Small sample endometrial cancer target region sketching method

Abstract

The invention relates to a small sample endometrial cancer target region sketching method, belongs to the technical field of target region sketching, and solves the problem of low segmentation accuracy caused by a small sample data training model in the prior art. The method comprises the following steps: constructing a plurality of data centers, wherein each data center comprises a multi-channel segmentation model and a multi-modal endometrial cancer sample set; pre-training the multi-channel segmentation model based on the plurality of data centers to obtain a pre-trained endometrial cancer target region segmentation model; performing fine adjustment on the pre-trained endometrial cancer target region segmentation model on a target data set to obtain a trained endometrial cancer target region segmentation model; and segmenting a to-be-segmented multi-modal endometrial cancer image based on the trained endometrial cancer target region segmentation model to obtain a target region sketching result of the to-be-segmented multi-modal endometrial cancer image. 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 delineation, and in particular, to a method for delineating the target area of endometrial cancer with small samples. 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, the target segmentation after surgery is of crucial significance 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] Training a model solely with data collected from a certain medical center is prone to overfitting due to the small sample size, and thus cannot accurately perform target segmentation. Summary of the Invention

[0004] In view of the above analysis, the embodiments of the present invention aim to provide a method for delineating the target area of endometrial cancer with small samples to solve the problem of low segmentation accuracy caused by training models with existing small sample data.

[0005] On the one hand, the embodiments of the present invention provide a method for delineating the target area of endometrial cancer with small samples, including the following steps:

[0006] Construct a plurality of data centers, each data center including a multi-channel segmentation model and a multi-modal endometrial cancer sample set;

[0007] Pre-train the multi-channel segmentation model based on the plurality of data centers to obtain a pre-trained endometrial cancer target segmentation model;

[0008] Fine-tune the pre-trained endometrial cancer target segmentation model on a target data set to obtain a trained endometrial cancer target segmentation model;

[0009] Segment the multi-modal endometrial cancer image to be segmented based on the trained endometrial cancer target segmentation model to obtain the target delineation result of the multi-modal endometrial cancer image to be segmented.

[0010] Based on a further improvement of the above method, the sample set includes a support set and a query set; pre-training the multi-channel segmentation model based on the plurality of data centers to obtain a pre-trained endometrial cancer target segmentation model includes:

[0011] S21. Randomly select N data centers, and record the parameters of the current multi-channel segmentation model as the initial parameters;

[0012] S22. For each selected data center, train the local multi-channel segmentation model based on the local support set and update the parameters of the local multi-channel segmentation model;

[0013] S23. For each selected data center, calculate the loss on the local query set based on the updated local multi-channel segmentation model;

[0014] S24. Update the initial parameters based on the losses of each selected data center; update the parameters of the multi-channel segmentation model of each data center to the initial parameters;

[0015] S25. Determine whether the stopping condition is reached. If so, stop training to obtain a pre-trained endometrial cancer target segmentation model. Otherwise, return to step S21.

[0016] Based on a further improvement of the above method, the following formula is used to update the initial parameters based on the losses of each selected data center:

[0017]

[0018] where, represents the gradient of the total loss L(φ) of N selected data centers, L(φ) represents the total loss of N selected data centers, Loss(θ i ) represents the 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 initial parameters.

[0019] Based on a further improvement of the above method, the multi-modal endometrial cancer samples include paired non-contrast CT images and contrast-enhanced CT images of endometrial cancer; some multi-modal endometrial cancer samples also include mask images corresponding to the non-contrast CT images and contrast-enhanced CT images;

[0020] The multi-channel 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;

[0021] The first segmentation module is used to perform target segmentation on the non-contrast CT image to obtain a first segmentation image;

[0022] The second segmentation module is used to perform target segmentation on the contrast-enhanced CT image to obtain a first segmentation image.

[0023] Based on further improvements to the above method, both the first segmentation module and the second segmentation module include a preprocessing module, an encoder, a bottleneck layer, a decoder, and a post-processing module;

[0024] The preprocessing module is used to preprocess the input image;

[0025] The encoder is used to gradually extract shallow features from the preprocessed image;

[0026] The bottleneck layer is used to extract deep features from the shallow features output by the encoder and output them to the decoder;

[0027] The decoder is used to decode according to the deep features and the shallow features of the corresponding layer of the encoder;

[0028] The post-processing module is used to generate a target area segmentation image according to the decoded features.

