A small sample endometrial cancer target area delineation method

Abstract

The present application relates to a kind of small sample endometrial carcinoma target delineation method, belong to target delineation technical field, solve the problem of low segmentation accuracy caused by small sample data training model in prior art.The method comprises the following steps: constructing multiple data centers, each data center includes multiple-channel segmentation model and multiple modal endometrial carcinoma sample set;The pre-training of the multiple-channel segmentation model is carried out based on the multiple data centers, and the pre-trained endometrial carcinoma target segmentation model is obtained;The pre-trained endometrial carcinoma target segmentation model is fine-tuned on target data set, and the trained endometrial carcinoma target segmentation model is obtained;Based on the trained endometrial carcinoma target segmentation model, the multiple modal endometrial carcinoma image to be segmented is segmented, and the target delineation result of the multiple modal endometrial carcinoma image to be segmented is obtained.The high-efficiency and accurate endometrial carcinoma target delineation is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target delineation, and particularly relates to a small sample endometrial carcinoma target delineation method. BACKGROUND

[0002] Endometrial carcinoma is one of the most common malignant tumors of the female reproductive system, and its incidence is increasing globally. In China, with the aging of the population and changes in lifestyle, the number of new cases of endometrial carcinoma is also increasing. This makes accurate diagnosis and effective treatment of endometrial carcinoma an important medical need. For endometrial carcinoma patients, postoperative target segmentation is crucial in subsequent treatment planning (such as radiotherapy planning). Accurate target segmentation can ensure that treatment methods such as radiotherapy precisely act on the tumor area, reduce damage to the surrounding normal tissue, improve the quality of life of patients and improve prognosis.

[0003] Training a model with data collected by a certain medical center alone is prone to overfitting due to small sample size, and the model cannot accurately segment the target area. SUMMARY

[0004] In view of the above analysis, the embodiments of the present application aim to provide a small sample endometrial carcinoma target delineation method to solve the problem of low segmentation accuracy caused by training a model with small sample data.

[0005] In one aspect, the embodiments of the present application provide a small sample endometrial carcinoma target delineation method, comprising the following steps:

[0006] Constructing a plurality of data centers, each data center comprising a multi-channel segmentation model and a multi-modal endometrial carcinoma sample set;

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

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

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

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

[0011] S21, randomly select N data centers, and denote the parameters of the current multi-channel segmentation model as 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 loss 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 a stop condition is reached, if yes, stop training to obtain a pre-trained endometrial cancer target region segmentation model, otherwise, return to step S21.

[0016] Based on the further improvement of the above method, the initial parameters are updated based on the loss of each selected data center using the following formula:

[0017]

[0018] wherein, 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.

[0019] Based on the further improvement of the above method, the multi-modal endometrial cancer sample includes a pair of endometrial cancer plain CT image and enhanced CT image; part of the multi-modal endometrial cancer sample further includes a mask image corresponding to the plain CT image and the enhanced CT image;

[0020] The multi-channel segmentation model comprises a first segmentation module and a second segmentation module; the first segmentation module and the second segmentation module are identical in structure;

[0021] The first segmentation module is used for target region segmentation of the plain CT image to obtain a first segmentation image;

[0022] The second segmentation module is used for target region segmentation of the enhanced CT image to obtain a first segmentation image.

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

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

[0025] The encoder is configured to perform step-by-step shallow feature extraction on the preprocessed image;

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

[0027] The decoder is configured to decode the deep features and the shallow features of the corresponding layers of the encoder;

[0028] The post-processing module is configured to generate a target region segmentation image based on the decoded features.

[0029] Based on the further improvement of the above method, the loss is calculated using the following formula:

[0030] Loss = L1 + L2 + L3

[0031] Wherein, L1 represents the multi-modal image segmentation loss, L2 represents the multi-modal cross loss, and L3 represents the contrast loss.

