CN109344833B - Medical image segmentation method, segmentation system and computer-readable storage medium - Google Patents

Abstract

The present disclosure provides a medical image segmentation method, a segmentation system, and a computer-readable storage medium, the method including the steps of: establishing a network model, and training the network model to obtain a trained network model; and predicting the image level labels and the pixel level labels by using the trained model, and finally realizing accurate segmentation of the image by using the predicted image level labels and the pixel level labels and the feature mapping of the reconstructed image. The technical scheme that this application provided has the advantage of accurate segmentation.

Description

Medical image segmentation method, segmentation system and computer-readable storage medium

Technical Field

The invention relates to the technical field of computers and medical treatment, in particular to a medical image segmentation method, a medical image segmentation system and a computer-readable storage medium.

Background

During the medical treatment, accurate medical image segmentation can help doctors to better diagnose and treat patients. In a medical image segmentation model based on deep learning, the dependence on a large amount of pixel-level annotation data is obvious. However, the image acquisition of the pixel-level labeling is difficult, and the segmentation precision and generalization capability of the model are severely limited. The training data that can be utilized in practice are a small number of pixel-level annotation samples and a large number of low-cost image-level annotation samples. The image-level labeling sample lacks description of pixel-level information, and only image-level features of an image can be extracted in the traditional supervised learning process, so that the pixel-level features are difficult to extract, and the training model is difficult to realize pixel-level segmentation of the image.

Disclosure of Invention

The embodiment of the invention provides a medical image segmentation method, a segmentation system and a computer readable storage medium, which can realize pixel-level segmentation of an image.

In a first aspect, an embodiment of the present invention provides a medical image segmentation method, including the following steps:

establishing a network model, and training the network model to obtain a trained network model;

and predicting the image level labels and the pixel level labels by using the trained model, and finally realizing accurate segmentation of the image by using the predicted image level labels and the pixel level labels and the feature mapping of the reconstructed image.

A medical image segmentation system, the system comprising:

the establishing module is used for establishing a network model and training the network model to obtain a trained network model;

the prediction module is used for predicting the image-level labels and the pixel-level labels by using the trained model;

and the segmentation module is used for realizing accurate segmentation of the image by utilizing the predicted image level label and the pixel level label and the feature mapping of the reconstructed image.

In a third aspect, a computer-readable storage medium is provided, which stores a program for electronic data exchange, wherein the program causes a terminal to execute the method provided in the first aspect.

Compared with the prior art, the scheme provided by the invention constructs the image reconstruction module, extracts the image reconstruction characteristics, makes up the deficiency of image-level label information and reduces the uncertainty influence of pixel-level labels. The invention solves the problem of neglecting the individual characteristics of the image in the existing image segmentation method, provides an individual learning framework based on the combination of the common characteristics and the individual characteristics, and improves the segmentation precision of the pixel level image. According to the scheme provided by the invention, the image reconstruction model is continuously utilized to carry out personalized learning in the testing stage, the influence of the image reconstruction model on the segmentation precision is analyzed, and personalized image segmentation is realized for image data. The network disclosed by the patent can solve the problem of pixel level segmentation of a target image under the condition of high precision on the basis of fusing image level features and other auxiliary features by mining the mapping relation between the image level tags and the pixel level tags.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

Fig. 1 is a technical flow diagram of a medical image segmentation method.

Fig. 1a is a schematic flow chart of a medical image segmentation method provided in the present application.

Fig. 1b is a schematic structural diagram of a medical image segmentation system provided in the present application.

Fig. 2 is a diagram of a medical image segmentation network architecture based on multitask learning.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The terms "first," "second," "third," and "fourth," etc. in the description and claims of the invention and in the accompanying drawings are used for distinguishing between different objects and not for describing a particular order. Furthermore, the terms "include" and "have," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those steps or elements listed, but may alternatively include other steps or elements not listed, or inherent to such process, method, article, or apparatus.

Reference herein to "an embodiment" means that a particular feature, result, or characteristic described in connection with the embodiment can be included in at least one embodiment of the invention. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is explicitly and implicitly understood by one skilled in the art that the embodiments described herein can be combined with other embodiments.

Medical images are complex and various, and the segmentation boundaries are different. Accurate segmentation of regions in medical images is achieved, and a segmentation model is trained by medical image data with large quantity, high quality and pixel labels. However, most images archived in hospitals have only image tags, or even no tags. The high-quality medical image pixel label needs manual labeling by a practitioner, is time-consuming, labor-consuming, very expensive and difficult to obtain on a large scale. In order to fully utilize medical image data of a large number of image labels and improve image segmentation precision, a medical image segmentation scheme based on dual learning is provided.

