Multi-organ segmentation system based on partial annotated dataset co-learning

Abstract

The application discloses a kind of multi-organ segmentation systems based on partial annotation dataset collaborative learning, comprising: histogram matching module: for realizing the image dataset from different centers has same gray scale distribution;Single organ segmentation model collaborative learning module: for improving the segmentation performance of each single organ segmentation model, and realizing that different organs labeled in different datasets can be aware of each other, improve the quality of generated pseudo label;Multi-organ segmentation model collaborative learning module: for making full use of the label in different datasets, improve the segmentation performance of each multi-organ segmentation model;Multi-organ segmentation model distillation module: for learning from multiple multi-organ segmentation models to obtain a final multi-organ segmentation model.The application realizes a multi-organ segmentation model based on partially annotated medical image dataset, greatly reduces the dependence on fully annotated dataset, and relieves the annotation difficulty of doctors.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image segmentation, in particular to a multi-organ segmentation system based on partial annotation dataset collaborative learning. BACKGROUND

[0002] In recent years, the development of deep learning has greatly improved the accuracy of multi-organ segmentation compared to traditional methods. However, deep learning relies on a large amount of annotated data, and obtaining annotations for multiple organs simultaneously is very difficult and time-consuming in the multi-organ segmentation task, resulting in a relatively small number of available multi-organ segmentation datasets.

[0003] On the other hand, there are many publicly available single-organ datasets in reality, such as kidney datasets, liver datasets, and pancreas datasets, which are more numerous than multi-organ datasets. Therefore, using single-organ datasets from different centers to train a multi-organ model has gradually become an important research direction.

[0004] To make full use of these partially annotated datasets, one intuitive idea is to train a single-organ segmentation model for each dataset, and then sequentially obtain the output of each organ on the multi-organ dataset. However, this approach requires training multiple models for multiple organs, which is time-consuming and memory-intensive, and the prior positional relationship between different organs is not fully utilized. Another approach is to use trained single-organ models to generate pseudo-labels for other datasets, resulting in a fully annotated dataset containing pseudo-labels, and then training a multi-organ segmentation model using these datasets. However, due to the domain differences between partially annotated datasets and multi-organ datasets, the quality of the generated pseudo-labels is poor, and the performance of the multi-organ model trained based on the pseudo-labels is very low.

[0005] Based on the above two ideas, there have been many improved solutions in recent years, mainly divided into two categories. One is the model-based method, which mainly embeds conditional encoding into the network to obtain a unified model for sequentially generating the annotation of each organ, or is based on continuous learning and federated learning. The other is the pseudo-label-based method, which generates pseudo-labels by training single-organ models and then corrects them, and then trains a multi-organ segmentation model based on the pseudo-label-containing dataset. Their performance has been greatly improved compared to before, but the model-based method can only sequentially obtain the annotation of each organ and cannot segment multiple organs at once. The pseudo-label-based method does not solve the domain differences between different datasets, and the organ prior information provided by the labels in different datasets is not fully utilized in the training process.

[0006] Based on the above problems, the present application proposes a novel collaborative learning framework for learning a multi-organ segmentation model from partially annotated datasets. SUMMARY

[0007] The purpose of the present application is to provide a multi-organ segmentation system based on partial annotation dataset collaborative learning to solve the problems existing in the prior art.

[0008] To achieve the above purpose, the present application provides a multi-organ segmentation system based on partial annotation dataset collaborative learning, comprising:

[0009] a histogram matching module, a single-organ segmentation model collaborative learning module, a multi-organ segmentation model collaborative learning module, and a multi-organ segmentation model distillation module.

[0010] The histogram matching module is used to process images in a plurality of partial annotation datasets to obtain a plurality of matching datasets.

[0011] The single-organ segmentation model collaborative learning module is used to obtain a plurality of full-label datasets according to a single-organ segmentation model and the plurality of matching datasets.

[0012] The multi-organ segmentation model collaborative learning module is used to train a plurality of multi-organ segmentation models according to the plurality of full-label datasets.