[0032] Based on the further improvement of the above method, the multi-modal cross loss is calculated using the following formula:

[0033]

[0034] Wherein, P(k) represents the number of samples without corresponding mask images in the query set of the i-th data center, P k1 (s,j) represents the element value of 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 of 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 of the first segmentation image.

[0035] Based on the further improvement of the above method, the contrast loss is calculated in the following manner:

[0036] For each sample without a corresponding mask image in the query set, the first segmentation image and the second segmentation image of the sample are used as positive sample pairs; the first segmentation image of the sample and the second segmentation image of other samples without a corresponding mask image in the query set are used as negative sample pairs; the positive sample pairs and negative sample pairs corresponding to each sample without a corresponding mask image in the query set are obtained;

[0037] The contrast loss is calculated based on the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set.

[0038] Based on the further improvement of the above method, the contrast loss is calculated based on the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set in the following way:

[0039]

[0040] wherein, represents the number of samples without a corresponding mask image in the query set of the i-th data center, 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 an adjustment parameter, and dice(·,·) represents a DICE loss.

[0041] Based on the further improvement of the above method, the multi-modal image segmentation loss is calculated using the following formula:

[0042]

[0043] wherein, represents the number of samples with a corresponding mask image 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, G1 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, 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.

[0044] Compared with the prior art, the uterus endometrial carcinoma target region segmentation method of the present application comprises the following steps: S1, constructing a plurality of data centers, each data center comprising a multi-channel segmentation model and a multi-modal uterus endometrial carcinoma sample set; S2, pre-training the multi-channel segmentation model based on the plurality of data centers to obtain a pre-trained uterus endometrial carcinoma target region segmentation model; and S3, fine-tuning the pre-trained segmentation model on a target data set to obtain a target region segmentation model suitable for the target data set.

[0045] The above technical solutions can be combined with each other in the present application to achieve more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification or be understood by implementing the present application. The purposes and other advantages of the present application can be achieved and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0046] The accompanying drawings are included to provide a further understanding of the present application, and are incorporated herein and constitute a part of the specification. The drawings illustrate embodiments of the present application and, together with the specification, serve to explain the principles of the present application.

[0047] Figure 1 A flowchart of the uterus endometrial carcinoma target region segmentation method of the present application embodiment. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present application will be specifically described below in conjunction with the drawings, wherein the drawings constitute a part of the present application and are used to explain the principles of the present application together with the embodiments of the present application, and are not used to limit the scope of the present application.

[0049] One specific embodiment of the present application discloses a small-sample uterus endometrial carcinoma target region segmentation method, as shown in the figure, comprising the following steps: Figure 1

[0050] S1, constructing a plurality of data centers, each data center comprising a multi-channel segmentation model and a multi-modal uterus endometrial carcinoma sample set;

[0051] S2, pre-training the multi-channel segmentation model based on the plurality of data centers to obtain a pre-trained uterus endometrial carcinoma target region segmentation model;

[0052] ​S3, fine-tuning the pre-trained endometrial cancer target segmentation model on the target data set to obtain a trained endometrial cancer target segmentation model;

[0053] S4, segmenting the multi-modal endometrial cancer image to be segmented based on the trained endometrial cancer target segmentation model to obtain a target region delineation result of the multi-modal endometrial cancer image to be segmented.

[0054] In implementation, each medical center collects postoperative multi-modal image data of endometrial cancer patients in the center to construct a multi-modal endometrial cancer sample set, and each center is equipped with a multi-channel segmentation model with the same structure and initial parameters, thereby constructing a data center, i.e., each data center has a local multi-channel segmentation model and a local sample set.