[0029] Based on further improvements to the above method, the following formula is used to calculate the loss:

[0030] Loss = L 1 + L 2 + L 3

[0031] Where, L 1 represents the multi-modal image segmentation loss, L 2 represents the multi-modal cross loss, L 3 represents the contrast loss.

[0032] Based on further improvements to the above method, the following formula is used to calculate the multi-modal cross loss:

[0033]

[0034] Where, represents the number of samples in the query set of the i-th data center that do not have 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 without a 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 without a corresponding mask image, and m represents the number of elements in the first segmentation image.

[0035] Based on further improvements to the above method, the following method is used to calculate the contrast loss:

[0036] For each sample in the query 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 query set that have no corresponding mask image as negative sample pairs; obtain the positive and negative sample pairs corresponding to each sample in the query set that has no corresponding mask image.

[0037] Calculate the contrastive loss based on the positive and negative sample pairs corresponding to each sample in the query set that has no corresponding mask image.

[0038] Based on a further improvement of the above method, calculate the contrastive loss based on the positive and negative sample pairs corresponding to each sample in the query set that has no corresponding mask image in the following way:

[0039]

[0040] Among them, represents the number of samples in the query 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 that has no corresponding mask image, P k2 represents the second segmentation image of the k-th sample that has no corresponding mask image, P j2 represents the second segmentation image of the j-th sample that has no corresponding mask image, τ represents the adjustment parameter, and dice(·,·) represents the DICE loss.

[0041] Based on a further improvement of the above method, use the following formula to calculate the multi-modal image segmentation loss:

[0042]

[0043] Among them, represents the number of samples in the query set of the i-th data center that have corresponding mask images, P k1 represents the first segmentation image of the k-th sample in the query set that has a corresponding mask image, G 1 represents the mask image corresponding to the plain CT image of the k-th sample in the query set that has a corresponding mask image, P k2 represents the second segmentation image of the k-th sample in the query set that has a corresponding mask image, G 2 represents the mask image corresponding to the enhanced CT image of the k-th sample in the query set that has a corresponding mask image, α 1 and α 2 represent the weight parameters.

[0044] Compared with the prior art, the present invention constructs multiple data centers, pre-trains the multi-channel segmentation model based on the multiple data centers, obtains the pre-trained endometrial cancer target segmentation model, and fine-tunes the pre-trained segmentation model on the target data set, so that the multiple data centers jointly train the segmentation model instead of mixing the data together for training, thereby not requiring the sample distribution of each data center to be consistent, and obtains the pre-trained endometrial cancer target segmentation model, and then fine-tunes it on the target data set, so as to quickly train and obtain the target segmentation model suitable for the target data set, and the performance of the model will not be affected by the small amount of data in the target data set, thereby improving the segmentation performance of the model on the target data set. For the center where the target data set is located, the multi-modal endometrial cancer image to be segmented is segmented based on the trained endometrial cancer target segmentation model, and the target segmentation result of the multi-modal endometrial cancer image to be segmented can be quickly and accurately obtained, thereby improving the efficiency of target segmentation.

[0045] 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

[0046] 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;

[0047] Figure 1 The flowchart of the method for delineating the target area of ​​endometrial cancer in a small sample according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0048] 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.

[0049] A specific embodiment of the present invention discloses a method for delineating a target region of endometrial cancer in a small sample, such as Figure 1 As shown, the following steps are included:

[0050] S1. Build multiple data centers, each of which includes a multi-channel segmentation model and a multi-modal endometrial cancer sample set;

[0051] S2. Pre-training the multi-channel segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer target area segmentation model;

[0052] S3. Fine-tune the pre-trained endometrial cancer target area segmentation model on the target data set to obtain a trained endometrial cancer target area segmentation model;

[0053] S4. Based on the trained endometrial cancer target area segmentation model, segment the multi-modal endometrial cancer image to be segmented to obtain the target area delineation result of the multi-modal endometrial cancer image to be segmented.