[0013] The multi-organ segmentation model distillation module is used to learn a final multi-organ segmentation model from the plurality of multi-organ segmentation models.

[0014] Optionally, the histogram matching module comprises:

[0015] a gray scale distribution calculation unit, which is configured to normalize all images in a source domain and a target domain to make the gray scale values of the images be integer values within a preset interval, and then calculate the average gray scale distribution of the images in the target domain to obtain the average gray scale distribution of the images in the multi-organ dataset; the source domain includes a plurality of partial annotation datasets;

[0016] a balancing unit, which is configured to define a corresponding normalized histogram based on the number of rows and columns of the images in the source domain and the target domain and the number of pixels of different intensity values, and perform histogram equalization on the normalized histogram;

[0017] a matching unit, which is configured to adjust the gray scale values obtained by the histogram equalization to the integer range, so that the gray scale values of the source images and the target images have a minimum difference and a mapping relationship is established, and a plurality of matching datasets corresponding to the plurality of partial annotation datasets are obtained according to the mapping relationship.

[0018] Optionally, the single-organ segmentation model collaborative learning module comprises:

[0019] The model training unit trains the single-organ segmentation model using the matching data set belonging to the same organ as the single-organ segmentation model.

[0020] The label generation unit inputs the matching data set not belonging to the same organ as the single-organ segmentation model to the trained single-organ segmentation model to generate pseudo labels and obtain a plurality of full-label data sets.

[0021] Optionally, the model training unit introduces a push-away loss in addition to the segmentation loss in the process of training the model.

[0022] The push-away loss includes a label difference loss and a feature difference loss.

[0023] The total loss function is calculated as follows:

[0024]

[0025]

[0026]

[0027]

[0028] wherein, is the segmentation loss calculated by the single-organ segmentation model, is the label difference loss output by the single-organ segmentation model, is the feature difference loss output by the single-organ segmentation model, n, m = 1, 2, … K, data set P m , P n corresponding real label represents the prediction result obtained by applying the model g m to the data set P n , λ col-pre and λ col-feature represent the cooperative learning hyperparameters of the single-organ segmentation model, and the loss calculated by the segmentation model on the data set P m is , and the KL divergence loss is represents the feature obtained by applying the segmentation model trained on the P m data set to the P n data set, represents the feature obtained by applying the segmentation model trained on the P n data set to the P n data set.

[0029] Optionally, the multi-organ segmentation model cooperative learning module comprises:

[0030] The multi-organ segmentation training unit cooperatively trains a plurality of multi-organ segmentation models by using a plurality of full-label data sets.

[0031] The loss calculation unit introduces a loss function in the model training process, including a segmentation loss and a pull loss.

[0032] Optionally, the pull loss in the loss calculation unit includes a label difference loss and a feature difference loss.

[0033] The calculation formula of the total loss function is as follows:

[0034]

[0035]

[0036]

[0037]

[0038] wherein, is a segmentation loss calculated by the multi-organ segmentation model, is a label difference loss output by the multi-organ segmentation model, is a feature difference loss output by the multi-organ segmentation model, is a full-labeled data set containing pseudo labels. denotes G m acts on the output result of the organ n generated. denotes the real label of the organ n in the G denotes G m acts on the extracted feature. beta col-pre and beta col-feature denote the cooperative learning hyperparameters of the multi-organ segmentation model.

[0039] Optionally, the multi-organ segmentation model distillation module takes a plurality of multi-organ models as teacher models, obtains a student model through a knowledge distillation technology, and takes the student model as the final multi-organ segmentation model. The loss function is defined as follows:

[0040]

[0041] wherein, G i is the multi-organ segmentation model of the third step, F is the multi-organ segmentation model obtained by final distillation, X t is a target domain data set, is the calculated KL divergence loss.

