Multi-data center training system for endometrial cancer target segmentation model

Abstract

The invention relates to a multi-data center training system for an endometrial cancer target region segmentation model, belongs to the technical field of target region sketching, and solves the problem of low performance of a segmentation model in the prior art. The system comprises a main control end and a plurality of data centers, wherein each data center stores a local multi-mode endometrial cancer sample set; a target sample set is stored on the main control end; constructing a target region segmentation model at the main control end, and sending the target region segmentation model to each data center; the main control end controls the plurality of data centers to train a local target region segmentation model based on the local multi-mode endometrial cancer sample set, and parameters of the target region segmentation model of the main control end are updated based on the parameters of the local target region segmentation model of the plurality of data centers to obtain a pre-trained target region segmentation model; and the main control end carries out fine tuning on the pre-trained target region segmentation model based on the target sample set to obtain a trained endometrial cancer target region segmentation model. The efficient and accurate delineation of the endometrial cancer target region is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of target segmentation, and particularly to a training system for an endometrial cancer target segmentation model of multiple data centers. Background Art

[0002] Endometrial cancer is one of the common malignant tumors in the female reproductive system, and its incidence rate shows an upward trend globally. In China, with the aging of the population and the change of lifestyle, the number of new cases of endometrial cancer is also increasing continuously. This makes the accurate diagnosis and effective treatment of endometrial cancer an important medical need. For endometrial cancer patients, target segmentation after surgery is crucial in the subsequent treatment plan formulation (such as radiotherapy plan). Accurate target segmentation can ensure that treatment means such as radiotherapy act precisely on the tumor area, reduce the damage to surrounding normal tissues, improve the quality of life of patients and improve the prognosis.

[0003] The data volume of endometrial cancer patients collected by different medical centers is not large and there are differences. Simply training a model with the data collected by a certain medical center, due to the small sample size, the model is prone to overfitting and cannot accurately perform image semantic segmentation, that is, it cannot accurately perform target segmentation, resulting in low model performance. If the data of each center is mixed together for training, it is required that the delineation rules of different centers are consistent. When the delineation rules between centers are inconsistent or the delineation styles are different, the training performance is usually poor, that is, the model performance is low. Summary of the Invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a training system for an endometrial cancer target segmentation model of multiple data centers to solve the problem of low performance of the existing segmentation model.

[0005] On the one hand, an embodiment of the present invention provides a training system for an endometrial cancer target segmentation model of multiple data centers, and the system includes a main control end and multiple data centers:

[0006] Each of the data centers stores a local multi-modal endometrial cancer sample set; the main control end stores a target sample set;

[0007] Build a target segmentation model on the main control end and send the target segmentation model to each of the data centers;

[0008] The main control end controls the multiple data centers to train the local target segmentation model based on the local multi-modal endometrial cancer sample set, and updates the parameters of the target segmentation model on the main control end based on the local target segmentation model parameters of the multiple data centers to obtain a pre-trained target segmentation model;

[0009] The master control end fine-tunes the pre-trained target area segmentation model based on the target sample set to obtain a trained endometrial cancer target area segmentation model.

[0010] Based on the further improvement of the above system, the master control end controls the multiple data centers to train the local target area segmentation model based on the local multi-modal endometrial cancer sample set, and updates the parameters of the target area segmentation model of the master control end based on the parameters of the local target area segmentation models of the multiple data centers to obtain a pre-trained target area segmentation model:

[0011] S21. The master control end randomly selects N data centers and sends training instructions to the selected N data centers;

[0012] S22. After each selected data center receives the training instructions, it trains the local target area segmentation model based on the local multi-modal endometrial cancer sample set, calculates the training loss of the local target area segmentation model, and sends the calculated training loss to the master control end;

[0013] S23. The master control end updates the parameters of the local target area segmentation model according to the training losses of the selected N data centers, and sends the updated parameters of the local target area segmentation model to each data center; determines whether the stop condition is reached at present. If so, stops training to obtain a pre-trained target area segmentation model; otherwise, the master control end executes step S21.

