A multi-task supervised distributed federated learning CT imaging method and system

Abstract

The application discloses a multi-task supervised distributed federated learning CT imaging method and system, wherein the multi-task supervised distributed federated learning CT imaging method is to perform low-dose CT image reconstruction on a local paired detection data-image data pair dataset acquired under different local tasks and a cloud paired standardized CT detection data-image data pair dataset through three steps. According to the application, the image center reconstructs the final CT image according to the optimized local network parameters obtained in step (2) and the projection data in the local paired detection data-image data pair. The multi-task supervised distributed federated learning CT imaging method and system can improve the model performance of the cloud and the local while protecting the data privacy, and solves the statistical problem of the federated learning by considering the correlation between different local tasks.

Description

Technical Field

[0001] The present invention relates to the field of medical CT imaging technology, and in particular to a multi-task supervised distributed federated learning CT imaging method and a multi-task supervised distributed federated learning CT imaging system. Background Art

[0002] X-ray computed tomography (CT) technology has been widely used in the field of medical diagnosis, providing considerable convenience for clinical diagnosis and treatment.

[0003] Currently, CT scans use X-ray beams to scan a certain thickness of a certain part of the human body. The harmful effects of radiation exposure are significant. Excessive CT doses can induce leukemia, cancer, and other genetic diseases. Reducing the tube current is a common clinical approach to reduce radiation dose, but this approach has the drawback of significantly reducing image quality.

[0004] Spectral CT, CT perfusion imaging, and cone-beam CT are all subcategories of X-ray computed tomography. Spectral CT generally refers to an imaging method that uses multi-spectral information to improve image quality or provide new image information. CT perfusion imaging differs from dynamic scanning in that it involves a rapid intravenous bolus injection of contrast agent and continuous CT scanning of the region of interest, thereby obtaining a time-density curve for the region of interest. Various perfusion parameter values ​​are then calculated using different mathematical models, enabling more effective quantitative reflection of changes in local tissue blood perfusion. This is a cutting-edge technology in the field of CT applications and is of great significance for clarifying the blood supply of lesions. Cone-beam CT (CBCT) is often used to image the maxillofacial region, marking the transition from 2D to 3D imaging of the maxillofacial region, both in data acquisition and image reconstruction, and truly expanding the role of imaging from diagnosis to image guidance for surgery and surgical procedures. Multi-task learning of spectral CT, CT perfusion imaging, and cone-beam CT can obtain additional information that helps improve model performance and increase data utilization efficiency by mining the relationships between tasks. However, multi-task learning has always been plagued by the problem of data privacy leakage.

[0005] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide a multi-task supervised distributed federated learning CT imaging method and a multi-task supervised distributed federated learning CT imaging system to solve the shortcomings of the existing technology. Summary of the Invention

[0006] One of the purposes of the present application is to provide a multi-task supervised distributed federated learning CT imaging method to avoid the shortcomings of the prior art. The multi-task supervised distributed federated learning CT imaging method can improve the model performance of the cloud and the local while protecting the data privacy, and solve the statistical problem of federated learning by considering the correlation between different local tasks.

[0007] The above-mentioned purposes of the present application are achieved by the following technical measures:

[0008] The present application provides a multi-task supervised distributed federated learning CT imaging method, comprising the following steps:

[0009] Step (1), acquiring a local paired probe data-image data pair dataset acquired by an image center under different local tasks, and acquiring a cloud paired standardized CT probe data-image data pair dataset, wherein the local tasks are traditional CT imaging, perfusion CT imaging, spectral CT imaging or CBCT imaging;

[0010] Step (2), inputting the paired standardized CT local paired probe data-image data pair into the cloud basic network for supervised learning training to obtain a global imaging model and global network parameters, inputting the global network parameters and the corresponding local paired probe data-image data pair into the local basic network for supervised learning training to obtain local network parameters, and then using a federated learning strategy to interactively train the local network parameters to obtain optimized local network parameters;

[0011] Step (3), reconstructing the final CT image according to the optimized local network parameters obtained in step (2) and the projection data in the corresponding local paired probe data-image data pair.