[0042] The technical effect of the present application is:

[0043] (1) Using a histogram matching method, images from different data sets have the same gray scale distribution, and the domain difference is eliminated in the spatial domain;

[0044] (2) Using collaborative learning between single organ segmentation models containing different labels, different organs contained in different data sets can be aware of each other, improving the performance of each single organ segmentation model and the quality of the pseudo label of the organ generated on other data sets;

[0045] (3) Using collaborative learning between multi-organ segmentation models containing pseudo labels and full labels, so that the labels contained in different data sets can be utilized, improving the performance of each multi-organ segmentation model. Finally, multiple models are distilled as a student model, which is the final multi-organ segmentation model;

[0046] (4) Using partially labeled data sets to implement a multi-organ segmentation model, greatly reducing the dependence on fully labeled data sets and alleviating the labeling difficulty of doctors. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein in their entirety. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0048] Figure 1 is a schematic diagram of the system structure in the embodiments of the present application; DETAILED DESCRIPTION

[0049] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0051] Embodiment one

[0052] As Figure 1As shown, the embodiment provides a multi-organ segmentation system based on partial annotation dataset collaborative learning, which includes four modules: (I) the histogram matching module realizes that multiple datasets from different centers have the same gray distribution, and eliminates domain differences in the spatial domain; (II) the single-organ segmentation model collaborative learning module trains multiple single-organ segmentation models, the single-organ models collaboratively learn, fully utilize the annotated data in other datasets, and generate pseudo labels of the organ for other datasets after training; (III) the multi-organ segmentation model collaborative learning module trains multiple multi-organ segmentation models based on the fully annotated dataset containing pseudo labels, the multi-organ segmentation models collaboratively learn, and fully utilize the labels in other datasets; (IV) the multi-organ segmentation model distillation module distills a student model as the final multi-organ segmentation model after the multiple multi-organ segmentation models are trained. The overall architecture of the system is as shown in Figure 1 As shown, the embodiment provides a multi-organ segmentation system based on partial annotation dataset collaborative learning, which includes four modules: (I) the histogram matching module realizes that multiple datasets from different centers have the same gray distribution, and eliminates domain differences in the spatial domain; (II) the single-organ segmentation model collaborative learning module trains multiple single-organ segmentation models, the single-organ models collaboratively learn, fully utilize the annotated data in other datasets, and generate pseudo labels of the organ for other datasets after training; (III) the multi-organ segmentation model collaborative learning module trains multiple multi-organ segmentation models based on the fully annotated dataset containing pseudo labels, the multi-organ segmentation models collaboratively learn, and fully utilize the labels in other datasets; (IV) the multi-organ segmentation model distillation module distills a student model as the final multi-organ segmentation model after the multiple multi-organ segmentation models are trained. The overall architecture of the system is as shown in

[0053] (I) The histogram matching module that realizes that multiple datasets from different centers have the same gray distribution and eliminates domain differences in the spatial domain: calculate the average gray distribution of the multi-organ dataset images Then each image in the single-organ dataset is matched with the average gray distribution to obtain the matched image;

[0054] (II) The single-organ segmentation model collaborative learning module that trains multiple single-organ segmentation models, the single-organ models collaboratively learn, fully utilize the annotated data in other datasets, and generate pseudo labels of the organ for other datasets after training: train multiple single-organ segmentation models, each model has a "push away" loss from other datasets in addition to the supervision loss of its own label during training, and each model generates pseudo labels of the organ for other datasets after training;

[0055] (III) The multi-organ segmentation model collaborative learning module that trains multiple multi-organ segmentation models based on the fully annotated dataset containing pseudo labels, the multi-organ segmentation models collaboratively learn, and fully utilize the labels in other datasets: train multiple multi-organ segmentation models using the fully annotated dataset containing pseudo labels, each model has a "pull close" loss from other datasets in addition to the supervision loss of its own label during training;

[0056] (IV) The multi-organ segmentation model distillation module that distills a student model as the final multi-organ segmentation model after the multiple multi-organ segmentation models are trained: after obtaining the multiple multi-organ segmentation models, these multi-organ segmentation models are used as teacher models, and a student model is extracted as the final multi-organ segmentation model.