Currently, the most common examination modality in hospitals is primarily magnetic resonance imaging. Among them, the magnetic resonance Diffusion Tensor Imaging (DTI) is a latest detection technology, which can provide a variety of signals to quantitatively evaluate the damaged condition of the spinal cord, and can reflect the change of the microstructure of the spinal cord, and has higher sensitivity compared with the conventional T2 WI. The diagnosis of medical image information is achieved by detecting whether a region of interest (ROI) signal is abnormally changed or not, and the like. The accuracy of the region of interest segmentation directly affects the diagnosis. For the purpose of improving the diagnosis effect, the scheme provides a novel medical image segmentation algorithm for segmenting the region of interest. The model can be trained by only needing a small number of images with pixel-level labels and a large amount of data with image-level labels. According to the scheme, the expensive pixel-level image labeling work can be reduced, and meanwhile, the description of the image in the diagnosis process of a doctor can be extracted as an image label, so that the image data of the image-level label can be acquired in a large scale.

Referring to fig. 1a, fig. 1a provides a medical image segmentation method, which, referring to fig. 1a, comprises the steps of:

s101, establishing a network model, and training the network model to obtain a trained network model;

the network model in step S101 includes: an infrastructure network A and an infrastructure network B, the infrastructure network A comprising: a deep residual network and a dense convolutional neural network; the base network B includes: a capsule network structure model.

And S102, predicting the image level labels and the pixel level labels by using the trained model, and finally realizing accurate segmentation of the image by using the predicted image level labels and the pixel level labels and the feature mapping of the reconstructed image.

According to the scheme provided by the invention, an image reconstruction module is constructed, the image reconstruction characteristics are extracted, the defect of image-level label information is made up, and the uncertainty influence of pixel-level labels is reduced. The invention solves the problem of neglecting the individual characteristics of the image in the existing image segmentation method, provides an individual learning framework based on the combination of the common characteristics and the individual characteristics, and improves the segmentation precision of the pixel level image. According to the scheme provided by the invention, the image reconstruction model is continuously utilized to carry out personalized learning in the testing stage, the influence of the image reconstruction model on the segmentation precision is analyzed, and personalized image segmentation is realized for image data. The network disclosed by the patent can solve the problem of pixel level segmentation of a target image under the condition of high precision on the basis of fusing image level features and other auxiliary features by mining the mapping relation between the image level tags and the pixel level tags.

Optionally, the training of the network model in step S101 to obtain a trained network model specifically includes:

tagging data with high quality pixel level I^f，L^f，T^fAdjusting parameters of the basic network A, the pixel label prediction network B, the image label prediction network C and the image reconstruction network D through a minimum loss function; the Loss function comprises a Loss function Loss corresponding to pixel label prediction^map(Bo，L^f) Loss function Loss of image tag^tag(Co，T^f) Euclidian Loss function Loss for image reconstruction^img(Do，I^f) (ii) a Wherein Bo, Co and Do represent the predicted pixel-level label, image-level label and reconstructed image, respectively;

annotating data with image level { I) that is easier to acquire^w，T^wAnd realizing further training of the model by minimizing an image label loss function and an image reconstruction loss function, and realizing pixel-level label L by integrating image-level label information after further training is finished^wPredicting; after the pixel level label information of the weakly marked data is obtained, the predicted L is calculated^wAnd as a true value, converting weakly labeled image data into fully labeled image data, and performing supervised training on the whole network model by using all the image data to obtain a trained network model.

Optionally, the supervised training includes:

according to the multi-step training method of multi-task learning, the parameter change of the network is smooth, and the Loss function comprises a softmax Loss function Loss predicted by a pixel label^map(Ao,L^f) Softmax Loss function Loss of image tag^tag(Bo,T^f) And Euler Loss function Loss of image reconstruction^img(Co,I^f)；

In the semi-supervised learning phase, we fine-tune the model with all full-label and weak-label image data. All image data may be represented as { I, L, T }, where I ═ I^f,I^w}，L＝{L^f,L^wT ═ T^f,T^wThen the optimization problem for the whole network can be converted into:

where θ represents a parameter of the entire network, α₁,β₁Representing the parameters of the output and hidden layers a1 and B1, respectively, of the DPN.

