Training method of endometrial cancer postoperative CTV segmentation model

Abstract

The invention relates to a training method of an endometrial cancer postoperative CTV segmentation model, belongs to the technical field of target region segmentation, and solves the problem of low segmentation accuracy in the prior art. The method comprises the following steps: constructing a plurality of data centers, wherein each data center comprises a multi-modal segmentation model and a multi-modal endometrial cancer postoperative CTV sample set; pre-training the multi-modal segmentation model based on the plurality of data centers to obtain a pre-trained endometrial cancer postoperative CTV segmentation model; screening style features based on the multi-modal endometrial cancer postoperative CTV sample set of the multiple data centers; and for a target data center, performing fine tuning on the pre-trained endometrial cancer postoperative CTV segmentation model based on the center style features to obtain a trained endometrial cancer postoperative CTV segmentation model. And efficient and accurate endometrial cancer postoperative CTV segmentation is realized.

Description

Technical Field

[0001] The invention relates to the technical field of target area delineation, and in particular to a training method for a CTV segmentation model after endometrial cancer surgery. Background Art

[0002] The morbidity and mortality of endometrial cancer (EC) are increasing year by year, and radiotherapy (VBT) has become one of the best radiotherapy methods for EC after surgery. Using CT or MR images to outline the clinical target volume (CTV) is very important for the accuracy of dose distribution and reducing the patient's acute radiation reaction. Accurate target segmentation can ensure that treatment methods such as radiotherapy act accurately on the tumor area, reduce damage to surrounding normal tissues, improve the patient's quality of life and improve prognosis.

[0003] Existing deep neural network models for medical image segmentation, such as nnUNet, are suitable for the segmentation of most organs, but the model performance is usually poor for small sample data sets and difficult clinical target areas without obvious boundaries. If there are data sets from other centers for the same task, pre-training can be performed on other centers, and then fine-tuned on the small sample data set through transfer learning.

[0004] However, different data centers have different styles. The existing technology does not consider the styles of different data centers when training the postoperative CTV segmentation model for endometrial cancer, so the trained model has low accuracy. Summary of the invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a training method for a postoperative CTV segmentation model for endometrial cancer, so as to solve the problem of low segmentation accuracy in the prior art.

[0006] On the one hand, an embodiment of the present invention provides a training method for a postoperative CTV segmentation model for endometrial cancer, comprising the following steps: Build multiple data centers, each of which includes a multimodal segmentation model and a multimodal endometrial cancer postoperative CTV sample set; Pre-training the multimodal segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer postoperative CTV segmentation model; Screening style features of postoperative CTV sample sets for multimodal endometrial cancer based on multiple data centers; For the target data center, the pre-trained endometrial cancer postoperative CTV segmentation model is fine-tuned based on the central style features to obtain a trained endometrial cancer postoperative CTV segmentation model.

[0007] Based on the further improvement of the above method, the samples of the multimodal endometrial cancer postoperative CTV dataset include paired plain scan CT images and enhanced CT images of endometrial cancer; some samples also include mask images corresponding to the plain scan CT images and enhanced CT images; The multimodal postoperative CTV sample set for endometrial cancer based on multiple data centers uses the following method to screen style features: For each data center's multimodal postoperative CTV sample set for endometrial cancer, samples with corresponding mask images are extracted, and the plain scan CT image and enhanced CT image of the sample are intersected with the corresponding mask image to obtain the plain scan CTV internal image and the enhanced CTV internal image; the plain scan CTV internal image and the enhanced CTV internal image are spliced ​​to obtain the sample's CTV internal image; Calculate multiple radiomic features of images within each CTV; For each data center, a matrix of each radiomics feature of each CTV in the data center is constructed based on each radiomics feature of the image in the CTV of the data center. The style representativeness of each radiomics feature is calculated based on each radiomics feature matrix of each data center; A preset number of radiomics features with greater style representativeness are selected as style features.

[0008] Based on the further improvement of the above method, the following formula is used to calculate the style representativeness of each radiomics feature: ; in, Indicates The style representativeness of the radiomics features, N represents the number of data centers, represents the KL divergence, represents the first A radiomics feature matrix, represents the jth data center A radiomics feature matrix.