[0014] Based on the further improvement of the above system, the master control end updates the parameters of the local target area segmentation model according to the training losses of the selected N data centers, including:

[0015] The master control end calculates the pre-update parameters of each selected data center according to the training losses of the selected N data centers, and sends the pre-update parameters of each selected data center to the corresponding data center;

[0016] Each selected data center pre-updates the local target area segmentation model according to the pre-update parameters, calculates the pre-update loss based on the pre-updated local target area segmentation model, and sends the pre-update loss to the master control end;

[0017] The master control end calculates the weights of the N data centers based on the pre-update losses of the N data centers;

[0018] Update the parameters of the local target area segmentation model based on the training losses of the N data centers and the weights of the N data centers.

[0019] Based on the further improvement of the above system, the master control end calculates the pre-update parameters of each selected data center according to the training losses of the selected N data centers, including:

[0020] The first pre-update parameters of the N selected data centers are calculated based on the training losses of the N selected data centers;

[0021] For the i-th selected data center, the second pre-update parameters of the i-th selected data center are obtained based on the training losses of the other N - 1 selected data centers;

[0022] The first pre-update parameters and the second pre-update parameters of the i-th selected data center constitute the pre-update parameters of the i-th selected data center.

[0023] Based on a further improvement of the above system, each selected data center pre-updates the local target segmentation model according to the pre-update parameters, and calculates the pre-update loss based on the pre-updated local target segmentation model, including:

[0024] Each selected data center pre-updates the local target segmentation model based on the first pre-update parameters and the second pre-update parameters respectively, to obtain a first pre-update model and a second pre-update model;

[0025] Calculate the loss on the test set of the local multi-modal endometrial cancer sample set based on the first pre-update model to obtain the first pre-update loss;

[0026] Calculate the loss on the test set of the local multi-modal endometrial cancer sample set based on the second pre-update model to obtain the second pre-update loss;

[0027] The difference between the second pre-update loss and the first pre-update loss gives the pre-update loss.

[0028] Based on a further improvement of the above system, the weights of each selected data center are calculated using the following formula based on the pre-update losses of the N data centers:

[0029]

[0030] where, q i represents the data quality of the i-th selected data center, δ i represents the pre-update loss of the i-th selected data center, and γ represents a regulation parameter.

[0031] Based on a further improvement of the above system, the multi-modal endometrial cancer sample set includes a training set and a test set; the training set is used to train the local target segmentation model, and the test set is used to calculate the training loss of the local target segmentation model;

[0032] The multi-modal endometrial cancer samples include paired plain CT images and enhanced CT images of endometrial cancer; some multi-modal endometrial cancer samples also include mask images corresponding to the paired plain CT images and enhanced CT images of endometrial cancer;

[0033] The training loss is calculated using the following formula:

[0034] Loss = L 1 + L 2 + L 3

[0035] where L 1 represents the multi-modal image segmentation loss, L 2 represents the multi-modal cross loss, and L 3 represents the contrast loss.

[0036] Based on the further improvement of the above system, the multi-modal cross loss is calculated using the following formula:

[0037]

[0038] where represents the number of samples in the test set of the i-th data center that do not have corresponding mask images, and P k1 (s, j) represents the element value at the (s, j) position in the first segmentation image of the k-th sample that does not have a corresponding mask image, and P k2 (s, j) represents the element value at the (s, j) position in the second segmentation image of the k-th sample that does not have a corresponding mask image, and m represents the number of elements in the first segmentation image.

[0039] Based on the further improvement of the above system, the contrast loss is calculated in the following manner:

[0040] For each sample in the test set that does not have a corresponding mask image, the first segmentation image and the second segmentation image of this sample are used as a positive sample pair; the first segmentation image of this sample and the second segmentation image of other samples in the test set that do not have corresponding mask images are used as negative sample pairs; positive and negative sample pairs corresponding to each sample in the test set that does not have a corresponding mask image are obtained;

[0041] The contrast loss is calculated based on the positive and negative sample pairs corresponding to each sample in the test set that does not have a corresponding mask image.