Referring to fig. 1b, fig. 1b provides a medical image segmentation system comprising:

the establishing module 201 is used for establishing a network model, and training the network model to obtain a trained network model;

the prediction module 202 is configured to predict the image-level labels and the pixel-level labels by using the trained model;

and the segmentation module 203 is used for realizing accurate segmentation of the image by utilizing the predicted image-level label and the pixel-level label and the feature mapping of the reconstructed image.

Optionally, the establishing module 201 is specifically configured to utilize the heightQuality of pixel level labeling data { I^f，L^f，T^fAdjusting parameters of the basic network A, the pixel label prediction network B, the image label prediction network C and the image reconstruction network D through a minimum loss function; the Loss function comprises a Loss function Loss corresponding to pixel label prediction^map(Bo，L^f) Loss function Loss of image tag^tag(Co，T^f) Euclidian Loss function Loss for image reconstruction^img(Do，I^f) (ii) a Wherein Bo, Co and Do represent the predicted pixel-level label, image-level label and reconstructed image, respectively;

annotating data with image level { I) that is easier to acquire^w，T^wRealizing further training of the model by minimizing an image label loss function and an image reconstruction loss function, and realizing prediction of a pixel-level label Lw by integrating image-level label information after the further training is finished; after the pixel level label information of the weakly marked data is obtained, the predicted L is calculated^wAnd as a true value, converting weakly labeled image data into fully labeled image data, and performing supervised training on the whole network model by using all the image data to obtain a trained network model.

The application of the deep convolutional neural network in the image task is greatly developed, and the application performance of the deep neural network is further expanded by the deep residual error network (ResNet) and the intensive convolutional neural network (DenseNet). The dual-channel network (DPN) integrates the advantages of ResNet and DenseNet, and the two channel networks are respectively used as the two channel networks to be fused, so that the existing characteristics can be recycled, new characteristics can be explored, and the basic framework can be expressed as follows:

h^k＝g^k(r^k), (4)

wherein x is^kAnd y^kRespectively representing the information extracted at the k-th step for each channel, f_t ^k(. and v)_t(. h) is a feature learning function, equation (1) represents dense connected channels so that the channel can explore new features, equation (2) represents reuse of common features in residual channels, equation (3) fuses features in dense channels with features in residual channels, the final feature transformation is implemented in equation (4), h^kRepresenting the current state that can be used for mapping or prediction for the next step.

In order to utilize data only with an image-level label, the traditional method directly utilizes the image label to predict the pixel label to realize image segmentation, but the information contained in the image label is too simple to recover the more accurate pixel label. When the segmentation model is trained based on a small amount of fully labeled image data and a large amount of weakly labeled image data, the image-level label of the weakly labeled image data lacks description of each pixel point, so that the image-level label has larger uncertainty when being expanded into the pixel-level label. Existing studies rarely consider the effect of this uncertainty on pixel-level segmentation.

Therefore, in order to fully utilize information contained in a picture, an input original image is reconstructed on the basis of auxiliary prediction of a pixel level label by using an image level label, the reconstructed picture is compared with the input original image, network model parameters are further optimized, dependence of medical image segmentation on pixel level annotation data amount is reduced, the cost of a digital image set is reduced, and pixel level segmentation of the image is realized by using a small amount of pixel level image annotation data and a large amount of image level annotation data.

According to the scheme provided by the invention, the used data set comprises a small amount of high-quality pixel label data sets and a large amount of image label data sets, and four network modules are constructed: the method comprises a basic network A, a pixel label prediction network B, an image label prediction network C and an image reconstruction network D. The basic content is divided into the following four steps:

1. and (5) constructing and initializing a network model. The basic network A adopts a two-channel network, and comprises a deep residual error network and a dense convolutional neural network. The two networks have more hidden layers and are pre-trained with a common data set in order to obtain better initialization parameters. And the pixel label prediction network B adopts a capsule network structure model to extract pixel-level characteristics. And the image label prediction network C adopts a convolutional neural network to extract image-level features and perform corresponding label prediction. And the network D firstly fuses the output layers of the networks B and C, and then adopts a multilayer convolutional neural network to realize the reconstruction of the image. The number of layers of networks B, C and D is small compared to the base network a, so that its parameters can be randomly initialized.