[0054] During implementation, each medical center collects multi-modal image data of endometrial cancer patients after surgery in the center to construct a multi-modal endometrial cancer sample set. At the same time, each center is equipped with a multi-channel segmentation model with the same structure and initial parameters, so as to construct a data center, that is, each data center has a local multi-channel segmentation model and a local sample set.

[0055] During implementation, the target data set is the data set of a new data center. The target data set is a small-sample data set. Simply training a model based on a small-sample data set is prone to overfitting, and the performance of the trained model is poor.

[0056] Compared with the prior art, the small-sample endometrial cancer target area delineation method provided in this embodiment constructs multiple data centers, pre-trains a multi-channel segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer target area segmentation model, and fine-tunes the pre-trained segmentation model on the target data set. Thus, multiple data centers jointly train the segmentation model instead of mixing the data for training, so it does not require the sample distributions of each data center to be consistent. By obtaining the pre-trained endometrial cancer target area segmentation model and then fine-tuning it on the target data set, a target area segmentation model suitable for the target data set can be quickly trained, and the performance of the model on the target data set will not be affected by the small amount of data in the target data set. Therefore, the segmentation performance of the model on the target data set is improved, and the accuracy of image semantic segmentation is improved. For the center where the target data set is located, based on the trained endometrial cancer target area segmentation model, segment the multi-modal endometrial cancer image to be segmented, and the target area segmentation result of the multi-modal endometrial cancer image to be segmented can be quickly and accurately obtained, improving the efficiency of target area segmentation.

[0057] Specifically, the 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 are divided into the support set, and a part of the samples are divided into the query set.

[0058] Specifically, pre-training the multi-channel segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer target area segmentation model includes:

[0059] S21. Randomly select N data centers, and record the parameters of the current multi-channel segmentation model as the initial parameters;

[0060] S22. For each selected data center, train the local multi-channel segmentation model based on the local support set and update the parameters of the local multi-channel segmentation model.

[0061] S23. For each selected data center, calculate the loss on the local query set based on the updated local multi-channel segmentation model.

[0062] S24. Update the initial parameters based on the losses of each selected data center; update the parameters of the multi-channel segmentation model of each data center to the initial parameters.

[0063] S25. Determine whether the stop condition is reached. If so, stop the training to obtain the pre-trained endometrial cancer target segmentation model; otherwise, return to step S21.

[0064] During implementation, first randomly select N data centers for model training in the current round. The parameters of the current multi-channel segmentation models of each data center are the same, denoted as the initial parameter ф.

[0065] For each selected data center, first train the local multi-channel segmentation model based on the support set in the local dataset and update the parameters of the local multi-channel segmentation model. During implementation, multi-step training can be performed, that is, the parameters of the local multi-channel segmentation model are updated multiple times, and the parameters of the last update are denoted as θ, where θ i represents the parameters of the local multi-channel segmentation model of the i-th data center.

[0066] After updating the parameters of the local multi-channel segmentation model, for each selected data center, calculate the loss on the local query set based on the updated local multi-channel segmentation model.

[0067] The multi-modal endometrial cancer samples include paired non-contrast CT images and contrast-enhanced CT images of endometrial cancer. Some multi-modal endometrial cancer samples also include mask images corresponding to the non-contrast CT images and contrast-enhanced CT images.

[0068] The multi-channel segmentation model can adopt a deep learning model.

[0069] The multi-channel 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.

[0070] The first segmentation module is used to perform target segmentation on the non-contrast CT image to obtain a first segmentation image.

[0071] The second segmentation module is used to perform target segmentation on the contrast-enhanced CT image to obtain a first segmentation image.

[0072] During implementation, due to the large workload and low efficiency of manual delineation, some samples in the sample set are allowed to have no annotation 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.

[0073] Specifically, both the first segmentation module and the second segmentation module include a preprocessing module, an encoder, a bottleneck layer, a decoder, and a postprocessing module;

[0074] The preprocessing module is used to preprocess the input image;

[0075] The encoder is used to gradually extract shallow features from the preprocessed image;

[0076] The bottleneck layer is used to extract deep features from the shallow features output by the encoder and output them to the decoder;

[0077] The decoder is used to decode according to the deep features and the shallow features of the corresponding layer of the encoder;

[0078] The postprocessing module is used to generate a target segmentation image according to the decoded features.