[0042] Based on the further improvement of the above system, the contrast loss is calculated based on the positive and negative sample pairs corresponding to each sample in the test set that does not have a corresponding mask image in the following manner:

[0043]

[0044] where represents the number of samples in the test set of the i-th data center that do not have corresponding mask images, and P k1The first segmentation image representing the k-th sample without a corresponding mask image, P k2 The second segmentation image representing the k-th sample without a corresponding mask image, P j2 The second segmentation image representing the j-th sample without a corresponding mask image, τ represents a tuning parameter, and dice(·,·) represents the DICE loss.

[0045] Compared with the prior art, the present invention constructs a master control end and multiple data centers. The master control end and multiple data centers maintain their own sample sets, improving data security. At the same time, the master control end constructs a target area segmentation model and sends it to each data center; the master control end controls multiple data centers to pre-train the local target area segmentation model based on the local sample set, and then updates the parameters of the target area segmentation model of the master control end to realize the training of the target area segmentation model of the master control end, obtaining a pre-trained target area segmentation model. The master control end then fine-tunes the pre-trained target area segmentation model on the target data set, so as to quickly train a target area segmentation model suitable for the target data set, and will not affect the performance of the model due to the small amount of data in the target data set, thus improving the segmentation performance of the model on the target data set.

[0046] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can be made obvious from the description, or understood by implementing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the content specifically pointed out in the description and the drawings. Description of the Drawings

[0047] The drawings are only for the purpose of showing specific embodiments, and are not considered as a limitation of the present invention. Throughout the drawings, the same reference signs represent the same components;

[0048] Figure 1 It is a block diagram of a training system for an endometrial cancer target area segmentation model with multiple data centers according to an embodiment of the present invention. Detailed Embodiments

[0049] The following will specifically describe the preferred embodiments of the present invention in conjunction with the drawings, where the drawings form a part of this application and are used together with the embodiments of the present invention to explain the principle of the present invention, and are not used to limit the scope of the present invention.

[0050] A specific embodiment of the present invention discloses a training system for an endometrial cancer target area segmentation model with multiple data centers, as Figure 1 shown, the system includes a master control end and multiple data centers;

[0051] Each of the data centers stores a local multi-modal endometrial cancer sample set; the master control end stores a target sample set;

[0052] Build a target segmentation model on the master control end and send the target segmentation model to each of the data centers;

[0053] The master control end controls the multiple data centers to train the local target segmentation model based on the local multi-modal endometrial cancer sample set, and updates the parameters of the target segmentation model on the master control end based on the local target segmentation model parameters of the multiple data centers to obtain a pre-trained target segmentation model;

[0054] The master control end fine-tunes the pre-trained target segmentation model based on the target sample set to obtain a trained endometrial cancer target segmentation model.

[0055] During implementation, each data center collects multi-modal image data of endometrial cancer patients after surgery in this center to build a multi-modal endometrial cancer sample set.

[0056] The master control end also has a local multi-modal endometrial cancer sample set, that is, the target data set. The sample set on the master control end is a small-sample data set. Training a model solely based on a small-sample data set is prone to overfitting, and the performance of the trained model is poor.

[0057] Compared with the prior art, the training system for the endometrial cancer target segmentation model with multiple data centers provided in this embodiment improves data security by building a master control end and multiple data centers, and the master control end and multiple data centers maintain their own sample sets. At the same time, the master control end builds a target segmentation model and sends it to each data center; the master control end controls the multiple data centers to pre-train the local target segmentation model based on the local sample set, and then updates the parameters of the target segmentation model on the master control end to realize the training of the target segmentation model on the master control end to obtain a pre-trained target segmentation model. The master control end then fine-tunes the pre-trained target segmentation model on the target data set, so as to quickly train a target segmentation model suitable for the target data set, and will not affect the performance of the model due to the small amount of data in the target data set, thus improving the segmentation performance of the model on the target data set.