2. And (5) initially training the network model. Tagging data with high quality pixel level I^f，L^f，T^fAnd adjusting parameters of the basic network A, the pixel label prediction network B, the image label prediction network C and the image reconstruction network D through a minimum loss function. The Loss function comprises a Loss function Loss corresponding to pixel label prediction^map(Bo，L^f) Loss function Loss of image tag^tag(Co, Tf) and Euclidian Loss function Loss for image reconstruction^img(Do，I^f). Where Bo, Co and Do represent the predicted pixel level label, the image level label and the reconstructed image, respectively.

3. Pixel level label prediction. To label data with easier-to-acquire image level I^w，T^wAnd (5) realizing further training of the model by minimizing an image label loss function and an image reconstruction loss function. After the training is finished, pixel-level label L is realized by integrating image-level label information^wAnd (4) predicting.

4. And converting into full-labeled data. After the pixel level label information of the weakly marked data is obtained, the predicted L is calculated^wAs a true value, weakly annotated image data is converted into fully annotated image data. Thus, the entire network model is modeled using the entire image dataAnd (5) supervision training. All image data may be represented as { I, L, T }, where I ═ I^f，I^w}，L＝{L^f，L^wT ═ T^f，T^w}. Then the optimization problem for the entire network can be translated into:

argmin_θ{Loss^maP(Bo,L)+Loss^tag(Co,T)+Loss^img(Do,I))

where theta represents a parameter of the entire network. The optimization of all network model parameters can be realized by optimizing the objective functions.

Through the converted full label data, the scheme updates the parameters of the network structure by using the loss function of the minimum reconstructed image, and after all network parameters are updated, the pixel label of a new picture can be predicted, so that the pixel level segmentation of the image is realized.

For a small number of fully annotated images { I^f，L^f，T^fAnd a large number of weakly labeled images with only image-level labels I^w，T^wThe scheme of the invention provides an individualized segmentation model based on weak supervised learning.

The model comprises four modules: the system comprises a basic network A, a pixel label prediction network B, an image label prediction network C and an image reconstruction network D.

The scheme of the invention comprises a technical process as shown in figure 1. The method comprises a model construction stage, a model training stage and a model testing stage:

stage of model construction

The model adopts a multi-task deep learning framework, realizes the training and prediction of the model by optimizing three target tasks, and the output layer of the model comprises the output of an image label, the output of a pixel label and the output of a reconstructed image. Each output layer corresponds to three sub-networks, which are abbreviated as a network, B network and C network, and the three sub-networks are built on the basic skeleton network, and the model structure is shown in fig. 2.

In order to utilize data only with an image-level label, the traditional method directly utilizes the image label to predict the pixel label to realize image segmentation, but the information contained in the image label is too simple to recover the more accurate pixel label. Therefore, in order to fully utilize the information contained in the picture, the input original image is reconstructed on the basis of the auxiliary prediction pixel level label by using the image level label, and the reconstructed picture is compared with the input original image, so that the network model parameters are further optimized. Thus, we construct three sub-networks, corresponding to the a, B and C networks in fig. 2, respectively. Each subtask can adjust the parameters of the skeleton network through the sharing of the basic skeleton network. The network model constructed is specifically set forth as follows:

a dense convolutional network and a residual convolutional network are adopted in the network A, new features are intensively searched in a connected mode, the flexibility of the features is improved, and a small network structure model can realize the representation learning of a complex feature space. The deep residual error network can repeatedly utilize the previous characteristics, reduce model redundancy and reduce training difficulty.

And the B network adopts a capsule network mechanism. Capsule pass samples labeled image level { I^w，T^wPerforming convolution operation to obtain a characteristic image, wherein the characteristic image comprises a convolution layer, a Primarycaps layer, a Digitcaps layer and a decoding layer, each capsule represents a function, activated vectors are output, and the vectors represent pixel level segmentation labels searched by the capsules.

And the C network structure is a convolutional neural network, and the convolutional neural network structure comprises a convolutional layer, a down-sampling layer and a full-connection layer. Each layer has a plurality of feature maps, each feature map extracting a feature of the input by a convolution filter, each feature map having a plurality of neurons. And enhancing the original signal characteristics and reducing noise through convolution operation, and extracting the image level label of the training data.

Model training phase

Because the data set is divided into full label image data and weak label image data with only image-level labels, in the training process, the training is carried out in three stages: firstly, pre-training (supervised learning) a network model by using data of high-quality pixel-level labels; second, fine tuning is then performed with all data (semi-supervised learning). And thirdly, converting the full label data.