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

[0056] Compared with the prior art, the small sample endometrial cancer target delineation method provided in the embodiment trains a multi-channel segmentation model based on multiple data centers, pre-trains the multi-channel segmentation model based on the multiple data centers to obtain a pre-trained endometrial cancer target segmentation model, fine-tunes the pre-trained segmentation model on a target data set, thereby training the segmentation model by the multiple data centers instead of mixing the data together for training, so that the sample distribution of each data center does not need to be consistent, the pre-trained endometrial cancer target segmentation model is obtained, and the target data set is fine-tuned, thereby quickly training a target data set applicable target segmentation model, which does not affect the performance of the model due to the small amount of data in the target data set, thereby improving the segmentation performance of the model on the target data set and improving the accuracy of image semantic segmentation. For the center where the target data set is located, the trained endometrial cancer target segmentation model is used to segment the multi-modal endometrial cancer image to be segmented, so that the target segmentation result of the multi-modal endometrial cancer image to be segmented can be quickly and accurately obtained, and the efficiency of target segmentation is improved.

[0057] Specifically, the sample set of each data center includes a support set and a query set, i.e., a part of the samples in the sample set are divided into a support set, and a part of the samples are divided into a query set.

[0058] Specifically, the pre-training of the multi-channel segmentation model based on the multiple data centers includes:

[0059] S21, randomly selecting N data centers, and recording the parameters of the current multi-channel segmentation model as initial parameters;

[0060] S22, for each selected data center, training the local multi-channel segmentation model based on the local support set, updating the parameters of the local multi-channel segmentation model;

[0061] S23, for each selected data center, calculating the loss on the local query set based on the updated local multi-channel segmentation model;

[0062] 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;

[0063] S25, determining whether a stop condition is reached, if yes, stopping training to obtain the pre-trained endometrial cancer target region segmentation model, otherwise, returning to step S21.

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

[0065] For each selected data center, first, the local multi-channel segmentation model is trained based on the support set in the local data set, and the parameters of the local multi-channel segmentation model are updated. In implementation, multi-step training can be performed, that is, the parameters of the local multi-channel segmentation model are updated multiple times, and the last updated parameters are marked as θ, 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, each selected data center calculates the loss on the local query set based on the updated local multi-channel segmentation model.

[0067] The multi-modal endometrial cancer sample includes a pair of endometrial cancer plain CT image and enhanced CT image, and part of the multi-modal endometrial cancer sample further includes a mask image corresponding to the plain CT image and the enhanced CT image.

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

[0069] The multi-channel segmentation model comprises a first segmentation module and a second segmentation module; the first segmentation module and the second segmentation module are of the same structure;

[0070] The first segmentation module is used for target region segmentation on the plain CT image to obtain a first segmentation image;

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

[0072] In 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, no corresponding mask image. Therefore, the multi-modal endometrial carcinoma sample set includes pairs of endometrial carcinoma plain CT images and enhanced CT images, and part of the pairs of endometrial carcinoma plain CT images and enhanced CT images correspond to the mask images.

[0073] Specifically, the first segmentation module and the second segmentation module each include a preprocessing module, an encoder, a bottleneck layer, a decoder, and a post-processing module.

[0074] The preprocessing module is configured to pre-process the input image.

[0075] The encoder is configured to perform step-by-step shallow feature extraction on the pre-processed image.

[0076] The bottleneck layer is configured to extract deep features from the shallow features output by the encoder and output to the decoder.

[0077] The decoder is configured to decode the deep features and the shallow features of the corresponding layers of the encoder.

[0078] The post-processing module is configured to generate a target region segmentation image according to the decoded features.

[0079] In implementation, first, a preprocessing module (Pre Block) is used to pre-process the input image to extract features, and then a plurality of down-sampling layers and convolution modules of the encoder are used to perform step-by-step shallow feature extraction and gradually reduce the size of the feature map. The preprocessing module includes a 2D convolution layer and a LeakyReLU activation layer. Each down-sampling layer includes a 2D convolution layer. Each convolution module includes two instance normalization layers, two convolution layers with a stride of 1, and two Leaky ReLU activation functions.