[0058] Specifically, the sample set of each data center includes a training set and a test set, that is, a part of the samples in the sample set are divided into a training set, and a part of the samples are divided into a test set. The training set is used to train the local target segmentation model, and the test set is used to calculate the training loss of the local target segmentation model.

[0059] Each multi-modal endometrial cancer sample includes a pair of plain CT images and enhanced CT images of endometrial cancer. Some multi-modal endometrial cancer samples also include mask images corresponding to the plain CT images and enhanced CT images.

[0060] During implementation, due to the large workload and low efficiency of manual delineation, some samples in the sample set are allowed to have no labeled data, that is, there is no corresponding mask image. Therefore, the multi-modal endometrial cancer sample set includes paired plain CT images and enhanced CT images of endometrial cancer, as well as mask images corresponding to some paired plain CT images and enhanced CT images of endometrial cancer.

[0061] The target segmentation model can adopt a deep learning model.

[0062] Specifically, the target segmentation model includes a first segmentation module and a second segmentation module; the first segmentation module and the second segmentation module have the same structure;

[0063] The first segmentation module is used to perform target segmentation on the plain CT image to obtain a first segmentation image;

[0064] The second segmentation module is used to perform target segmentation on the enhanced CT image to obtain a first segmentation image.

[0065] The main control end sends the constructed target segmentation model to each data center, including the model structure and model parameters. Initially, the parameters of the target segmentation model at the main control end and the target segmentation models at each data center are the same.

[0066] During implementation, the first segmentation module and the second segmentation module can adopt a 3D-Unet structure.

[0067] Specifically, the main control end controls multiple data centers to train the local target segmentation model based on the local multi-modal endometrial cancer sample set in the following manner, and updates the parameters of the target segmentation model at the main control end based on the parameters of the local target segmentation models at multiple data centers to obtain a pre-trained target segmentation model:

[0068] S21. The main control end randomly selects N data centers and sends training instructions to the selected N data centers;

[0069] S22. After each selected data center receives the training instructions, it trains the local target segmentation model based on the local multi-modal endometrial cancer sample set, calculates the training loss of the local target segmentation model, and sends the calculated training loss to the main control end;

[0070] S23. The main control end updates the parameters of the local target segmentation model according to the training losses of the selected N data centers, and sends the updated parameters of the local target segmentation model to each data center; determines whether the current stop condition is reached. If so, stops training and obtains a pre-trained target segmentation model; otherwise, the main control end executes step S21.

[0071] During implementation, the master control end first randomly selects N data centers for model training in the current round, that is, it sends training instructions to the selected N data centers. At this time, the parameters of the target area segmentation model at the master control end and the target area segmentation models of each data center are the same, denoted as the initial parameter ф.

[0072] For each selected data center, after receiving the training instruction, it first trains the local target area segmentation model based on the training set in the local multi-modal endometrial cancer sample set and updates the parameters of the local target area segmentation model. During implementation, multi-step training can be performed, that is, the parameters of the local target area segmentation model are updated multiple times, and the parameters of the last update are denoted as θ, θ i represents the parameters of the local target area segmentation model of the i-th data center.

[0073] After updating the parameters of the local target area segmentation model, each selected data center calculates the training loss on the test set based on the updated local target area segmentation model.

[0074] During implementation, for the i-th data center, the samples on the test set are input into the local target area segmentation model (the model parameters are θ i ), and the training loss is calculated based on the first segmentation image and the second segmentation image output by the model.

[0075] Specifically, for the i-th data center, the following formula is used to calculate the training loss Loss based on the updated local target area segmentation model:

[0076] Loss = L 1 + L 2 + L 3 (1)

[0077] where, L 1 represents the multi-modal image segmentation loss, L 2 represents the multi-modal cross loss, L 3 represents the contrast loss.

[0078] Specifically, for the labeled samples, that is, the samples with corresponding mask images, the segmentation loss is calculated based on the first segmentation image and the second segmentation image obtained by the first segmentation module and the second segmentation module.