And (3) a supervision learning stage:

in a full-supervised learning training stage, image data with full labels are adopted to train a model, and the training process comprises three steps:

firstly, training a DPN, an A network and a B network simultaneously by using an image label and a pixel label;

fixing the above learned network parameters, and learning the parameters of the C network by reconstructing the image;

and updating the parameters of the network structure simultaneously.

According to the multi-step training method of multi-task learning, the parameter change of the network is smooth. The Loss function comprises a softmax Loss function Loss predicted by a pixel label^map(Ao,L^f) Softmax Loss function Loss of image tag^tag(Bo,T^f) And Euler Loss function Loss of image reconstruction^img(Co,I^f)。

A semi-supervised learning stage:

in the semi-supervised learning phase, we fine-tune the model with all full-label and weak-label image data. All image data may be represented as { I, L, T }, where I ═ I^f,I^w}，L＝{L^f,L^wT ═ T^f,T^w}. Then the optimization problem for the entire network can be translated into:

where θ represents a parameter of the entire network, α₁,β₁Representing the parameters of the output and hidden layers a1 and B1, respectively, of the DPN.

And (3) full tag data conversion stage:

for full-label data, we can achieve fine-tuning of all network model parameters by directly optimizing the first part of the equation. For weak label image data, missing pixel label L^wCan utilize image level tags T^wAnd (5) assisting in inference. Fixing other parameters, updating parameters of both A1 and B1 network structures with a function that minimizes the loss of reconstructed images, therebyAnd predicting the pixel label and the image label, and updating all network parameters by using the predicted pixel label and the predicted image label. After all network parameters are updated, pixel label L is carried out on a new picture^wAnd (4) completing the training phase of the model.

Stage of model testing

In the testing stage of the model, based on the trained model, preliminary segmentation of the medical image can be realized by fusing image-level label information and pixel-level label information. The trained model can extract the common characteristics of the test data set for image segmentation, and the individual characteristics of different image data are ignored. Therefore, in order to make the segmentation more accurate, it is necessary to consider the personality characteristics of the test image.

In the testing stage, the input image is reconstructed, the image reconstruction loss function is minimized, the personalized network model learning of the testing image is realized, then the learned model is used for predicting the image level label and the pixel level label, and finally the predicted image level label, the predicted pixel level label and the feature mapping of the reconstructed image are used for realizing the accurate segmentation of the image.

Embodiments of the present invention also provide a computer storage medium, wherein the computer storage medium stores a computer program for electronic data exchange, and the computer program enables a computer to execute part or all of the steps of any one of the medical image segmentation methods as recited in the above method embodiments.

It should be noted that, for simplicity of description, the above-mentioned method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present invention is not limited by the order of acts, as some steps may occur in other orders or concurrently in accordance with the invention. Further, those skilled in the art should also appreciate that the embodiments described in the specification are exemplary embodiments and that the acts and modules illustrated are not necessarily required to practice the invention.

In the foregoing embodiments, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus may be implemented in other manners. For example, the above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units is only one type of division of logical functions, and there may be other divisions when actually implementing, for example, a plurality of units or components may be combined or may be integrated into another system, or some features may be omitted, or not implemented. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection of some interfaces, devices or units, and may be an electric or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit may be implemented in the form of hardware, or may be implemented in the form of a software program module.

The integrated units, if implemented in the form of software program modules and sold or used as stand-alone products, may be stored in a computer readable memory. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a memory and includes several instructions for causing a computer device (which may be a personal computer, a server, a network device, or the like) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned memory comprises: a U-disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a removable hard disk, a magnetic or optical disk, and other various media capable of storing program codes.

Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by associated hardware instructed by a program, which may be stored in a computer-readable memory, which may include: flash Memory disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

The above embodiments of the present invention are described in detail, and the principle and the implementation of the present invention are explained by applying specific embodiments, and the above description of the embodiments is only used to help understanding the method of the present invention and the core idea thereof; meanwhile, for a person skilled in the art, according to the idea of the present invention, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present invention.