[0012] Preferably, the above step (2) specifically comprises:

[0013] Step (2.1), inputting the cloud paired standardized CT projection data-image pair dataset into the cloud basic network for supervised learning training to obtain a global imaging model and global network parameters;

[0014] Step (2.2), distributing the global network parameters to all image centers, inputting the global network parameters and the corresponding local paired probe data-image data pair into the local basic network for supervised learning training to obtain a local imaging model and local network parameters corresponding to each local task;

[0015] Step (2.3), aggregating the local network parameters of each imaging center to obtain aggregated model parameters, then uploading the aggregated model parameters to the cloud, inputting the aggregated model parameters into the cloud-based network, further updating the global imaging model and global network parameters by updating the gradient parameters of the global loss function of the global imaging model, and distributing the updated global network parameters to all imaging centers;

[0016] Step (2.4), each local task respectively performs supervised learning training according to the updated global network parameters and the corresponding local paired probe data-image data pairs, and finally obtains the updated local imaging model and its parameters by iteratively updating the gradient parameters of the global loss function;

[0017] Step (2.5), judging the state of each local loss function and the global loss function, when each local loss function and the global loss function are converged, defining the current each local network parameter as the optimized local network parameter, and entering step (3); otherwise, returning to step (2.3).

[0018] In the step (2.3), the aggregated model parameters are obtained by formula (I) and formula (II):

[0019]

[0020] Wherein, N agg is the aggregated model parameter, n is the total sample number of all local task CT data sets, M k is the number of local paired probe data-image data pairs in the kth local imaging model, and K is the total number of local imaging models.

[0021] Preferably, the global loss function is represented by formula (III):

[0022]

[0023] Wherein, L cloud-globa is the global loss function, is the cloud-based global imaging model, is the probe data in the cloud-based paired standardized CT probe data-image data pair, is the image data in the cloud-based paired standardized CT probe data-image data pair, is the image generated by the cloud-based global imaging model, and i represents the i th cloud data set sequence number.

[0024] Preferably, the gradient parameters of the global loss function are represented by formula (IV):

[0025]

[0026] in are the global network parameters before updating, are the updated global network parameters, is the local network parameter of the k-th image center, P is the learning rate, is the gradient of the global model loss function.

[0027] Preferably, the above local loss function is represented by formula (V):

[0028]

[0029] Among them, L local is the loss function of the local imaging model of the k-th image center, is the local imaging model, is the i-th data sequence number of the detection data in the local paired detection data-image data pair in the local imaging model of the k-th image center, is the i-th data sequence number of the image data in the local paired detection data-image data pair in the local imaging model of the k-th image center, The corresponding i-th normal-dose CT image generated by the local imaging model for the k-th image center.

[0030] When the local task is conventional CT imaging or perfusion CT imaging, A single-channel local imaging model is used; when the local task is CBCT imaging or spectral CT imaging, A multi-channel local imaging model is adopted.

[0031] Preferably, the gradient parameter of the above local loss function is represented by formula (VI):

[0032]

[0033] in, is the local network parameter of the k-th image center before updating, is the local network parameter of the k-th image center after update, is the gradient of the local imaging model loss function at the k-th image center.

[0034] In the step (2.2), supervised learning uses the image data in the cloud-based paired standardized CT projection data-image data pairs as label data.

[0035] In the step (2.4), supervised learning is performed using the image data in the local paired detection data-image data pair as label data.

[0036] Preferably, the above-mentioned local paired detection data-image data pairs are CT detection data under low-dose scanning conditions-high-quality image data pairs obtained through iterative reconstruction algorithm; or low-dose detection data-normal-dose image data pairs obtained through simulation.

[0037] Preferably, the above-mentioned cloud-paired standardized CT detection data-image data pair is a CT detection data of a phantom under low-dose scanning conditions-a CT image data pair of a phantom under high-dose scanning conditions, a CT detection data under low-dose scanning conditions-a high-quality CT image data pair obtained by an iterative reconstruction algorithm, or a low-dose CT detection data-normal-dose CT image data pair obtained by synthesis.

[0038] Preferably, the above-mentioned imaging center includes at least a data set of local paired detection data-image data pairs collected under two local tasks.

[0039] Preferably, there are at least two imaging centers.

[0040] Preferably, the cloud-based network and the local basic network are both residual networks.

[0041] Another object of the present invention is to overcome the shortcomings of the prior art and provide a multi-task supervised distributed federated learning CT imaging system. This multi-task supervised distributed federated learning CT imaging system can improve the performance of cloud and local models while protecting data privacy, and solve the statistical difficulties of federated learning by considering the correlation between different local tasks.