[0009] Based on the further improvement of the above method, for the target data center, the pre-trained postoperative endometrial cancer CTV segmentation model is fine-tuned based on the center style feature, including: The plain scan CT image and enhanced CT image of the training sample in the target data center are respectively intersected with the corresponding mask image to obtain the plain scan CTV image and the enhanced CTV image; the plain scan CTV image and the enhanced CTV image are spliced ​​to obtain the real CTV image of the sample; The plain scan CT image and the enhanced CT image of the training sample are input into the pre-trained endometrial cancer postoperative CTV segmentation model to obtain a first segmented image and a second segmented image; the plain scan CT image and the first segmented image are intersected to obtain a predicted plain scan CTV internal image, the enhanced CT image and the second segmented image are intersected to obtain a predicted enhanced CTV internal image, and the predicted plain scan CTV internal image and the predicted enhanced CTV internal image are spliced ​​to obtain a predicted CTV internal image of the sample; The loss is calculated based on the real CTV image and the predicted CTV image of the sample, the parameters of the pre-trained endometrial cancer postoperative CTV segmentation model are updated, and the pre-trained endometrial cancer postoperative CTV segmentation model is fine-tuned.

[0010] Based on the further improvement of the above method, the sample set includes a support set and a query set; the multimodal segmentation model is pre-trained based on the multiple data centers to obtain a pre-trained endometrial cancer postoperative CTV segmentation model, including: S21. Randomly select N data centers and record the parameters of the current multimodal segmentation model as initial parameters; S22. For each selected data center, training a local multimodal segmentation model based on a local support set, and updating parameters of the local multimodal segmentation model; S23. For each selected data center, calculate the loss on the local query set based on the updated local multimodal segmentation model; S24, updating the initial parameters based on the loss of each selected data center; updating the parameters of the multimodal segmentation model of each data center to the initial parameters; S25, judging whether the stopping condition is met, if so, stopping the training to obtain the pre-trained postoperative CTV segmentation model for endometrial cancer, otherwise, returning to step S21.

[0011] Based on the further improvement of the above method, the following formula is used to update the initial parameters based on the loss of each selected data center: ; ; in, represents the total loss of N selected data centers, Represents the total loss of N selected data centers The gradient of represents the loss of the i-th selected data center, represents the weight of the i-th selected data center, β represents the learning rate, Represents the initial parameters.

[0012] Based on the further improvement of the above method, the loss is calculated using the following formula: ; in, represents the multimodal image segmentation loss, represents the multimodal cross loss, represents contrast loss.

[0013] Based on the further improvement of the above method, the multimodal cross loss is calculated using the following formula: ; in, Indicates the number of samples that do not have corresponding mask images in the query set of the i-th data center, In the first segmented image, the kth sample without a corresponding mask image is represented The element value at position, In the second segmented image, the kth sample without a corresponding mask image is represented The element value of the position, m represents the number of elements of the first segmented image.

[0014] Based on the further improvement of the above method, the contrast loss is calculated in the following way: For each sample in the query 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 query set without a corresponding mask image are used as a negative sample pair; the positive sample pair and the negative sample pair corresponding to each sample in the query set without a corresponding mask image are obtained; The contrast loss is calculated based on the positive and negative pairs corresponding to each sample in the query set that does not have a corresponding mask image.

[0015] Based on the further improvement of the above method, the contrast loss is calculated based on the positive sample pairs and negative sample pairs corresponding to each sample in the query set without a corresponding mask image in the following way: ; in, Indicates the number of samples that do not have corresponding mask images in the query set of the i-th data center, represents the first segmented image of the kth sample without a corresponding mask image, represents the second segmented image of the kth sample without a corresponding mask image, 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.