[0057] The four modules are further described below.

[0058] The first module is a histogram matching module which eliminates domain difference in spatial domain and makes the multiple data sets from different centers have the same gray scale distribution. The working content of the module includes:

[0059] Given a source domain P = {P i | i = 1,..., K} includes K partially labeled data sets, where Any source domain image set is Where represents the data set P i , and the corresponding label image set is Given a target domain image set

[0060] First, normalize all images so that their gray scale values are integer values in the interval [0, L-1] (in the present application, L = 256). Then, average the gray scale of the target domain images Take as the matching standard. The normalized histogram of P l is defined as p l (g l ) = n l / HW, l = 0,..., L-1, where g l represents the lth intensity value, n l is the number of pixels in image with intensity g q , and H and W are the number of rows and columns of image . Similarly, the normalized histogram of the target image is represented as p q (g q ), q = 0, 1,..., L-1.

[0061] Then, perform histogram equalization on the histogram of the source image , and round the obtained gray scale value to the integer range [0, L-1]. Here, the histogram equalization is defined as the transformation T1, T2 of the following formula:

[0062]

[0063] Similarly, perform histogram equalization on the histogram p q (g q ) of the average target intensity image , and adjust the obtained gray scale value to the integer range [0, L-1]:

[0064]

[0065] For each value, to make and as close as possible, find the corresponding g q value and store the mapping. Finally, according to the mapping relationship between the found source domain image and the gray value of the target domain average image, the histogram matched image

[0066] (ii) Single organ segmentation model collaborative learning module (II), the working content includes:

[0067] After completing the histogram matching, K single organ segmentation models {g m |m = 1,..., K} can be trained on the K partially labeled data sets in the source domain, and then the trained models are used to generate pseudo labels for other data sets to obtain K fully labeled data sets. Specifically, in the training process of each single organ segmentation model, in addition to the main segmentation loss, there is also a collaborative loss from other data sets, which we call the push-away loss, including label-level loss and feature-level loss.

[0068] Taking the kidney and liver as an example, the main loss of the kidney segmentation model is calculated from the network prediction result and the real label of the kidney. When the kidney segmentation model acts on the liver data set, the prediction result of the kidney will be obtained. Since different organs are mutually exclusive, the prediction result of the kidney and the label of the liver should not overlap. Therefore, we propose a difference loss at the label level . In addition, a difference loss at the feature level is introduced to push away the features extracted by the kidney model in the liver data set from the features extracted by the liver model. Through the push-away loss, not only can the kidney be more accurately segmented, but the kidney model can also perceive the existence of the liver, thereby improving the quality of generating pseudo labels in the next stage. Extending it to all training data sets, the push-away loss at two levels is calculated on any two data sets. Given the mth single organ segmentation model g m , the specific calculation of the loss function is as follows:

[0069]

[0070]

[0071]

[0072]

[0073] wherein, segmentation loss calculated for single-organ segmentation model, label difference loss output by single-organ segmentation model, feature difference loss output by single-organ segmentation model, n, m = 1, 2, … K, dataset P m , P n corresponding real label denotes the prediction result obtained by applying the model g m to the dataset P n , λ col-pre and λ col-feature denote the cooperative learning hyperparameters of the single-organ segmentation model, the loss calculated for the segmentation model on the dataset P m is , the KL divergence loss is denotes the feature obtained by applying the segmentation model trained on the P m dataset to the P n dataset, denotes the feature obtained by applying the segmentation model trained on the P n dataset to the P n dataset.