[0080] On the contrary, the decoder includes a plurality of up-sampling layers and convolution modules, the convolution module structure of the decoder is consistent with that of the encoder, so that the features can be decoded step by step and the size of the feature map can be gradually increased. Deconvolution can be used as the up-sampling layer. After the up-sampling layer and the convolution module, the post-processing module (Post Block) performs target region segmentation. The network structure of the post-processing module is the same as that of the preprocessing module.

[0081] In implementation, the bottleneck layer uses a 1x1 convolution kernel to reduce and increase dimensions, reduces the number of channels of the input features, increases the number of channels to the original number, reduces the dimension of the features, extracts deep features, and outputs to the decoder.

[0082] In implementation, the corresponding encoder and decoder are fused with the shallow features and deep features on the corresponding layers of the decoder through the skip connection, so that the decoder can obtain more high-resolution information when upsampling, and then more perfectly restore the detailed information in the original image, and improve the segmentation accuracy.

[0083] In 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 according to 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 loss is calculated using the following formula:

[0085] Loss=L1+L2+L3 (1)

[0086] Wherein, L1 represents the multi-modal image segmentation loss, L2 represents the multi-modal cross loss, and L3 represents the contrast loss.

[0087] Specifically, for the labeled sample, that is, the sample with the corresponding mask image, 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 multi-modal image segmentation loss is calculated using the following formula:

[0089]

[0090] Wherein, P k1 represents the number of samples with corresponding mask images in the query set of the i th data center, P k2 represents the first segmentation image of the k th sample with corresponding mask images in the query set, G1 represents the mask image corresponding to the plain CT image of the k th sample with corresponding mask images in the query set, P k1 represents the second segmentation image of the k th sample with corresponding mask images in the query set, G1 represents the mask image corresponding to the enhanced CT image of the k th sample with corresponding mask images in the query set, and α1 and α2 represent weight parameters.

[0091] In implementation, α1 and α2 are used to control false negatives and false positives respectively, and adjusting α1 and α2 can control the trade-off between false negatives and false positives. |P k1 ∩G1| represents the intersection of P k1 and G1

[0092] For samples without labels, i.e., without corresponding mask images, the loss is calculated by multi-modal mutual learning and contrastive learning to obtain a multi-modal cross-loss and a contrastive loss.

[0093] Specifically, the multi-modal cross-loss is calculated using the following formula:

[0094]

[0095] wherein, represents the number of samples without corresponding mask images in the query set of the i-th data center, P k1 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 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.

[0096] Through the multi-modal cross-loss, the first segmentation module and the second segmentation module are trained to segment images as similar as possible.

[0097] Specifically, the contrastive loss is calculated in the following manner:

[0098] For each sample without a corresponding mask image in the query set, the first segmentation image and the second segmentation image of the sample are taken as a positive sample pair; the first segmentation image of the sample and the second segmentation image of other samples without a corresponding mask image in the query set are taken as a negative sample pair; the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set are obtained.

[0099] The contrastive loss is calculated based on the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set.

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

[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 taken as the positive sample pair (P k1 , P k2 ) corresponding to the sample. Then, the first segmentation image P k1 and the second segmentation image of other samples in the query set constitute the negative sample pair (P k1 , P j2 ) corresponding to the sample, j≠k.

[0102] Then, the contrastive loss is calculated based on the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set in each data center in the following manner:

[0103]

[0104] wherein, represents the number of samples without a corresponding mask image in the query set of the i-th data center, 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 an adjustment parameter, and dice(·,·) represents a DICE loss.

[0105] The contrastive loss makes the distance between positive samples closer and the distance between negative samples farther, thereby deepening the mutual learning between different modalities and improving the segmentation performance of the model.

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

[0107] In implementation, because the sample quantity and data quality are different in different data centers, the loss in the N data centers has different weights in updating the initial parameters.