[0016] Compared with the prior art, the present invention constructs multiple data centers, pre-trains the multimodal segmentation model based on the multiple data centers, obtains the pre-trained endometrial cancer postoperative CTV segmentation model, and screens the style features based on the multimodal endometrial cancer postoperative CTV sample set of the multiple data centers, thereby fine-tuning the pre-trained CTV segmentation model on the target data center based on the features that best represent different data centers, so that the model is more adapted to the style of the target data center, and the segmentation accuracy of the model on the target data center is improved. In addition, the multiple data centers jointly train the segmentation model instead of mixing the data together for training, so that the sample distribution of each data center is not required to be consistent. By obtaining the pre-trained endometrial cancer postoperative CTV segmentation model, and then fine-tuning it on the target data set, the CTV segmentation model applicable to the target data set is quickly trained, and the performance of the model is not affected by the small amount of data in the target data set, thereby improving the segmentation performance of the model on the target data center. For the target data center, the multimodal endometrial cancer image to be segmented is segmented based on the trained endometrial cancer postoperative CTV segmentation model, and the CTV segmentation result of the multimodal endometrial cancer image to be segmented can be quickly and accurately obtained, thereby improving the efficiency of CTV segmentation.

[0017] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components; Figure 1 The flowchart is a method for training a postoperative CTV segmentation model for endometrial cancer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0020] A specific embodiment of the present invention discloses a training method for a postoperative CTV segmentation model for endometrial cancer. Figure 1 As shown, the following steps are included: S1. Build multiple data centers, each of which includes a multimodal segmentation model and a multimodal endometrial cancer postoperative CTV sample set; S2. Pre-training the multimodal segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer target area postoperative CTV segmentation model; S3, style features of postoperative CTV sample set for multimodal endometrial cancer based on multiple data centers; S4. For the target data center, fine-tune the pre-trained endometrial cancer postoperative CTV segmentation model based on the central style feature to obtain a trained endometrial cancer postoperative CTV segmentation model.

[0021] During implementation, each medical center collects multimodal image data of postoperative endometrial cancer patients in the center to construct a multimodal endometrial cancer postoperative CTV sample set. At the same time, each center is equipped with a multimodal segmentation model with the same structure and initial parameters to build a data center, that is, each data center has a local multimodal segmentation model and a local sample set.

[0022] During implementation, the target data center is a data set of a new data center. The data in the target data set center is a small sample data set. The training model of a small sample data set is prone to overfitting, and the performance of the trained model is poor.

[0023] Compared with the prior art, the training method of the postoperative CTV segmentation model for endometrial cancer provided in this embodiment constructs multiple data centers, pre-trains the multimodal segmentation model based on the multiple data centers, obtains the pre-trained postoperative CTV segmentation model for endometrial cancer, and screens the style features based on the multimodal postoperative CTV sample set for endometrial cancer based on the multiple data centers, thereby fine-tuning the pre-trained CTV segmentation model on the target data center based on the features that best represent different data centers, so that the model is more adapted to the style of the target data center, and the segmentation accuracy of the model on the target data center is improved. In addition, the segmentation model is trained jointly by multiple data centers instead of mixing the data together for training, so that the sample distribution of each data center is not required to be consistent. By obtaining the pre-trained postoperative CTV segmentation model for endometrial cancer, and then fine-tuning it on the target data set, the CTV segmentation model applicable to the target data set is quickly trained, and the performance of the model is not affected by the small amount of data in the target data set, thereby improving the segmentation performance of the model on the target data center. For the target data center, the multimodal endometrial cancer images to be segmented are segmented based on the trained postoperative CTV segmentation model for endometrial cancer, so that the CTV segmentation results of the multimodal endometrial cancer images to be segmented can be obtained quickly and accurately, thereby improving the efficiency of CTV segmentation.

[0024] Specifically, the postoperative CTV sample set of each data center includes a support set and a query set, that is, a part of the samples in the sample set is divided into a support set, and a part of the samples is divided into a query set.

[0025] Specifically, pre-training the multimodal segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer postoperative CTV segmentation model includes: S21. Randomly select N data centers and record the parameters of the current multimodal segmentation model as initial parameters; S22. For each selected data center, training a local multimodal segmentation model based on a local support set, and updating parameters of the local multimodal segmentation model; S23. For each selected data center, calculate the loss on the local query set based on the updated local multimodal segmentation model; S24, updating the initial parameters based on the loss of each selected data center; updating the parameters of the multimodal segmentation model of each data center to the initial parameters; S25, judging whether the stopping condition is met, if so, stopping the training to obtain the pre-trained postoperative CTV segmentation model for endometrial cancer, otherwise, returning to step S21.