[0079] During implementation, first, the preprocessing module (Pre Block) is used to preprocess the feature extraction of the input image, and then the shallow feature extraction is gradually carried out through multiple downsampling layers and convolution modules of the encoder, and the size of the feature map is gradually reduced. The preprocessing module includes a 2D convolutional layer and a LeakyReLU activation layer. Each downsampling layer contains a 2D convolutional layer. Each convolutional module contains two instance normalization layers, two convolutional layers with a stride of 1, and two Leaky ReLU activation functions.

[0080] On the contrary, the decoder includes multiple upsampling layers and convolution modules. The structure of the convolution module in the decoder is the same as that in the encoder, so that the features can be gradually decoded and the size of the feature map can be gradually increased. Transposed convolution can be used as the upsampling layer. After passing through the upsampling layer and the convolution module, the postprocessing module (Post Block) performs target segmentation. The network structure of the postprocessing module is the same as that of the preprocessing module.

[0081] During implementation, the bottleneck layer uses a 1x1 convolutional kernel for dimensionality reduction and dimensionality increase, reduces the number of channels of the input features and then increases them back to the original number of channels, reduces the dimension of the features, extracts deep features and outputs them to the decoder.

[0082] During implementation, through skip connections, the shallow and deep features on the corresponding layers of the encoder and decoder are fused, enabling the decoder to obtain more high-resolution information during upsampling, thereby more perfectly restoring the detailed information in the original image and improving the segmentation accuracy.

[0083] During implementation, for the i-th data center, the samples on the local query set are input into the local multi-channel segmentation model (the model parameters are θ i ), and the loss is calculated based on the first segmentation image and the second segmentation image output by the model.

[0084] Specifically, for the i-th data center, the loss is calculated on the local query set based on the updated local multi-channel segmentation model, and the following formula is specifically used to calculate the loss:

[0085] Loss = L 1 + L 2 + L 3 (1)

[0086] 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.

[0087] Specifically, for the labeled samples, that is, the samples with corresponding mask images, the multi-modal 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.

[0088] Specifically, the following formula is used to calculate the multi-modal image segmentation loss:

[0089]

[0090] where represents the number of samples with corresponding mask images in the query set of the i-th data center, P k1 represents the first segmentation image of the k-th sample with a corresponding mask image in the query set, G 1 represents the mask image corresponding to the plain CT image of the k-th sample with a corresponding mask image in the query set, P k2 represents the second segmentation image of the k-th sample with a corresponding mask image in the query set, G 1 represents the mask image corresponding to the enhanced CT image of the k-th sample with a corresponding mask image in the query set, α 1 and α 2 represent weight parameters.

[0091] During implementation, α 1 and α 2For separately controlling false negatives and false positives 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 The intersection

[0092] For samples without labels, that is, samples without corresponding mask images, the loss is calculated through multi-modal mutual learning and contrast learning to obtain the multi-modal cross-loss and contrast loss.

[0093] Specifically, the following formula is used to calculate the multi-modal cross-loss:

[0094]

[0095] where represents the number of samples in the query set of the i-th data center without corresponding mask images, and P k1 (s, j) represents the element value at the position (s, j) in the first segmentation image of the k-th sample without a corresponding mask image, and P k2 (s, j) represents the element value at the position (s, j) in the second segmentation image of the k-th sample without a corresponding mask image, and m represents the number of elements in the first segmentation image.

[0096] 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.

[0097] Specifically, the following method is used to calculate the contrast loss:

[0098] For each sample in the query set without 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 without corresponding mask images in the query set are used as negative sample pairs; positive sample pairs and negative sample pairs corresponding to each sample in the query set without a corresponding mask image are obtained;

[0099] 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.

[0100] For example, the number of samples in the query set of the i-th data center without corresponding mask images is ones.