[0074] (III) Multi-organ model cooperative learning module (III), the working content of which includes:

[0075] After the second stage, we obtain the full-labeled dataset containing pseudo labels where In the set only one organ label is true, and the other labels are pseudo labels generated by other models. In order to make full use of the labels in each dataset and improve the performance of each multi-organ segmentation model, we perform cooperative learning to train multiple multi-organ segmentation models {G m | m = 1, …, K}. Specifically, in addition to the main segmentation loss , each multi-organ segmentation model also has cooperative loss with other models, which we call “pulling close” loss, which also includes label-level loss and feature-level loss Take the liver, kidney and pancreas in the kidney dataset as an example, the kidney label is real, and the liver and pancreas labels are both pseudo labels generated by other models. The main segmentation loss on the kidney dataset is calculated between the prediction result of the kidney multi-organ segmentation model G kid and all its three labels. The liver and pancreas multi-organ segmentation models G liv and G pan both calculate the label-level similarity loss on the kidney dataset, and the similarity loss includes two parts, for G livOn the one hand, we calculate the segmentation loss between its kidney prediction results and the true kidney label; on the other hand, we calculate the segmentation loss between its liver and pancreas prediction results and G kid The segmentation loss between the liver and pancreas prediction results is pan The same is true for G. liv and G pan The features extracted on the kidney dataset should be consistent with G kid The extracted features are similar. Through similarity loss, the predictions and features of different multi-organ segmentation models are often the same, thereby further reducing the domain differences at the model level. The specific calculation of the loss function is as follows:

[0076]

[0077]

[0078]

[0079]

[0080] in, Segmentation loss calculated for the multi-organ segmentation model, is the label difference loss output by the multi-organ segmentation model, is the feature difference loss output by the multi-organ segmentation model, It is a fully annotated dataset with pseudo labels. Represents G m Acts on The output of organ n is generated. express The true label of organ n, Indicates G m Acts on The extracted features. col-pre and β col-feature Represents the collaboratively learned hyperparameters of the multi-organ segmentation model.

[0081] (IV) Multi-organ segmentation model distillation module (IV), which includes:

[0082] After obtaining multiple multi-organ segmentation models, each model focuses on different aspects. For example, in the fully annotated kidney dataset with pseudo labels, the kidney label is the real label, while the other labels are pseudo labels, so the kidney segmentation accuracy is higher. In order to better learn the knowledge of the multi-organ model, we use the multi-organ model as the teacher model and distill a student model together as our multi-organ segmentation model F, the loss function The definition is as follows:

[0083]

[0084] wherein G i is the multi-organ segmentation model of the third step, F is the multi-organ segmentation model obtained by final distillation, X t is the target domain data set, is the calculated KL divergence loss.

[0085] The present application is trained on three single-organ data sets of liver, kidney and pancreas, and is verified on the BTCV data set. Table 1 shows the role of different modules in segmentation, and finally the DSC value of liver segmentation of our method is 92.87, the DSC value of kidney segmentation is 93.45, and the DSC value of pancreas segmentation is 76.58.

[0086] Compared with the method of directly generating pseudo labels, the pseudo labels obtained by the method of the present embodiment are of higher quality.

[0087] Table 2 gives the segmentation accuracy after adding label-level loss and feature-level loss in the push-pull loss, and from the data in the table, it can be seen that the label-level loss and the feature-level loss play a certain role in improving the segmentation accuracy.

[0088] Table 3 gives the comparison results of our method and other methods, and the single training is directly training three single-organ models acting on the target data set, and the joint training is training three single-organ models, then each model generates pseudo labels of the organ for other data sets, and then trains a multi-organ segmentation model based on the pseudo labels, DoDNet and Co-training are methods proposed in 2019 to solve the problem of partially labeled multi-organ segmentation, and from the data in the table, it can be seen that the present method is obviously superior to the comparison methods, and the segmentation accuracy of each organ is improved to a certain extent. The results of the ablation test of each module are shown in Table 1.

[0089] Table 1

[0090]

[0091] Table 2

[0092]

[0093] Table 3

[0094]

[0095] Embodiment Two

[0096] The present application provides a multi-organ segmentation system based on partial labeled data set collaborative learning, comprising:

[0097] The histogram matching module, the single-organ segmentation model cooperative learning module, the multi-organ segmentation model cooperative learning module, and the multi-organ segmentation model distillation module;

[0098] The histogram matching module is configured to process images in the partial labeling data sets to obtain a plurality of matching data sets.