[0079] Specifically, the following formula is used to calculate the multi-modal image segmentation loss:

[0080]

[0081] where, represents the number of samples with corresponding mask images in the test set of the i-th data center, P k1Denote the first segmentation image of the k-th sample in the test set that has a corresponding mask image, G 1 Denote the mask image corresponding to the plain CT image of the k-th sample in the test set that has a corresponding mask image, P k2 Denote the second segmentation image of the k-th sample in the test set that has a corresponding mask image, G 2 Denote the mask image corresponding to the enhanced CT image of the k-th sample in the test set that has a corresponding mask image, α 1 and α 2 Denote the weight parameter.

[0082] During implementation, α 1 and α 2 are used to control false negatives and false positives respectively, and adjusting α 1 and α 2 can control the trade-off between false negatives and false positives. |P k1 ∩G 1 | represents the intersection of P k1 and G 1 .

[0083] For samples without labels, that is, without corresponding mask images, the loss is calculated through multi-modal mutual learning and contrast learning to obtain the multi-modal cross loss and the contrast loss.

[0084] Specifically, the following formula is used to calculate the multi-modal cross loss:

[0085]

[0086] where, represents the number of samples in the test set of the i-th data center that have no corresponding mask images, P k1 (s,j) represents the element value at the (s,j) position in the first segmentation image of the k-th sample that has no corresponding mask image, P k2 (s,j) represents the element value at the (s,j) position in the second segmentation image of the k-th sample that has no corresponding mask image, and m represents the number of elements in the first segmentation image.

[0087] Through the multi-modal cross loss, the images segmented by the first segmentation module and the second segmentation module are trained to be as similar as possible.

[0088] Specifically, the following method is used to calculate the contrast loss:

[0089] For each sample in the test set that has no corresponding mask image, use the first segmentation image and the second segmentation image of this sample as a positive sample pair; use the first segmentation image of this sample and the second segmentation image of other samples in the test set that have no corresponding mask image as negative sample pairs; obtain the positive sample pairs and negative sample pairs corresponding to each sample in the test set that has no corresponding mask image;

[0090] Calculate the contrastive loss based on the positive sample pairs and negative sample pairs corresponding to each sample in the test set that has no corresponding mask image.

[0091] For example, the number of samples in the test set of the i-th data center that have no corresponding mask image is pieces.

[0092] For the k-th sample without a corresponding mask image, take the first segmentation image P k1 and the second segmentation image P k2 predicted by the model as the positive sample pair (P k1 , P k2 ) corresponding to this sample. Then, form the negative sample pair corresponding to this sample by using the first segmentation image P k1 and the second segmentation images of other samples in the test set, (P k1 , P j2 ), j ≠ k.

[0093] Then, calculate the contrastive loss based on the positive sample pairs and negative sample pairs corresponding to each sample in the test set that has no corresponding mask image in the following way:

[0094]

[0095] where represents the number of samples in the test set of the i-th data center that have no corresponding mask image, P k1 represents the first segmentation image of the k-th sample without a corresponding mask image, P k2 represents the second segmentation image of the k-th sample without a corresponding mask image, P j2 represents the second segmentation image of the j-th sample without a corresponding mask image, τ represents the adjustment parameter, and dice(·, ·) represents the DICE loss.

[0096] Through the contrastive loss, the distance between positive samples becomes closer and the distance between negative samples becomes farther, thereby deepening the mutual learning between different modalities and improving the segmentation performance of the model.

[0097] After each selected data center calculates the training loss of the local target area segmentation model, send the calculated training loss to the master control end.

[0098] The master control end updates the parameters of the local target segmentation model according to the training losses of the selected N data centers, and sends the parameters of the updated local target segmentation model to each data center for synchronous update.

