a high-frequency information database of CT data shared by the cloud center. Step (3): Input the high-frequency information database of CT data obtained in step (2) into the cloud basic network for training to obtain the global network model and global network parameters. Distribute the global network parameters to all local data centers. Each local data center inputs the global network parameters and the original CT measurement image data obtained in step (1) into the local basic network for unsupervised training to obtain the local network model and local network parameters. Aggregate the local network parameters of all local data centers and then use the federated learning strategy for interactive training to finally obtain the optimized local network parameters. Step (4): The local data center reconstructs the final CT image based on the optimized local network parameters obtained in step (3) and the original CT measurement image data corresponding to step (1). The paired normalized CT low-dose image data-high-dose image data pair is a low-dose CT image reconstructed under low-dose scanning conditions and a high-quality CT image data pair reconstructed under high-dose scanning conditions; The high-frequency feature extraction operation is either Fourier transform or wavelet transform; The high-frequency characteristic is the noise distribution characteristic.

2. The distributed CT intelligent imaging method according to claim 1, characterized in that: In step (2), the CT data high-frequency information database Where K is the number of local data centers, and B is the high-frequency feature information set of the k-th local data center. k Equation (Ⅰ) represents: in, This is a high-frequency feature extraction operation. For low-dose CT image data in the original CT measurement image data, M k The total amount of high-frequency information contributed by the original CT measurement image data in the k-th local data center, where i is the sequence number of the i-th original CT measurement image data.

3. The distributed CT intelligent imaging method according to claim 2, characterized in that: Step (3) specifically includes: Step (3.1): Input the high-frequency information database of CT data obtained in step (2) into the cloud-based basic network for training to obtain the global network model and global network parameters; Step (3.2): Distribute the global network parameters to all local data centers. Each local data center inputs the global network parameters and the original CT measurement image data obtained in step (1) into the local basic network for unsupervised learning training to obtain the local network model and local network parameters corresponding to each local data center. Step (3.3): Aggregate the local network parameters of each local data center to obtain aggregated model parameters, then upload the aggregated model parameters to the cloud, and the cloud center inputs the aggregated model parameters into the cloud basic network for updating to obtain the updated global network model and updated global network parameters, and distribute the updated global network parameters to all local data centers. Step (3.4): Each local data center performs unsupervised learning training based on the updated global network parameters and the corresponding original CT measurement image data. By iteratively updating the gradient parameters of the local loss function, the updated local imaging model and the updated local network parameters are finally obtained. Step (3.5): Determine the state of the loss function of the local network model of each local data center. When the loss function of all local network models meets the termination condition, define the local network parameters of the current local network model as the optimized local network parameters and proceed to step (4); otherwise, return to step (3.3).

4. The distributed CT intelligent imaging method according to claim 3, characterized in that: The local loss function is the loss function. At that time, loss function Equation (II) represents: in, Let be the loss function of the local network model for the k-th local data center. For the local network model of the k-th local data center, The loss function is used to generate the normal dose image for the k-th local data center network model. The gradient parameter is represented by equation (Ⅲ): in, The parameters of the local network model before the update of the k-th local data center are: Let be the updated local network model parameters for the k-th local data center, and λ be the learning rate. The loss function for the local network model of the k-th local data center The gradient.

5. The distributed CT intelligent imaging method according to claim 3, characterized in that: The local loss function is the loss function. At that time, loss function Equation (Ⅳ) represents: in, Let be the loss function of the local network model for the k-th local data center. For the local network model of the k-th local data center, The loss function is used to generate the normal dose image for the k-th local data center network model. The gradient parameter is represented by equation (V): in, The parameters of the local network model before the update of the k-th local data center are: Let be the updated local network model parameters for the k-th local data center, and λ be the learning rate. The loss function for the local network model of the k-th local data center The gradient.

6. The distributed CT intelligent imaging method according to claim 3, characterized in that: The local loss function is the loss function. and loss function The steps include: Step (a) Using the loss function Update the loss function. Equation (VI) represents: in, Let be the loss function of the local network model for the k-th local data center. For a local network model with k local data centers, The loss function is used to generate the normal dose image for the k-th local data center network model. The gradient parameter is represented by equation (Ⅶ): in, The parameters of the local network model before the update of the k-th local data center are: Let be the updated local network model parameters for the k-th local data center, and λ be the learning rate. The loss function for the local network model of the k-th local data center The gradient; Step (b): Using the high-frequency information provided by the globally shared data distribution information database, new image data is generated, as represented by equation (VIII): in, For high-frequency feature extraction operations inverse transform, This represents the newly generated image data; i is the sequence number of the i-th image data. Step (c): Pass the new image data generated in step (b) through a loss function. Update the loss function. Equation (IX) represents: in, Let be the loss function of the local network model for the k-th local data center. The gradient parameter is represented by equation (X): in The loss function for the local network model of the k-th local data center The gradient.

7. The distributed CT intelligent imaging method according to any one of claims 4 to 6, characterized in that: In step (3.3), the aggregation model parameters are obtained using equation (XI): Among them, H agg N represents the parameters of the aggregation model. k Let be the number of training data points for the local network model in the k-th local data center. These are the local network parameters for the k-th local data center.

8. The distributed CT intelligent imaging method according to claim 1, characterized in that: The local dataset Equation (XII) represents:

9. The distributed CT intelligent imaging method according to claim 3, characterized in that: The training termination condition is convergence or reaching the maximum number of training iterations.

10. A distributed intelligent CT imaging system, characterized in that: The distributed CT intelligent imaging method as described in any one of claims 1 to 9 is adopted.

Images