[0026] During implementation, N data centers are randomly selected to perform the current round of model training. The parameters of the current multimodal segmentation model of each data center are the same, marked as initial parameters ф.

[0027] For each selected data center, the local multimodal segmentation model is first trained based on the support set in the local data set, and the parameters of the local multimodal segmentation model are updated. In implementation, multi-step training can be performed, that is, the parameters of the local multimodal segmentation model are updated multiple times, and the parameters of the last update are recorded as , represents the parameters of the local multimodal segmentation model of the i-th data center.

[0028] After updating the parameters of the local multimodal segmentation model, each selected data center calculates the loss on the local query set based on the updated local multimodal segmentation model.

[0029] The samples of the multimodal postoperative CTV dataset for endometrial cancer include paired plain scan CT images and enhanced CT images of endometrial cancer, and some samples also include mask images corresponding to the plain scan CT images and enhanced CT images; The multimodal segmentation model comprises a first segmentation module and a second segmentation module; the first segmentation module and the second segmentation module have the same structure; The first segmentation module is used to perform target area segmentation on the plain scan CT image to obtain a first segmented image; The second segmentation module is used to perform target area segmentation on the enhanced CT image to obtain a first segmented image.

[0030] During implementation, due to the large workload and low efficiency of manual delineation, it is allowed that some samples in the sample set have no labeled data, that is, no corresponding mask images. Therefore, the samples of the multimodal endometrial cancer sample set include paired plain scan CT images and enhanced CT images of endometrial cancer, as well as mask images corresponding to some paired plain scan CT images and enhanced CT images of endometrial cancer.

[0031] During implementation, the first segmentation model and the second segmentation model may adopt an existing image segmentation model, such as a Unet model.

[0032] In implementation, for the i-th data center, the samples on the local query set are input into the local multimodal segmentation model (the model parameters are ), the loss is calculated based on the first segmentation image and the second segmentation image output by the model.

[0033] Specifically, for the i-th data center, the loss is calculated on the local query set based on the updated local multimodal segmentation model. The loss is calculated using the following formula: ; in, represents the multimodal image segmentation loss, represents the multimodal cross loss, represents contrast loss.

[0034] Specifically, for labeled samples, that is, samples with corresponding mask images, the multimodal 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.

[0035] Specifically, the multimodal image segmentation loss is calculated using the following formula: ; in, Indicates the number of samples with corresponding mask images in the query set of the i-th data center, represents the first segmented image of the kth sample in the query set that has a corresponding mask image, represents the mask image corresponding to the plain scan CT image of the kth sample with a corresponding mask image in the query set, represents the second segmented image of the kth sample in the query set that has a corresponding mask image, represents the mask image corresponding to the enhanced CT image of the kth sample with a corresponding mask image in the query set, and Represents the weight parameter.

[0036] When implementing, and Used to control false negatives and false positives, respectively, and adjust and The trade-off between false negatives and false positives can be controlled. express and The intersection of .

[0037] For samples without labels, that is, without corresponding mask images, the loss is calculated through multimodal mutual learning and contrastive learning to obtain multimodal cross loss and contrastive loss.

[0038] Specifically, the multimodal cross loss is calculated using the following formula: ; in, Indicates the number of samples that do not have corresponding mask images in the query set of the i-th data center, In the first segmented image, the kth sample without a corresponding mask image is represented The element value at position, In the second segmented image, the kth sample without a corresponding mask image is represented The element value of the position, m represents the number of elements of the first segmented image.

[0039] Through multimodal cross loss, the images segmented by the first segmentation module and the second segmentation module are trained to be as similar as possible.