[0042] The above-mentioned purpose of the present invention is achieved through the following technical measures:

[0043] A multi-task supervised distributed federated learning CT imaging system is provided, which adopts the above-mentioned multi-task supervised distributed federated learning CT imaging method.

[0044] The present invention discloses a multi-task supervised distributed federated learning CT imaging method and system, wherein the multi-task supervised distributed federated learning CT imaging method comprises the following steps: step (1), obtaining a data set of local paired detection data-image data pairs collected by an imaging center under different local tasks, and obtaining a data set of paired standardized CT detection data-image data pairs on the cloud, wherein the local tasks are traditional CT imaging, perfusion CT imaging, energy spectrum CT imaging or CBCT imaging; step (2), inputting the paired standardized CT local paired detection data-image data pairs into a basic network for supervised learning training to obtain a global imaging model and global network parameters, and each imaging center performs supervised learning training on the basic network according to the global network parameters and the corresponding local paired detection data-image data pairs to obtain local network parameters, and then adopts a federated learning strategy for interactive training to finally obtain optimized local network parameters; step (3), the imaging center reconstructs the final CT image according to the optimized local network parameters obtained in step (2) and the projection data in the corresponding local paired detection data-image data pairs. The present invention can improve the model performance of the cloud and the local while protecting data privacy, and solve the statistical problem of federated learning by considering the correlation between different local tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention is further described with reference to the accompanying drawings, but the contents in the drawings do not constitute any limitation to the present invention.

[0046] Figure 1 This is a flowchart of a multi-task supervised distributed federated learning CT imaging method of the present invention.

[0047] Figure 2 This is the image reconstructed from the data in Example 2. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described with reference to the following examples.

[0049] Example 1

[0050] A multi-task supervised distributed federated learning CT imaging method, such as Figure 1 As shown, the following steps are included:

[0051] Step (1), obtaining a dataset of local paired detection data-image data pairs collected by an imaging center under different local tasks, and obtaining a dataset of cloud-based paired standardized CT detection data-image data pairs, wherein the local tasks are conventional CT imaging, perfusion CT imaging, spectral CT imaging, or CBCT imaging;

[0052] Step (2), inputting the paired standardized CT local paired probe data-image data pairs into the cloud-based network for supervised learning training to obtain a global imaging model and global network parameters, each imaging center inputting the global network parameters and corresponding local paired probe data-image data pairs into the local-based network for supervised learning training to obtain local network parameters, and then using a federated learning strategy to interactively train the final optimized local network parameters;

[0053] Step (3), reconstructing a final CT image according to the optimized local network parameters obtained in step (2) and projection data in the corresponding local paired probe data-image data pairs.

[0054] In step (2), the following steps are included:

[0055] Step (2.1), inputting the cloud paired standardized CT projection data-image pairs dataset into the cloud-based network for supervised learning training to obtain a global imaging model and global network parameters;

[0056] Step (2.2), distributing the global network parameters to all imaging centers, inputting the global network parameters and corresponding local paired probe data-image data pairs into the local-based network for supervised learning training to obtain a local imaging model and local network parameters corresponding to each local task;

[0057] Step (2.3), aggregating the local network parameters of each imaging center to obtain aggregated model parameters, then uploading the aggregated model parameters to the cloud, inputting the aggregated model parameters into the cloud-based network, further updating the global imaging model and global network parameters by updating the gradient parameters of the global loss function of the global imaging model, and distributing the updated global network parameters to all imaging centers;

[0058] Step (2.4), each local task performs supervised learning training according to the updated global network parameters and corresponding local paired probe data-image data pairs, and finally obtains updated local imaging models and updated local network parameters by iteratively updating the gradient parameters of the local loss function;

[0059] Step (2.5), judging the state of each local loss function and global loss function, when each local loss function and global loss function is converged, the current local network parameters are defined as the optimized local network parameters, and step (3) is entered; otherwise, return to step (2.3).

[0060] In step (2.2), the image data in the cloud paired standardized CT projection data-image data pairs is used as label data for supervised learning.

[0061] In step (2.4), the supervised learning is to use the image data in the local paired probe data-image data pair as the label data.