[0099] The single-organ segmentation model cooperative learning module is configured to obtain a plurality of full labeling data sets according to the single-organ segmentation model and the plurality of matching data sets.

[0100] The multi-organ segmentation model cooperative learning module is configured to train a plurality of multi-organ segmentation models according to the plurality of full labeling data sets.

[0101] The multi-organ segmentation model distillation module is configured to learn a final multi-organ segmentation model from the plurality of multi-organ segmentation models.

[0102] Specifically, the histogram matching module comprises:

[0103] The gray scale distribution calculation unit is configured to normalize all images in a source domain and a target domain to make the gray scale values of the images be integer values in a preset interval, and then calculate an average gray scale distribution of the images in the target domain to obtain an average gray scale distribution of the images in the multi-organ data set.

[0104] The equalization unit is configured to define a corresponding normalized histogram based on the number of rows and columns of the images in the source domain and the target domain and the number of pixels of different intensity values, and perform histogram equalization on the normalized histogram.

[0105] The matching unit is configured to adjust the gray scale values obtained by the histogram equalization to the integer range, so that the gray scale values of the source images and the target images have a minimum difference and a mapping relationship is established, and a plurality of matching data sets corresponding to the plurality of partial labeling data sets are obtained according to the mapping relationship.

[0106] Specifically, the single-organ segmentation model cooperative learning module comprises:

[0107] The model training unit is configured to train the single-organ segmentation model using the matching data set belonging to the same organ as the single-organ segmentation model.

[0108] The label generation unit is configured to input the matching data set not belonging to the same organ as the single-organ segmentation model to the trained single-organ segmentation model to generate pseudo labels and obtain a plurality of full labeling data sets.

[0109] Specifically, the model training unit introduces a push-away loss in addition to a segmentation loss in the process of training the model.

[0110] The push-away loss includes a label difference loss and a feature difference loss.

[0111] The total loss function The calculation formula is as follows:

[0112]

[0113]

[0114]

[0115]

[0116] Wherein, is the segmentation loss calculated by the single organ segmentation model, is the label difference loss output by the single organ segmentation model, is the feature difference loss output by the single organ segmentation model, n, m = 1, 2,..., K data set P m , P n Corresponding real label Indicates the prediction result obtained by the model g m acting on the data set P n , λ col-pre And λ col-feature Indicate the cooperative learning hyperparameters of the single organ segmentation model, the loss calculated by the segmentation model on the data set P m is , the KL divergence loss is Indicates the feature obtained by the segmentation model trained on the P m data set acting on the P n data set, Indicates the feature obtained by the segmentation model trained on the P n data set acting on the P n data set.

[0117] Specifically, the multi-organ segmentation model cooperative learning module comprises:

[0118] A multi-organ segmentation training unit cooperatively trains a plurality of multi-organ segmentation models using a plurality of full-label data sets;

[0119] A loss calculation unit introduces a loss function in the model training process, including a segmentation loss and a pull-in loss.

[0120] Specifically, the pull-in loss in the loss calculation unit includes a label difference loss and a feature difference loss.

[0121] The calculation formula of the total loss function is as follows:

[0122]

[0123]

[0124]

[0125]

[0126] wherein, a segmentation loss calculated for the multi-organ segmentation model, a label difference loss output by the multi-organ segmentation model, a feature difference loss output by the multi-organ segmentation model, a full-labeled dataset containing pseudo labels. denotes G m acts on the output result of the organ n generated. denotes the true label of the organ n in the denotes G m acts on the extracted feature. β col-pre and β col-feature denote the cooperative learning hyperparameters of the multi-organ segmentation model.

[0127] Specifically, the multi-organ segmentation model distillation module takes several multi-organ models as teacher models, obtains a student model through the knowledge distillation technology, and takes the student model as the final multi-organ segmentation model. The loss function is defined as follows:

[0128]

[0129] wherein, G i is the multi-organ segmentation model of the third step, F is the multi-organ segmentation model obtained by the final distillation, X t is the target domain dataset, is the calculated KL divergence loss.