[0040] Specifically, the contrast loss is calculated in the following manner: For each sample in the query 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 query set without a corresponding mask image are used as a negative sample pair; the positive sample pair and the negative sample pair corresponding to each sample in the query set without a corresponding mask image are obtained; The contrast loss is calculated based on the positive and negative pairs corresponding to each sample in the query set that does not have a corresponding mask image.

[0041] For example, there are samples without corresponding mask images in the query set of the i-th data center. indivual.

[0042] For the kth sample without a corresponding mask image, the first segmentation image predicted by the model is and the second segmented image As the positive sample pair corresponding to this sample Then, the first segmented image The second segmented image of other samples in the query set that do not have corresponding mask images constitutes a negative sample pair corresponding to the sample. .

[0043] Then, based on the positive and negative pairs corresponding to each sample in the query set that does not have a corresponding mask image, the contrast loss is calculated in the following way: ; in, Indicates the number of samples that do not have corresponding mask images in the query set of the i-th data center, represents the first segmented image of the kth sample without a corresponding mask image, represents the second segmented image of the kth sample without a corresponding mask image, 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.

[0044] Through contrast loss, the distance between positive samples is made closer and the distance between negative samples is made farther, thereby deepening the mutual learning between different modalities and improving the segmentation performance of the model.

[0045] After calculating the losses in the N data, the initial parameters are updated based on the losses in the N data centers.

[0046] During implementation, due to the different sample sizes and data qualities of different data centers, the losses of N data centers have different weights on the initial parameter update.

[0047] Specifically, the weight of each selected data center is obtained in the following manner: S241, pre-updating parameters of the multimodal segmentation model of the N selected data centers based on the losses of the N selected data centers; S242. For each selected data center, calculate the marginal loss on the local query set based on the pre-updated local multimodal segmentation model; S243. Calculate the weight of each selected data center based on the marginal loss of each selected data center.

[0048] During implementation, the parameters of the multimodal segmentation model of the N selected data centers are pre-updated based on the losses of the N selected data centers, including: Obtaining a first pre-update parameter based on the loss calculation 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 losses of the other N-1 selected data centers; For the i-th selected data center, the local multimodal segmentation model is pre-updated 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.

[0049] Specifically, the first pre-update parameter is calculated based on the losses of the N selected data centers using the following formula: ; ; in, represents the total loss of N selected data centers, Indicates total loss The gradient of represents the loss of the i-th selected data center, β represents the learning rate, represents the initial parameters, Represents the first pre-update parameter.

[0050] Specifically, the second pre-updated parameter of the ith selected data center is obtained by using the following formula based on the losses of the other N-1 selected data centers: ; ; in, represents the total loss of N-1 selected data centers, Indicates total loss The gradient of represents the loss of the i-th selected data center, β represents the learning rate, represents the initial parameters, represents the second pre-updated parameter of the i-th selected data center.

[0051] During implementation, the parameters of the local multimodal segmentation model of the i-th selected data center are updated to the first pre-updated parameters to obtain the first pre-updated model (parameters are ), the parameters of the local multimodal segmentation model of the i-th selected data center are updated to the second pre-updated parameters to obtain the second pre-updated model (parameters are ).

[0052] Specifically, for the i-th selected data center, the marginal loss is calculated on the local query set based on the pre-updated local multimodal segmentation model, including: Calculate the loss on the local query set based on the first pre-updated model to obtain a first loss; Calculate the loss on the local query set based on the second pre-updated model to obtain a second loss; The difference between the second loss and the first loss gives the marginal loss.

[0053] During implementation, the first loss and the second loss can be calculated according to the formula (1).

[0054] Specifically, the weight of each selected data center is calculated based on the marginal loss of each selected data center using the following formula: ; ; in, represents the data quality of the i-th selected data center, represents the marginal loss of the i-th selected data center, and γ represents the adjustment parameter.

[0055] After obtaining the weights, the initial parameters are updated based on the losses of the N data centers using the following formula: ; ; in, represents the total loss of N selected data centers, Represents the total loss of N selected data centers The gradient of represents the loss of the i-th selected data center, represents the weight of the i-th selected data center, β represents the learning rate, Represents the initial parameters.