[0099] Specifically, the master control end updates the parameters of the local target segmentation model according to the training losses of the selected N data centers, including:

[0100] S231. The master control end calculates the pre-update parameters of each selected data center according to the training losses of the selected N data centers, and sends the pre-update parameters of each selected data center to the corresponding data center;

[0101] S232. Each selected data center pre-updates the local target segmentation model according to the pre-update parameters, calculates the pre-update loss based on the pre-updated local target segmentation model, and sends the pre-update loss to the master control end;

[0102] S233. The master control end calculates the weights of the N data centers based on the pre-update losses of the N data centers;

[0103] S234. Update the parameters of the local target segmentation model based on the training losses of the N data centers and the weights of the N data centers.

[0104] Specifically, the master control end calculates the pre-update parameters of each selected data center according to the training losses of the selected N data centers, including:

[0105] S2311. Calculate the first pre-update parameters of the N selected data centers based on the training losses of the N selected data centers;

[0106] S2312. For the i-th selected data center, obtain the second pre-update parameters of the i-th selected data center based on the training losses of the other N - 1 selected data centers;

[0107] The first pre-update parameters and the second pre-update parameters of the i-th selected data center constitute the pre-update parameters of the i-th selected data center.

[0108] Specifically, the first pre-update parameters are calculated using the following formula based on the losses of the N selected data centers:

[0109]

[0110] Among them, represents the gradient of the total loss L(φ), L(φ) represents the total loss of the N selected data centers, Loss(θ i ) represents the loss of the i-th selected data center, β represents the learning rate, φ represents the initial parameters, φ′1 Represents the first pre-update parameter.

[0111] For the i-th selected data center, the second pre-update parameter of the i-th selected data center obtained by the following formula based on the losses of the other N - 1 selected data centers:

[0112]

[0113] Wherein, Represents the total loss L i2 Gradient of L(φ), L i2 L(φ) represents the total loss of N - 1 selected data centers, Loss(θ j ) represents the loss of the i-th selected data center, β represents the learning rate, φ represents the initial parameter, and φ′ i2 Represents the second pre-update parameter of the i-th selected data center.

[0114] The first pre-update parameters in each selected data are the same.

[0115] The master control end sends the pre-update parameters corresponding to each selected data center to the corresponding selected data center. Each selected data center pre-updates the local target area segmentation model according to the pre-update parameters, calculates the pre-update loss based on the pre-updated local target area segmentation model, and sends the pre-update loss to the master control end.

[0116] Specifically, step S232 includes:

[0117] Each selected data center pre-updates the local target area segmentation model based on the first pre-update parameter and the second pre-update parameter respectively to obtain the first pre-update model and the second pre-update model;

[0118] Calculate the loss on the test set of the local multi-modal endometrial cancer sample set based on the first pre-update model to obtain the first pre-update loss;

[0119] Calculate the loss on the test set of the local multi-modal endometrial cancer sample set based on the second pre-update model to obtain the second pre-update loss;

[0120] The difference between the second pre-update loss and the first pre-update loss gives the pre-update loss.

[0121] During implementation, for the i-th selected data center, update the parameters of the local target area segmentation model to the first pre-update parameter to obtain the first pre-update model (parameters are φ′ 1 ), and update the parameters of the local target area segmentation model to the second pre-update parameter to obtain the second pre-update model (parameters are φ′ i2 ).

[0122] During implementation, the first pre-update loss is calculated based on the first pre-update model on the local test set, and the second pre-update loss is calculated based on the second pre-update model on the local test set. Both the first pre-update calculation and the second pre-update loss can be calculated according to the formula in Equation (1). The pre-update loss of the i-th selected data center is obtained by subtracting the first pre-update loss from the second pre-update loss.

[0123] The i-th selected data center sends the calculated pre-update loss to the master control end, and the master control end calculates the weights of the N data centers based on the pre-update losses of the N data centers.

[0124] During implementation, since the sample sizes and data qualities of different data centers are different, the weights of the losses of the N data centers for updating the initial parameters are different. By calculating the weights according to the data qualities of different data centers, the model parameters are dynamically calculated during each backpropagation, thereby increasing the weights of the data centers with high quality, and further improving the model segmentation accuracy and the model performance.