[0101] For the k-th sample without a corresponding mask image, the first segmentation image P k1 and the second segmentation image P k2 predicted by the model are used as the positive sample pair (P k1 , Pk2 )。Then, the first segmented image P k1 and the second segmented images of other samples in the query set form the negative sample pair corresponding to this sample, (P k1 , P j2 ), j ≠ k.

[0102] Then, based on the positive and negative sample 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:

[0103]

[0104] Among them, represents the number of samples in the query set of the i-th data center that do not have a corresponding mask image, P k1 represents the first segmented image of the k-th sample that does not have a corresponding mask image, P k2 represents the second segmented image of the k-th sample that does not have a corresponding mask image, P j2 represents the second segmented image of the j-th sample that does not have a corresponding mask image, τ represents the adjustment parameter, and dice(·, ·) represents the DICE loss.

[0105] Through the contrast 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.

[0106] After calculating the loss in N data, the initial parameters are updated based on the losses of N data centers.

[0107] During implementation, since the sample sizes and data qualities of different data centers are different, the weights of the losses of N data centers for updating the initial parameters are different.

[0108] Specifically, the weights of each selected data center are obtained in the following way:

[0109] S241. Pre-update the parameters of the multi-channel segmentation models of N selected data centers based on the losses of N selected data centers;

[0110] S242. For each selected data center, calculate the marginal loss on the local query set based on the pre-updated local multi-channel segmentation model;

[0111] S243. Calculate the weights of each selected data center based on the marginal losses of each selected data center.

[0112] During implementation, pre-updating the parameters of the multi-channel segmentation models of N selected data centers based on the losses of N selected data centers includes:

[0113] The first pre-update parameter is calculated based on the losses of N selected data centers;

[0114] For the i-th selected data center, the second pre-update parameter of the i-th selected data center is obtained based on the losses of the other N - 1 selected data centers;

[0115] For the i-th selected data center, the local multi-channel segmentation model is pre-updated 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.

[0116] Specifically, the first pre-update parameter is calculated using the following formula based on the losses of N selected data centers:

[0117]

[0118] where, represents the gradient of the total loss L(φ), L(φ) represents the total loss of 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 parameter, and μ′ 1 represents the first pre-update parameter.

[0119] Specifically, the second pre-update parameter of the i-th selected data center obtained using the following formula based on the losses of the other N - 1 selected data centers:

[0120]

[0121] where, represents the gradient of the total loss L i2 (φ), L i2 (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.

[0122] During implementation, the parameters of the local multi-channel segmentation model of the i-th selected data center are updated to the first pre-update parameter to obtain the first pre-update model (parameters are φ′ 1 ), and the parameters of the local multi-channel segmentation model of the i-th selected data center are updated to the second pre-update parameter to obtain the second pre-update model (parameters are φ′ i2 ).

[0123] Specifically, for the i-th selected data center, calculating the marginal loss based on the pre-updated local multi-channel segmentation model on the local query set includes:

[0124] Calculating the loss based on the first pre-updated model on the local query set to obtain the first loss;

[0125] Calculating the loss based on the second pre-updated model on the local query set to obtain the second loss;

[0126] The difference between the second loss and the first loss gives the marginal loss.

[0127] During implementation, both the first loss and the second loss can be calculated according to the formula in Equation (1).

[0128] Specifically, the weight of each selected data center is calculated using the following formula based on the marginal loss of each selected data center:

[0129]

[0130] where q i represents the data quality of the i-th selected data center, δ i represents the marginal loss of the i-th selected data center, and γ represents the adjustment parameter.

[0131] After obtaining the weights, the initial parameters are updated based on the losses of the N data centers using the following formula:

[0132]

[0133] where, represents the gradient of the total loss L(φ) of the N selected data centers, 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, ω i represents the weight of the i-th selected data center, β represents the learning rate, and φ represents the initial parameters.

[0134] It should be noted that loss(θ i ) represents the loss of the i-th selected data center, which is the loss calculated in step S23, that is, the loss calculated when the parameter is θ i .

[0135] By calculating the weights according to the data quality of different data centers and dynamically calculating the model parameters during each backpropagation, the weights of data centers with high quality are increased, thereby improving the model segmentation accuracy and the model performance.