[0062] It should be noted that the supervised learning refers to a kind of learning algorithm under the condition that the correct output of the data set is known, because the input and output are known, it means that there is a relationship between the input and output. The supervised learning algorithm is to find and summarize this "relationship". Common supervised algorithms include linear regression, neural network, decision tree, support vector machine, KNN, and naive Bayes algorithm. The present application discovers the internal mode or structure based on the relationship between the variables in the data by using the clustering algorithm.

[0063] In step (2.3), the aggregated model parameters are obtained by formula (I) and formula (II):

[0064]

[0065] Wherein, N agg is the aggregated model parameter, n is the total sample number of all local task CT data sets, M k is the sample number of the local paired probe data-image data pair in the kth local imaging model, and K is the total number of local imaging models.

[0066] The global loss function of the present application is represented by formula (III):

[0067]

[0068] Wherein, L cloud-global is the global loss function, is the cloud global imaging model, is the probe data in the cloud paired standardized CT probe data-image data pair, is the image data in the cloud paired standardized CT probe data-image data pair, is the image generated by the cloud global imaging model, and i represents the i th cloud data set serial number.

[0069] It should be noted that the data set of the cloud paired standardized CT probe data-image data pair of the present application is the data set of the conventional CT, which is suitable for different task models to train and fine-tune the cloud global model parameters, so as to help the training of the specific task in the kth image center.

[0070] The gradient parameter of the global loss function of the present application is represented by formula (IV):

[0071]

[0072] Wherein is the global network parameter before updating, are the updated global network parameters, is the local network parameter of the k-th image center, P is the learning rate, is the gradient of the global model loss function.

[0073] The local loss function of the present invention is represented by formula (V):

[0074]

[0075] Among them, L local-k is the loss function of the local imaging model of the k-th image center, is the local imaging model, is the i-th data sequence number of the detection data in the local paired detection data-image data pair in the local imaging model of the k-th image center, is the i-th data sequence number of the image data in the local paired detection data-image data pair in the local imaging model of the k-th image center, The corresponding i-th normal dose CT image generated by the local imaging model for the k-th image center.

[0076] When the local task is conventional CT imaging or perfusion CT imaging, A single-channel local imaging model is used; when the local task is CBCT imaging or spectral CT imaging, A multi-channel local imaging model is adopted.

[0077] Among them, the gradient parameter of the local loss function is expressed by formula (VI):

[0078]

[0079] in, is the local network parameter of the k-th image center before updating, is the local network parameter of the k-th image center after update, is the gradient of the local imaging model loss function at the k-th image center.

[0080] The local paired detection data-image data pair of the present invention is a CT detection data under low-dose scanning conditions-high-quality image data pair obtained through an iterative reconstruction algorithm; or a low-dose detection data-normal-dose image data pair obtained through simulation.

[0081] The "patient detection data under low-dose scanning conditions - high-quality image data obtained using an iterative reconstruction algorithm" pair includes: CT detection data of the patient under low-dose scanning conditions, specifically CT detection data obtained under mAs conditions lower than those of conventional scanning protocols (e.g., less than 100 mAs). The "high-quality image data obtained using an iterative reconstruction algorithm" specifically refers to high-quality CT images obtained by reconstructing low-dose CT detection data using an iterative reconstruction algorithm; the two constitute a matching data pair. The simulated low-dose CT detection data is obtained by adding appropriate Gaussian noise and Poisson noise to the CT detection data obtained through conventional scanning, and can closely simulate detection data obtained through clinical low-dose scanning protocols.

[0082] The cloud-paired standardized CT detection data-image data pair of the present invention is a pair of CT detection data of a phantom under low-dose scanning conditions and CT image data of a phantom under high-dose scanning conditions, a pair of CT detection data under low-dose scanning conditions and high-quality CT image data obtained through an iterative reconstruction algorithm, or a pair of low-dose CT detection data and normal-dose CT image data obtained through synthesis.

[0083] The image center of the present invention includes at least a dataset of local paired detection data and image data pairs collected during two local tasks. There are at least two image centers. The cloud-based network and the local base network of the present invention are both residual networks.

[0084] The global loss function and the local loss function of the present invention can be an L1 loss function or an L2 loss function, and the specific type is a perceptual loss function or a generative adversarial loss function.

[0085] It should be further explained that in this technical field, the role of the loss function is to determine the convergence of the network model. Different loss functions have different loss values, and there is no fixed threshold. The only way to determine whether the loss value has converged is to observe the downward trend of the loss value. If the trend does not change, the model can be judged to have converged. Whether to use a global loss function or a local loss function for judgment can be determined by those skilled in the art based on actual operational circumstances, and will not be elaborated on here.