[0125] Specifically, the master control end calculates the weights of each selected data center based on the pre-update losses of each selected data center using the following formula:

[0126]

[0127] where q i represents the data quality of the i-th selected data center, δ i represents the pre-update loss of the i-th selected data center, and γ represents the adjustment parameter.

[0128] After obtaining the weights, the master control end updates the parameters of the local target area segmentation model based on the training losses of the N data centers and the weights of the N data centers using the following formula:

[0129]

[0130] where represents the gradient of the total training loss L(φ) of the N selected data centers, L(φ) represents the total training loss of the N selected data centers, Loss(θ i ) represents the training loss of the i-th selected data center, ω i represents the weight of the i-th selected data center, β represents the learning rate, and φ represents the parameters of the local target area segmentation model.

[0131] It should be noted that Loss(θ i ) represents the training loss of the i-th selected data center, which is the loss calculated in step S22, that is, the training loss calculated when the parameter is θ i .

[0132] After the master control end updates the parameters of the local target segmentation model, it sends the updated parameters of the local target segmentation model to each data center to update the parameters of the local target segmentation model in each data center, thus completing one backpropagation of the target segmentation model.

[0133] After completing one backpropagation of the target segmentation model, if the current stop condition is reached, the training is ended to obtain the pre-trained target segmentation model. If not, it returns to step S21 to continue the training until the stop condition of the model training is reached. In implementation, the stop condition can be reaching the preset number of training times or the total training loss reaching the preset accuracy.

[0134] After obtaining the pre-trained target segmentation model, the master control end fine-tunes the pre-trained target segmentation model on the target data set to quickly obtain the trained endometrial cancer target segmentation model. Inputting the multi-modal endometrial cancer image to be segmented on the master control end into the trained endometrial cancer target segmentation model can quickly and accurately obtain the target area image. It improves the accuracy and efficiency of target segmentation and avoids the overfitting risk of small samples.

[0135] Those skilled in the art can understand that all or part of the processes for implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory, or a random access memory, etc.

[0136] As mentioned above, the above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A multi-data center endometrial cancer target segmentation model training system, characterized in that: The system includes a main control terminal and multiple data centers: Each of the data centers stores a local multimodal endometrial cancer sample set; the main control terminal stores a target sample set; Constructing a target area segmentation model at the main control end, and sending the target area segmentation model to each of the data centers; The main control end controls the multiple data centers to train the local target segmentation model based on the local multimodal endometrial cancer sample set, and updates the parameters of the target segmentation model of the main control end based on the local target segmentation model parameters of the multiple data centers to obtain a pre-trained target segmentation model; The main control end fine-tunes the pre-trained target segmentation model based on the target sample set to obtain a trained endometrial cancer target segmentation model.

2. The training system for the multi-data center endometrial cancer target segmentation model according to claim 1, characterized in that: The main control end controls the multiple data centers to train the local target segmentation model based on the local multimodal endometrial cancer sample set, and updates the parameters of the target segmentation model of the main control end based on the local target segmentation model parameters of the multiple data centers to obtain a pre-trained target segmentation model: S21, the master control terminal randomly selects N data centers and sends training instructions to the selected N data centers; S22. After receiving the training instruction, each selected data center trains the local target segmentation model based on the local multimodal endometrial cancer sample set, calculates the training loss of the local target segmentation model, and sends the calculated training loss to the main control end; S23. The main control end updates the parameters of the local target area segmentation model according to the training losses of the selected N data centers, and sends the updated parameters of the local target area segmentation model to each data center; determines whether the current stop condition is met, and if so, stops training to obtain a pre-trained target area segmentation model; otherwise, the main control end executes step S21.

3. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 2, characterized in that: The master control end updates the parameters of the local target segmentation model according to the training losses of the selected N data centers, including: The master control end calculates the pre-update parameters of each selected data center according to the training losses of the selected N data centers, and sends the pre-update parameters of each selected data center to the corresponding data center; Each selected data center pre-updates the local target segmentation model according to the pre-update parameters, calculates the pre-update loss based on the pre-updated local target segmentation model, and sends the pre-update loss to the master control end; The master calculates the weights of N data centers based on their pre-update losses; The parameters of the local target segmentation model are updated based on the training losses of the N data centers and the weights of the N data centers.

4. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 3, characterized in that: The master end calculates the pre-update parameters of each selected data center based on the training losses of the selected N data centers, including: Obtaining first pre-updated parameters of the N selected data centers based on the training losses of the N selected data centers; For the ith selected data center, a second pre-updated parameter of the ith selected data center obtained based on the training losses of the other N-1 selected data centers; The first pre-updated parameter and the second pre-updated parameter of the i-th selected data center constitute the pre-updated parameter of the i-th selected data center.

5. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 3, characterized in that: Each selected data center pre-updates the local target segmentation model according to the pre-update parameters, and calculates the pre-update loss based on the pre-updated local target segmentation model, including: Each selected data center pre-updates the local target area segmentation model based on the first pre-update parameter and the second pre-update parameter to obtain a first pre-update model and a second pre-update model; The loss is calculated based on the first pre-updated model on the test set of the local multimodal endometrial cancer sample set to obtain the first pre-updated loss; The second pre-updated loss is obtained by calculating the loss on the test set of the local multimodal endometrial cancer sample set based on the second pre-updated model; The difference between the second pre-update loss and the first pre-update loss yields the pre-update loss.

6. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 3, characterized in that: The weight of each selected data center is calculated based on the pre-updated loss of N data centers using the following formula: Among them, q i represents the data quality of the i-th selected data center, δ i represents the pre-update loss of the i-th selected data center, and γ represents the adjustment parameter.

7. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 3, characterized in that: The multimodal endometrial cancer sample set includes a training set and a test set; the training set is used to train a local target segmentation model, and the test set is used to calculate the training loss of the local target segmentation model; The multimodal endometrial cancer samples include a pair of plain scan CT images and enhanced CT images of endometrial cancer; some multimodal endometrial cancer samples also include a pair of plain scan CT images and mask images corresponding to the enhanced CT images of endometrial cancer; The training loss is calculated using the following formula: Loss = L1 + L2 + L3 Among them, L1 represents multimodal image segmentation loss, L2 represents multimodal cross loss, and L3 represents contrast loss.

8. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 7, characterized in that: The multimodal cross loss is calculated using the following formula: in, represents the number of samples without corresponding mask images in the test set of the i-th data center, P k1 (s, j) represents the element value at position (s, j) in the first segmented image of the kth sample that has no corresponding mask image, P k2 (s, j) represents the element value at the position (s, j) in the second segmented image of the kth sample without a corresponding mask image, and m represents the number of elements in the first segmented image.

9. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 7, characterized in that: The contrast loss is calculated in the following way: For each sample in the test set without a corresponding mask image, the first segmented image and the second segmented image of the sample are used as a positive sample pair; the first segmented image of the sample and the second segmented image of other samples in the test set without a corresponding mask image are used as negative sample pairs; Get the positive sample pair and negative sample pair corresponding to each sample in the test set that does not have a corresponding mask image; The contrast loss is calculated based on the positive and negative pairs corresponding to each sample in the test set that does not have a corresponding mask image.

10. The training system for the target segmentation model of endometrial cancer in multiple data centers according to claim 9, characterized in that: Based on the positive and negative pairs of each sample in the test set without a corresponding mask image, the contrast loss is calculated in the following way: in, represents the number of samples without corresponding mask images in the test set of the i-th data center, P k1 represents the first segmented image of the kth sample without a corresponding mask image, P k2 represents the second segmented image of the kth sample without a corresponding mask image, P j2 represents the second segmented image of the jth sample without a corresponding mask image, τ represents the adjustment parameter, and dice(·,·) represents the DICE loss.

Images