Claims (7)

1. A method of medical image segmentation, characterized in that the method comprises the steps of:

establishing a network model, and training the network model to obtain a trained network model;

predicting the image-level labels and the pixel-level labels by using the trained model, and finally realizing accurate segmentation of the image by using the predicted image-level labels and the pixel-level labels and the feature mapping of the reconstructed image;

the training of the network model to obtain the trained network model specifically comprises:

labeling data with pixel level { I^f，L^f，T^fParameter adjustment is carried out on the basic network, the pixel label prediction network, the image label prediction network and the image reconstruction network through a minimum loss function; the Loss function comprises a Loss function Loss corresponding to pixel label prediction^map(Bo，L^f) Loss function Loss of image tag^tag(Co，T^f) Euclidian Loss function Loss for image reconstruction^img(Do，I^f) (ii) a Wherein Bo, Co and Do represent the predicted pixel-level label, image-level label and reconstructed image, respectively;

annotating data { I) with image level^w，T^wRealizing further training of the model by minimizing an image label loss function and an image reconstruction loss function, and realizing pixel-level label L by integrating image-level label information after further training is finished^wPredicting; after the pixel level label information of the weakly marked data is obtained, the predicted L is calculated^wAnd as a true value, converting weakly labeled image data into fully labeled image data, and performing supervised training on the whole network model by using all the image data to obtain a trained network model.

2. The method of claim 1,

the network model includes: an infrastructure network A and an infrastructure network B, the infrastructure network A comprising: a deep residual network and a dense convolutional neural network; the base network B includes: a capsule network structure model.

3. The method of claim 1, the supervised training comprising:

according to the multi-step training method of multi-task learning, the parameter change of the network is smooth, and the Loss function comprises a softmax Loss function Loss predicted by a pixel label^map(Ao，L^f) Softmax Loss function Loss of image tag^tag(Bo，T^f) And Euler Loss function Loss of image reconstruction^img(Co，I^f)；

In the semi-supervised learning phase, we fine-tune the model using all full-labeled and weak-labeled image data, all denoted as { I, L, T }, where I ═ I {^f，I^w}，L＝{L^f，L^wT ═ T^f，T^wThen the optimization problem of the whole network is transformedComprises the following steps:

where θ represents a parameter of the entire network, α₁，β₁Representing the parameters of the output layer a1 and the hidden layer B1 of the DPN, respectively.

4. A medical image segmentation system, characterized in that the system comprises:

the establishing module is used for establishing a network model and training the network model to obtain a trained network model;

the prediction module is used for predicting the image-level labels and the pixel-level labels by using the trained model;

the segmentation module is used for realizing accurate segmentation of the image by utilizing the predicted image level label, the predicted pixel level label and the feature mapping of the reconstructed image;

the building block is particularly adapted to annotate data { I) with pixel level^f，L^f，T^fParameter adjustment is carried out on the basic network, the pixel label prediction network, the image label prediction network and the image reconstruction network through a minimum loss function; the Loss function comprises a Loss function Loss corresponding to pixel label prediction^map(Bo，L^f) Loss function Loss of image tag^tag(Co，T^f) Euclidian Loss function Loss for image reconstruction^img(Do，I^f) (ii) a Wherein Bo, Co and Do represent the predicted pixel-level label, image-level label and reconstructed image, respectively;

annotating data { I) with image level^w，T^wRealizing further training of the model by minimizing an image label loss function and an image reconstruction loss function, and realizing pixel-level label L by integrating image-level label information after further training is finished^wPredicting; after the pixel level label information of the weakly marked data is obtained, the predicted L is calculated^wAs a true value, weakly labeled image data is converted into fully labeled image data, and all of the image data is utilizedAnd carrying out supervision training on the whole network model by the image data to obtain a trained network model.

5. The system of claim 4,

the network model includes: an infrastructure network A and an infrastructure network B, the infrastructure network A comprising: a deep residual network and a dense convolutional neural network; the base network B includes: a capsule network structure model.

6. The system of claim 4, the supervised training comprising:

according to the multi-step training method of multi-task learning, the parameter change of the network is smooth, and the Loss function comprises a softmax Loss function Loss predicted by a pixel label^map(Ao，L^f) Softmax Loss function Loss of image tag^tag(Bo，T^f) And Euler Loss function Loss of image reconstruction^img(Co，I^f)；

In the semi-supervised learning phase, we fine-tune the model using all full-labeled and weak-labeled image data, all denoted as { I, L, T }, where I ═ I {^f，I^w}，L＝{L^f，L^wT ═ T^f，T^wThen, the optimization problem of the whole network is converted into:

where θ represents a parameter of the entire network, α₁，β₁Representing the parameters of the output layer a1 and the hidden layer B1 of the DPN, respectively.

7. A computer-readable storage medium storing a program for electronic data exchange, wherein the program causes a terminal to perform the method as provided in any one of claims 1-3.

Images