[0108] Specifically, the weight of each selected data center is obtained in the following manner:

[0109] S241, pre-update the parameters of the multi-channel segmentation model of the N selected data centers based on the loss of the 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 weight of each selected data center based on the marginal loss of each selected data center.

[0112] In implementation, the pre-update of the parameters of the multi-channel segmentation model of the N selected data centers based on the loss of the N selected data centers includes:

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

[0114] for the i-th selected data center, a second pre-update parameter of the i-th selected data center is obtained based on the loss 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 based on the losses of the N selected data centers using the following formula:

[0117]

[0118] in, Let L(φ) represent the gradient of the total loss, where L(φ) represents the total loss across the N selected data centers, and Loss(θ) represent the gradient of the total loss. i ) represents the loss of the i-th selected data center, β represents the learning rate, φ represents the initial parameters, and μ′1 represents the first pre-update parameters.

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

[0120]

[0121] in, L represents the total loss. i2 The gradient of (φ), L i2 (φ) represents the total loss of the N-1 selected data centers, Loss(θ) j Let φ' represent the loss for the i-th selected data center, β represent the learning rate, φ represent the initial parameters, and φ' represent the initial parameters. i2 This represents the second pre-update parameter for the i-th selected data center.

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

[0123] Specifically, for the i-th selected data center, the marginal loss is calculated on the local query set based on the pre-updated local multi-channel partitioning model, including:

[0124] The first loss is obtained by calculating the loss on the local query set based on the first pre-updated model;

[0125] The second loss is obtained by calculating the loss on the local query set based on the second pre-update model;

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

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

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

[0129]

[0130] wherein 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 an adjustment parameter.

[0131] After obtaining the weight, the initial parameter is updated based on the loss of the N data centers using the following formula:

[0132]

[0133] wherein 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 a learning rate, and φ represents the initial parameter.

[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, i.e., the loss calculated when the parameter θ i .

[0135] By calculating the weight according to the data quality of different data centers and dynamically calculating the model parameters each time of back propagation, the weight of the data center with high quality is improved, and thus the model segmentation accuracy and the model performance are improved.

[0136] After updating the initial parameter, the parameters of the local multi-channel segmentation model of all data centers are updated to the initial parameter, and thus one back propagation of the multi-channel segmentation model is completed.

[0137] After completing one back propagation of the multi-channel segmentation model, if the current stop condition is reached, the training is ended, and a pre-trained endometrial cancer target region segmentation model is obtained. If not, the step S21 is returned to continue the training until the stop condition of the model training is reached. In implementation, the stop condition can be that the preset training times are reached, or the total loss reaches the 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 segmentation model is fine-tuned on the target data set, so that a trained endometrial cancer target segmentation model can be quickly obtained. The multi-modal endometrial cancer image to be segmented in this center is input into the trained endometrial cancer target segmentation model, and the target region image can be quickly and accurately obtained.

[0139] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing relevant hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory, a random access memory, etc.

[0140] The above description is only a preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A small sample endometrial cancer target delineation method, characterized in that, The method comprises the following steps: constructing a plurality of data centers, each data center comprising a multi-channel segmentation model and a multi-modal endometrial cancer sample set; the structure and initial parameters of the multi-channel segmentation model of each data center are the same; 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; fine-tuning the pre-trained endometrial cancer target region segmentation model on a target data set to obtain a trained endometrial cancer target region segmentation model; segmenting a multi-modal endometrial cancer image to be segmented based on the trained endometrial cancer target region segmentation model to obtain a target region delineation result of the multi-modal endometrial cancer image to be segmented; the sample set comprises a support set and a query set; the pre-training of the multi-channel segmentation model based on the plurality of data centers to obtain a pre-trained endometrial cancer target region segmentation model comprises: S21, randomly selecting N data centers, and denoting the parameters of the current multi-channel segmentation model as initial parameters; S22, for each selected data center, training the local multi-channel segmentation model based on the local support set, and updating the parameters of the local multi-channel segmentation model; S23, for each selected data center, calculating 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 a stop condition is reached, if yes, stopping training to obtain a pre-trained endometrial cancer target region segmentation model, otherwise, returning to step S21.