[0056] It should be noted that represents the loss of the i-th selected data center, which is the loss calculated in step S23, i.e., in the parameter Calculate the loss when .

[0057] By calculating weights based on the data quality of different data centers, the model parameters are dynamically calculated each time back propagation occurs, thereby increasing the weights of data centers with high quality, thereby improving the model segmentation accuracy and improving model performance.

[0058] After the initial parameters are updated, the parameters of the local multimodal segmentation models of all data centers are updated to the initial parameters, thereby completing a back propagation of the multimodal segmentation model.

[0059] After completing the back propagation of the multimodal segmentation model, if the current stop condition is met, the training is terminated to obtain the pre-trained endometrial cancer postoperative CTV segmentation model. If not, return to step S21 to continue training until the stop condition of model training is met. During implementation, the stop condition can be reaching a preset number of training times or the total loss reaching a preset accuracy.

[0060] Since the image styles of different data centers are different, by screening the features that can represent the styles of different data centers, we can more accurately guide the fine-tuning of the pre-trained model, making the obtained model more adaptable to the new data center. Therefore, style features are screened based on the multimodal postoperative CTV sample set of endometrial cancer from multiple data centers.

[0061] Specifically, the multimodal postoperative CTV sample set of endometrial cancer based on multiple data centers uses the following method to screen style features: S31. For each data center's multimodal postoperative endometrial cancer CTV sample set, extract samples with corresponding mask images, intersect the plain scan CT image and enhanced CT image of the sample with the corresponding mask image, respectively, to obtain a plain scan CTV internal image and an enhanced CTV internal image; splice the plain scan CTV internal image and the enhanced CTV internal image to obtain a CTV internal image of the sample; S32, calculating multiple radiomics features of the images in each CTV; S33. For each data center, based on each radiomics feature of the image in each CTV of the data center, construct each radiomics feature matrix of the data center; S34, calculating the style representativeness of each radiomics feature based on each radiomics feature matrix of each data center; S35. Select a preset number of radiomics features with greater style representativeness as style features.

[0062] During implementation, for each data center, samples with corresponding mask images are first extracted, and the plain scan CT image and enhanced CT image of the sample are intersected with the corresponding mask image respectively to obtain the plain scan CTV image and the enhanced CTV image; the plain scan CTV image and the enhanced CTV image are spliced ​​to obtain the CTV image of the sample.

[0063] Then, multiple radiomics features of the image in each CTV are calculated. In implementation, the multiple radiomics features include energy, entropy, maximum value, minimum value, mean value, gray level co-occurrence matrix, gray level run length matrix, etc. of the image, which are not limited in the present invention. In implementation, a pyradiomics algorithm can be used to calculate multiple radiomics features.

[0064] For each data center, each radiomics feature matrix of the data center is constructed based on each radiomics feature of the images in each CTV of the data center.

[0065] For example, for the i-th data center, the first image in each CTV of the data center The radiomics feature was used as the first A row element of the radiomics feature matrix is ​​used to construct the first A radiomics feature matrix.

[0066] Then, the style representativeness of each radiomics feature is calculated based on each radiomics feature matrix in each data center.

[0067] Specifically, the style representativeness of each radiomics feature is calculated using the following formula: ; in, Indicates The style representativeness of the radiomics features, N represents the number of data centers, represents the KL divergence, represents the first A radiomics feature matrix, represents the jth data center A radiomics feature matrix.

[0068] During implementation, the KL divergence is used to calculate the The distribution difference of radiomics features in any two data centers is the greater the distribution difference of features in different data centers, the more it can represent the styles of different data centers.

[0069] According to the order of style representativeness from large to small, a preset number of radiomics features with large style representativeness are selected as style features. During implementation, the preset number is set according to the computational efficiency requirements.

[0070] For a new medical center, a small number of multimodal postoperative endometrial cancer CTV samples can be collected to construct a fine-tuned training sample set, and the pre-trained postoperative endometrial cancer CTV segmentation model can be fine-tuned based on the center style features to quickly obtain a trained postoperative endometrial cancer CTV segmentation model. The multimodal postoperative endometrial cancer multimodal images to be segmented on the center are input into the trained postoperative endometrial cancer CTV segmentation model to quickly and accurately obtain the target area image.