[0086] The multi-task supervised distributed federated learning CT imaging method can process different CT imaging (traditional CT, perfusion CT, spectral CT, CBCT imaging) local tasks according to the task requirements of each image center, and the local task drives the reconstructed image to achieve image reconstruction of different local tasks. The data protection mechanism of federated learning can effectively solve this problem, and multi-task learning can solve the statistical problem of federated learning, and the two complement each other to build an effective learning model. Through the combination of federated learning and multi-task learning in the technical scheme of the application, the model performance of the cloud and the local can be improved while protecting the data privacy, and the correlation between different local tasks is considered to solve the statistical problem of federated learning.

[0087] Embodiment 2

[0088] The application of a multi-task supervised distributed federated learning CT imaging method, such as Figure 2 .

[0089] The CT data set of the cloud-end paired standardized CT probe data-image data pair and the three local medical centers (i.e. image centers) local paired probe data-image data pairs collected in this embodiment are from three local hospitals of different CT manufacturers and different scanning protocols.

[0090] In the data construction process, 674 paired CT chest tomography images are used as cloud-end paired standardized CT probe data-image data pairs for supervised training of the cloud-end denoising model (i.e. global imaging model), of which 472 are used to construct a training set, 68 are used to construct a validation set, and 134 are used to construct a test set.

[0091] Paired data sets of three different devices and protocols are used, which are 1410 CBCT lung scan patient data, 1450 spectral CT abdominal scan patient data and 704 brain perfusion CT sequence scan patient data as local paired probe data-image data pairs, of which 70% of the data is used to construct a training set, 10% of the data is used to construct a validation set, and 20% of the data is used to construct a test set.

[0092] The training parameter settings are as follows: (1) the CT image block size is set to 64x64, and the step size is set to 64; (2) the learning rate and batch size of the global imaging model and the local imaging model are set to 0.001 and 64; (3) the number of rounds of model training between the cloud and the local end is set to 1 epoch.

[0093] The test results of this embodiment are as follows: Figure 2As shown in the figure, it shows the denoising results of three different scanning equipment data images, and gives the corresponding ROI magnification diagram. Among them, the first to third rows are CBCT lung scan data, spectral CT abdomen scan data and brain perfusion CT sequence scan data respectively. The first column is the simulated low-dose input image with Gaussian noise and Poisson noise, the second column is the denoising effect of the present invention, and the third column is the reference image. Figure 2 It can be seen that the present invention can suppress the noise influence caused by low-dose imaging to a certain extent, and the obtained image is close to the reference image.

[0094] Example 3

[0095] A multi-task supervised distributed federated learning CT imaging system adopts the multi-task supervised distributed federated learning CT imaging method as described in Example 1.

[0096] The multi-task supervised distributed federated learning CT imaging system of the present invention can improve the performance of cloud and local models while protecting data privacy, and solve the statistical problems of federated learning by considering the correlation between different local tasks.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A multi-task supervised distributed federated learning CT imaging method, characterized in that: The steps include: Step (1), obtaining a dataset of local paired detection data-image data pairs collected by an imaging center under different local tasks, and obtaining a dataset of cloud-based paired standardized CT detection data-image data pairs, wherein the local tasks are conventional CT imaging, perfusion CT imaging, spectral CT imaging, or CBCT imaging; Step (2) inputs the paired standardized CT local paired detection data-image data pairs into the cloud-based basic network, performs supervised learning training to obtain the global imaging model and global network parameters, and each imaging center inputs the global network parameters and the corresponding local paired detection data-image data pairs into the local basic network for supervised learning training to obtain the local network parameters, and then adopts the federated learning strategy for interactive training to finally obtain the optimized local network parameters; In step (3), the imaging center reconstructs the final CT image based on the optimized local network parameters obtained in step (2) and the projection data in the corresponding local paired detection data-image data pair.