[0130] The above merely describes the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any changes or replacements easily thought of by those skilled in the art within the technical range disclosed by the present application should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-organ segmentation system based on collaborative learning of partially labeled datasets, characterized by: include: Histogram matching module, single-organ segmentation model collaborative learning module, multi-organ segmentation model collaborative learning module, multi-organ segmentation model distillation module; The histogram matching module is used to process images in several partially labeled data sets to obtain several matching data sets; The single organ segmentation model collaborative learning module is used to obtain a plurality of fully labeled data sets according to the single organ segmentation model and the plurality of matching data sets; The multi-organ segmentation model collaborative learning module is used to train multiple multi-organ segmentation models according to the multiple fully labeled data sets; The multi-organ segmentation model distillation module is used to learn from multiple multi-organ segmentation models to obtain a final multi-organ segmentation model; The histogram matching module includes: a grayscale distribution calculation unit, which, given a source domain and a target domain, normalizes all images in the source domain and the target domain so that the grayscale values ​​of the images are integers within a preset range, and then averages the grayscale values ​​of the images in the target domain to obtain an average grayscale distribution of images in a plurality of organ datasets; the source domain includes a plurality of partially labeled datasets; An equalization unit, defining corresponding normalized histograms based on the number of rows and columns of the source domain image and the target domain image, and the number of pixels with different intensity values, respectively; and performing histogram equalization on the normalized histograms; A matching unit adjusts the grayscale values ​​obtained by histogram equalization to an integer range so that there is a minimum difference between the grayscale values ​​of the source image and the grayscale values ​​of the target image, establishes a mapping relationship, and obtains a plurality of matching data sets corresponding to the plurality of partially labeled data sets according to the mapping relationship; The single organ segmentation model collaborative learning module includes: The model training unit uses a matching dataset belonging to the same organ as the single organ segmentation model to train the single organ segmentation model; The label generation unit inputs the matching dataset that is not assigned to the same organ as the single-organ segmentation model into the trained single-organ segmentation model to generate pseudo labels and obtain several fully labeled datasets; In the process of model training, the model training unit introduces the push-off loss in addition to the segmentation loss; The push-off loss includes label difference loss and feature difference loss; Total loss function The calculation formula is as follows: in, Segmentation loss calculated for the single organ segmentation model, is the label difference loss output by the single organ segmentation model, is the feature difference loss output by the single organ segmentation model, n,m=1,2,...K, dataset P m ,P n The corresponding true label Represented by model g m Acting on dataset P n The prediction results obtained, λ col-pre and λ col-feature represents the collaborative learning hyperparameters of the single organ segmentation model, dataset P m The loss calculated by the segmentation model on The KL divergence loss is Indicates that in P m The segmentation model trained on the dataset acts on P n The features obtained on the dataset, Indicates that in P n The segmentation model trained on the dataset acts on P n Features obtained on the dataset; The multi-organ segmentation model collaborative learning module includes: Multi-organ segmentation training unit, which uses several fully labeled datasets to collaboratively train several multi-organ segmentation models; Loss calculation unit, which introduces loss functions during model training, including segmentation loss and approximation loss; The closing loss in the loss calculation unit includes label difference loss and feature difference loss; The calculation formula of the total loss function is as follows: in, Segmentation loss calculated for the multi-organ segmentation model, is the label difference loss output by the multi-organ segmentation model, is the feature difference loss output by the multi-organ segmentation model, is a fully annotated dataset with pseudo labels, Indicates G m Acts on The output of organ n is generated, express The true label of organ n, Indicates G m Acts on The extracted features, β col-pre and β col-feature Represents the collaboratively learned hyperparameters of the multi-organ segmentation model.

2. The multi-organ segmentation system based on collaborative learning of partially labeled datasets according to claim 1 is characterized in that The multi-organ segmentation model distillation module uses several multi-organ models as teacher models, obtains a student model through knowledge distillation technology, and uses the student model as the final multi-organ segmentation model.

Images