2. The small sample endometrial cancer target volume delineation method of claim 1, wherein, The initial parameters are updated based on the loss of each selected data center using the following formula: ； ； wherein, represents the total loss of N selected data centers the gradient of, represents the total loss of N selected data centers, represents the loss of the i-th selected data center, represents the weight of the i-th selected data center, represents the learning rate, represents the initial parameters.

3. The small sample endometrial cancer target volume delineation method of claim 1, wherein, The multi-modal endometrial cancer sample comprises a pair of endometrial cancer plain CT image and enhanced CT image; part of the multi-modal endometrial cancer sample further comprises a mask image corresponding to the plain CT image and the enhanced CT image; The multi-channel 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 for target region segmentation of the plain CT image to obtain a first segmentation image; The second segmentation module is used for target region segmentation of the enhanced CT image to obtain a first segmentation image.

4. The small sample endometrial cancer target volume delineation method of claim 3, wherein, Both the first segmentation module and the second segmentation module comprise a preprocessing module, an encoder, a bottleneck layer, a decoder and a post-processing module; The preprocessing module is used for preprocessing the input image; The encoder is used for step-by-step shallow feature extraction of the preprocessed image; The bottleneck layer is used for extracting deep features from the shallow features output by the encoder and outputting to the decoder; The decoder is used for decoding according to the deep features and the shallow features of the corresponding layers of the encoder; The post-processing module is used for generating a target region segmentation image according to the decoding features.

5. The small sample endometrial cancer target volume delineation method of claim 3, wherein, The loss is calculated using the following formula: ； wherein, denotes a multi-modal image segmentation loss, denotes a multi-modal cross loss, denotes a contrastive loss.

6. The small sample endometrial cancer target volume delineation method of claim 5, wherein, The multi-modal cross-loss is calculated using the following formula: ； wherein, represents the number of samples in the query set of the i-th data center that do not have a corresponding mask image, represents the element value of the position in the first segmentation image of the k-th sample that does not have a corresponding mask image, represents the element value of the position in the second segmentation image of the k-th sample that does not have a corresponding mask image, m represents the number of elements of the first segmentation image.​ 7. The small sample endometrial cancer target volume delineation method of claim 5, wherein, The contrast loss is calculated in the following manner: For each sample in the query set without a corresponding mask image, taking the first segmentation image and the second segmentation image of the sample as a positive sample pair; taking the first segmentation image of the sample and the second segmentation image of other samples in the query set without a corresponding mask image as a negative sample pair; Obtaining the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set; Based on the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set, a contrast loss is calculated.

8. The small sample endometrial cancer target volume delineation method according to claim 7, characterized in that, Based on the positive sample pair and the negative sample pair corresponding to each sample without a corresponding mask image in the query set, the contrast loss is calculated in the following manner: ； wherein, represents the number of samples in the i-th data center that have no corresponding mask image in the query set, represents the first segmentation image of the k-th sample that has no corresponding mask image, represents the second segmentation image of the k-th sample that has no corresponding mask image, represents the second segmentation image of the j-th sample that has no corresponding mask image, τ represents an adjustment parameter, and dice(·,·) represents a DICE loss.

9. The small sample endometrial cancer target volume delineation method of claim 5, wherein, The multi-modal image segmentation loss is calculated using the following formula: ； wherein, represents the number of samples in the query set of the i-th data center that have a corresponding mask image, represents the first segmentation image of the k-th sample in the query set that has a corresponding mask image, represents the mask image corresponding to the planar CT image of the k-th sample in the query set that has a corresponding mask image, represents the second segmentation image of the k-th sample in the query set that has a corresponding mask image, 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, and represents a weight parameter.

Images