2. The multi-task supervised distributed federated learning CT imaging method according to claim 1, characterized in that: The step (2) specifically includes: Step (2.1), inputting the cloud-based paired standardized CT projection data-image pair dataset into the cloud-based basic network for supervised learning training to obtain the global imaging model and global network parameters; Step (2.2): Distribute the global network parameters to all imaging centers, input the global network parameters and the corresponding local paired detection data-image data pairs into the local basic network for supervised learning training to obtain the local imaging model and local network parameters corresponding to each local task; Step (2.3): Aggregate the local network parameters of each imaging center to obtain the aggregate model parameters, then upload the aggregate model parameters to the cloud, input the aggregate model parameters into the cloud basic network, further update the global imaging model and global network parameters by updating the gradient parameters of the global loss function of the global imaging model, and distribute the updated global network parameters to all imaging centers; In step (2.4), each local task performs supervised learning training based on the updated global network parameters and the corresponding local paired detection data-image data pairs, and finally obtains the updated local imaging model and updated local network parameters by iteratively updating the gradient parameters of the local loss function; Step (2.5) determines the status of each local loss function and the global loss function. When both the local loss function and the global loss function converge, the current local network parameters are defined as the optimized local network parameters and the process goes to step (3). If not, the process goes back to step (2.3).

3. The multi-task supervised distributed federated learning CT imaging method according to claim 2, characterized in that: In the step (2.3), the aggregation model parameters are obtained by formula (I) and formula (II): Among them, N agg is the aggregation model parameter, n is the total number of samples in all local task CT datasets, M k is the number of local paired detection data-image data pairs in the kth local imaging model, and K is the total number of local imaging models.

4. The multi-task supervised distributed federated learning CT imaging method according to claim 3, characterized in that: The global loss function is expressed by formula (III): Among them, L cloud-global is the global loss function, It is a global imaging model in the cloud. Standardize the detection data in CT detection data-image data pairs for cloud pairing, Standardize the image data in CT detection data-image data pairs for cloud pairing, The image generated by the cloud global imaging model, i represents the sequence number of the i-th cloud dataset.

5. The multi-task supervised distributed federated learning CT imaging method according to claim 4, characterized in that: The gradient parameter of the global loss function is expressed by formula (IV): in are the global network parameters before updating, are the updated global network parameters, is the local network parameter of the k-th image center, P is the learning rate, is the gradient of the global model loss function.

6. The multi-task supervised distributed federated learning CT imaging method according to claim 3, characterized in that: The local loss function is expressed by formula (V): Among them, L loc is the loss function of the local imaging model of the k-th image center, is the local imaging model, is the i-th data sequence number of the detection data in the local paired detection data-image data pair in the local imaging model of the k-th image center, is the i-th data sequence number of the image data in the local paired detection data-image data pair in the local imaging model of the k-th image center, The corresponding i-th normal-dose CT image generated by the local imaging model for the k-th image center; When the local task is conventional CT imaging or perfusion CT imaging, A single-channel local imaging model is used; when the local task is CBCT imaging or spectral CT imaging, A multi-channel local imaging model is adopted.

7. The multi-task supervised distributed federated learning CT imaging method according to claim 6, characterized in that: The gradient parameter of the local loss function is expressed by formula (VI): in, is the local network parameter of the k-th image center before updating, is the local network parameter of the k-th image center after update, is the gradient of the local imaging model loss function at the k-th image center.

8. The multi-task supervised distributed federated learning CT imaging method according to claim 6, characterized in that: In the step (2.2), supervised learning is to use the image data in the cloud-based paired standardized CT projection data-image data pair as label data; In the step (2.4), supervised learning is performed using the image data in the local paired detection data-image data pair as label data.

9. The multi-task supervised distributed federated learning CT imaging method according to claim 6, characterized in that: The local paired detection data-image data pair is a CT detection data-high-quality image data pair obtained through an iterative reconstruction algorithm under low-dose scanning conditions; or a low-dose detection data-normal-dose image data pair obtained through simulation; The cloud-based paired standardized CT detection data-image data pair is a pair of CT detection data of a phantom under low-dose scanning conditions and CT image data of a phantom under high-dose scanning conditions, a pair of CT detection data under low-dose scanning conditions and high-quality CT image data obtained through an iterative reconstruction algorithm, or a pair of low-dose CT detection data obtained through synthesis and normal-dose CT image data; The imaging center includes at least a data set of local paired detection data-image data pairs collected under two local tasks; The number of the image centers is at least two; The cloud-based basic network and the local basic network are both residual networks.

10. A multi-task supervised distributed federated learning CT imaging system, characterized by: A multi-task supervised distributed federated learning CT imaging method as described in any one of claims 1 to 9 is adopted.

Images