[0136] After updating the initial parameters, update the parameters of the local multi-channel segmentation models of all data centers to the initial parameters, thereby completing one backpropagation of the multi-channel segmentation model.

[0137] After completing one backpropagation of the multi-channel segmentation model, if the current stopping condition is reached, end the training and obtain the pre-trained endometrial cancer target area segmentation model; if not, return to step S21 to continue training until the stopping condition of the model training is reached. In implementation, the stopping condition can be reaching a preset number of training times or the total loss reaching a preset accuracy.

[0138] For a new medical center, a small amount of multi-modal endometrial cancer samples can be collected to construct a target data set, and the pre-trained endometrial cancer target area segmentation model can be fine-tuned on the target data set to quickly obtain a trained endometrial cancer target area segmentation model. Input the multi-modal endometrial cancer image to be segmented at this center into the trained endometrial cancer target area segmentation model to quickly and accurately obtain the target area image.

[0139] Those skilled in the art can understand that all or part of the processes of implementing the above method 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 disk, an optical disc, a read-only memory or a random access memory, etc.

[0140] 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 by the protection scope of the present invention.

Claims

1. A method for delineating a target area of ​​endometrial cancer in a small sample, characterized in that: The following steps are involved: Build multiple data centers, each of which includes a multi-channel segmentation model and a multi-modal endometrial cancer sample set; Pre-training the multi-channel segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer target area segmentation model; Fine-tuning the pre-trained endometrial cancer target region segmentation model on the target data set to obtain a trained endometrial cancer target region segmentation model; The multimodal endometrial cancer image to be segmented is segmented based on the trained endometrial cancer target segmentation model to obtain the target area delineation result of the multimodal endometrial cancer image to be segmented.

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

3. The method for delineating the target area of ​​endometrial cancer in a small sample according to claim 2, characterized in that: The initial parameters are updated based on the loss of each selected data center using the following formula: in, represents the gradient of the total loss L(φ) of the N selected data centers, 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, ω i represents the weight of the i-th selected data center, β represents the learning rate, and φ represents the initial parameter.

4. The method for delineating the target area of ​​endometrial cancer in a small sample according to claim 2, characterized in that: 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 mask images corresponding to the plain scan CT images and enhanced CT images; The multi-channel 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; 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.

5. The method for delineating the target area of ​​endometrial cancer in a small sample according to claim 4, characterized in that: The first segmentation module and the second segmentation module both include a pre-processing module, an encoder, a bottleneck layer, a decoder and a post-processing module; The preprocessing module is used to preprocess the input image; The encoder is used to gradually extract shallow features from the preprocessed image; The bottleneck layer is used to extract deep features from shallow features output by the encoder and output them to the decoder; The decoder is used to decode according to the deep features and the shallow features of the corresponding layer of the encoder; The post-processing module is used to generate a target area segmentation image based on the decoded features.

6. The method for delineating the target area of ​​endometrial cancer in small sample according to claim 4, characterized in that: The 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.

7. The method for delineating the target area of ​​endometrial cancer in a small sample according to claim 6, characterized in that: The multimodal cross loss is calculated using the following formula: in, represents the number of samples that do not have corresponding mask images in the query 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.

8. The method for delineating the target area of ​​endometrial cancer in small sample according to claim 6, 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.

9. The method for delineating the target area of ​​endometrial cancer in a small sample according to claim 8, 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, represents the number of samples that do not have corresponding mask images in the query 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.

10. The method for delineating the target area of ​​endometrial cancer in small samples according to claim 6, characterized in that: The multimodal image segmentation loss is calculated using the following formula: in, represents the number of samples with corresponding mask images in the query set of the i-th data center, P k1 represents the first segmented image of the k-th sample with a corresponding mask image in the query set, G1 represents the mask image corresponding to the plain scan CT image of the k-th sample with a corresponding mask image in the query set, P k2 represents the second segmented image of the k-th sample with a corresponding mask image in the query set, G2 represents the mask image corresponding to the enhanced CT image of the k-th sample with a corresponding mask image in the query set, and α1 and α2 represent weight parameters.

Images