[0071] For the target data center, the pre-trained postoperative endometrial cancer CTV segmentation model is fine-tuned based on the central style features, including: S41, taking intersections of the plain scan CT image and the enhanced CT image of the training sample in the target data center with the corresponding mask image, respectively, to obtain a plain scan CTV in-image and an enhanced CTV in-image; splicing the plain scan CTV in-image and the enhanced CTV in-image to obtain a real CTV in-image of the sample; S42, inputting the plain scan CT image and the enhanced CT image of the training sample into the pre-trained endometrial cancer postoperative CTV segmentation model to obtain a first segmented image and a second segmented image; taking the intersection of the plain scan CT image and the first segmented image to obtain a predicted plain scan CTV internal image, taking the intersection of the enhanced CT image and the second segmented image to obtain a predicted enhanced CTV internal image, and splicing the predicted plain scan CTV internal image and the predicted enhanced CTV internal image to obtain a predicted CTV internal image of the sample; S43, calculating the loss based on the real CTV image and the predicted CTV image of the sample, updating the parameters of the pre-trained endometrial cancer postoperative CTV segmentation model, and fine-tuning the pre-trained endometrial cancer postoperative CTV segmentation model.

[0072] During implementation, the real CTV image of the training sample of the target data center is obtained in the manner of the aforementioned step S31; For each sample, the first segmentation image and the second segmentation image are obtained by the pre-trained postoperative CTV segmentation model for endometrial cancer, that is, the target area mask images corresponding to the plain scan CT image and the enhanced CT image, respectively. In the same way, the predicted CTV image of the sample is obtained.

[0073] The loss is calculated based on the real CTV image and the predicted CTV image of the sample, the parameters of the pre-trained endometrial cancer postoperative CTV segmentation model are updated, and the pre-trained endometrial cancer postoperative CTV segmentation model is fine-tuned. During implementation, an existing loss function, such as a cross entropy loss function, can be used.

[0074] During implementation, the model parameters can be updated once according to the loss of one training sample, or the parameters can be updated once according to the loss of a batch of training samples.

[0075] After reaching the stopping condition of the model training, the training is stopped to obtain a trained endometrial cancer postoperative CTV segmentation model. During implementation, the stopping condition may be reaching a preset number of training times, or the loss reaching a preset accuracy.

[0076] By fine-tuning based on radiomics features to better adapt to the style of the new data center, the accuracy of the postoperative CTV segmentation model for endometrial cancer on the new data center was improved.

[0077] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0078] The above description 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 conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A training method for a postoperative CTV segmentation model for endometrial cancer, characterized in that: The following steps are involved: Build multiple data centers, each of which includes a multimodal segmentation model and a multimodal endometrial cancer postoperative CTV sample set; Pre-training the multimodal segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer postoperative CTV segmentation model; Screening style features of multimodal postoperative CTV sample sets for endometrial cancer based on multiple data centers; For the target data center, the pre-trained endometrial cancer postoperative CTV segmentation model is fine-tuned based on the central style features to obtain a trained endometrial cancer postoperative CTV segmentation model.

2. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 1, characterized in that: The samples of the multimodal postoperative CTV dataset for endometrial cancer include paired plain scan CT images and enhanced CT images of endometrial cancer; some samples also include mask images corresponding to the plain scan CT images and enhanced CT images; The multimodal postoperative CTV sample set for endometrial cancer based on multiple data centers uses the following method to screen style features: For each data center's multimodal postoperative CTV sample set for endometrial cancer, samples with corresponding mask images are extracted, and the plain scan CT image and enhanced CT image of the sample are intersected with the corresponding mask image to obtain the plain scan CTV internal image and the enhanced CTV internal image; the plain scan CTV internal image and the enhanced CTV internal image are spliced ​​to obtain the sample's CTV internal image; Calculate multiple radiomic features of images within each CTV; For each data center, a matrix of each radiomics feature of each CTV in the data center is constructed based on each radiomics feature of the image in the CTV of the data center. The style representativeness of each radiomics feature is calculated based on each radiomics feature matrix of each data center; A preset number of radiomics features with greater style representativeness are selected as style features.

3. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 2, characterized in that: The style representativeness of each radiomics feature is calculated using the following formula: ； in, Indicates The style representativeness of the radiomics features, N represents the number of data centers, represents the KL divergence, represents the first A radiomics feature matrix, represents the jth data center A radiomics feature matrix.

4. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 1, characterized in that: For the target data center, the pre-trained postoperative endometrial cancer CTV segmentation model is fine-tuned based on the central style features, including: The plain scan CT image and enhanced CT image of the training sample in the target data center are respectively intersected with the corresponding mask image to obtain the plain scan CTV image and the enhanced CTV image; the plain scan CTV image and the enhanced CTV image are spliced ​​to obtain the real CTV image of the sample; The plain scan CT image and the enhanced CT image of the training sample are input into the pre-trained endometrial cancer postoperative CTV segmentation model to obtain a first segmented image and a second segmented image; the plain scan CT image and the first segmented image are intersected to obtain a predicted plain scan CTV internal image, the enhanced CT image and the second segmented image are intersected to obtain a predicted enhanced CTV internal image, and the predicted plain scan CTV internal image and the predicted enhanced CTV internal image are spliced ​​to obtain a predicted CTV internal image of the sample; The loss is calculated based on the real CTV image and the predicted CTV image of the sample, the parameters of the pre-trained endometrial cancer postoperative CTV segmentation model are updated, and the pre-trained endometrial cancer postoperative CTV segmentation model is fine-tuned.

5. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 1, characterized in that: The sample set includes a support set and a query set; the multimodal segmentation model is pre-trained based on the multiple data centers to obtain a pre-trained endometrial cancer postoperative CTV segmentation model, including: S21. Randomly select N data centers and record the parameters of the current multimodal segmentation model as initial parameters; S22. For each selected data center, training a local multimodal segmentation model based on a local support set, and updating parameters of the local multimodal segmentation model; S23. For each selected data center, calculate the loss on the local query set based on the updated local multimodal segmentation model; S24, updating the initial parameters based on the loss of each selected data center; updating the parameters of the multimodal segmentation model of each data center to the initial parameters; S25, judging whether the stopping condition is met, if so, stopping the training to obtain the pre-trained postoperative CTV segmentation model for endometrial cancer, otherwise, returning to step S21.

6. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 5, characterized in that: The initial parameters are updated based on the loss of each selected data center using the following formula: ； ； in, represents the total loss of N selected data centers, Represents the total loss of N selected data centers The gradient of represents the loss of the i-th selected data center, represents the weight of the i-th selected data center, β represents the learning rate, Represents the initial parameters.

7. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 5, characterized in that: The loss is calculated using the following formula: ； in, represents the multimodal image segmentation loss, represents the multimodal cross loss, represents contrast loss.

8. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 7, characterized in that: The multimodal cross loss is calculated using the following formula: ； in, Indicates the number of samples that do not have corresponding mask images in the query set of the i-th data center, In the first segmented image, the kth sample without a corresponding mask image is represented The element value at position, In the second segmented image, the kth sample without a corresponding mask image is represented The element value of the position, m represents the number of elements of the first segmented image.

9. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 7, characterized in that: The contrast loss is calculated in the following way: For each sample in the query 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 query set without a corresponding mask image are used as negative sample pairs; Get the positive sample pairs and negative sample pairs corresponding to each sample in the query set that has no corresponding mask image; The contrast loss is calculated based on the positive and negative pairs corresponding to each sample in the query set that does not have a corresponding mask image.

10. The training method for the postoperative CTV segmentation model for endometrial cancer according to claim 9, characterized in that: The contrast loss is calculated based on the positive and negative pairs corresponding to each sample in the query set that does not have a corresponding mask image in the following way: ； in, Indicates the number of samples that do not have corresponding mask images in the query set of the i-th data center, represents the first segmented image of the kth sample without a corresponding mask image, represents the second segmented image of the kth sample without a corresponding mask image, 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