Medical image segmentation method based on cross pseudo supervision

Abstract

The invention provides a medical image segmentation method based on cross pseudo supervision. The medical image segmentation method comprises the following implementation steps: acquiring a training and testing sample set; constructing a medical image segmentation network model of cross pseudo supervision based on strong and weak disturbance consistency, and performing iterative training on the medical image segmentation network model; and obtaining a medical image segmentation result. According to the invention, strong disturbance is carried out on a sample after weak disturbance through the first branch network, then image-level strong disturbance segmentation is carried out, meanwhile, feature-level strong disturbance segmentation is carried out on the sample after weak disturbance through the second branch network, and cross pseudo-supervision losses of the image-level strong disturbance segmentation sub-network and the feature-level strong disturbance segmentation sub-network are respectively calculated through the cross pseudo-supervision module. The method achieves the consistency of strong and weak disturbances, avoids the defect of training deviation accumulation caused by the excessive dependence on weak disturbances in the prior art, can better learn the complex edge features in the image, and effectively improves the segmentation precision compared with the prior art.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and relates to a medical image segmentation method, and specifically to a medical image segmentation method based on cross pseudo-supervision, which can be applied to the field of medical imaging. Background Art

[0002] In recent years, computer-aided diagnosis (CAD) has made great progress in the fields of medical imaging and diagnostic radiology due to the development of deep learning technology. Medical image segmentation can identify the pixels of organs or lesions from the background of medical images. This is not only an important part of CAD, but also one of the most challenging tasks in medical image analysis. Accurate medical image segmentation can effectively determine information such as the size and location of organs and lesions, thereby quantitatively evaluating the effects before and after treatment, helping doctors to quickly make more accurate clinical decisions, and greatly reducing the workload of imaging doctors.

[0003] However, accurate dense annotation needs to be carried out under the guidance of professional doctors, which requires the annotators to have highly specialized knowledge and experience. Therefore, it is difficult to obtain large-scale segmentation maps with dense labels. Semi-supervised medical image segmentation methods can reduce the cost of medical image annotation by jointly using a small amount of labeled data and a large amount of unlabeled data. However, most semi-supervised medical image segmentation methods focus on consistency regularization, which is usually based on the clustering assumption of semi-supervised learning: samples of the same category will be clustered in data clusters separated by low-density areas. In semantic segmentation tasks, the distribution of pixels is significantly different from the distribution of predicted categories in image classification tasks, and there are no obvious low-density areas to distinguish the categories of pixels. Choosing more meaningful image enhancement methods is the key to obtaining reliable segmentation performance on such tasks without obvious low-density areas. In the field of semi-supervised image segmentation, data enhancement methods that can be applied to specific image segmentation tasks and can significantly improve the performance of semi-supervised image segmentation are called strong perturbations. In contrast, classic data enhancement methods such as image flipping, image scaling, and image rotation are called weak perturbations. Designing strong perturbations suitable for specific tasks introduces additional supervisory information, thereby reducing the confirmation bias generated by the model when there are only weak perturbations.

[0004] In order to introduce cross pseudo-supervision into the medical image segmentation model to improve the segmentation performance of the model, a patent application with application publication number CN115359029A and titled "A semi-supervised medical image segmentation method based on a heterogeneous cross pseudo-supervised network" discloses a semi-supervised medical image segmentation method based on a heterogeneous cross pseudo-supervised network. The invention combines Unet and Swin-Unet in the HCPS network model for cross pseudo-supervised learning to improve the training efficiency and segmentation effect of the network. At the same time, cross pseudo-supervised learning is performed based on the combination of Unet and Swin-Unet to improve the global feature capture ability and local capture ability, and make full use of the existing labeled data to achieve a satisfactory segmentation effect. However, the defect is that only weak perturbation methods are used for data enhancement, which depends to a large extent on the accurate prediction of the model for weakly perturbed images. When there are fewer labeled medical images, the confirmation bias of the model will continue to accumulate during training, resulting in poor segmentation effect of the edge parts of the medical image. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and propose a cross-pseudo-supervised medical image segmentation method based on the consistency of strong and weak perturbations to solve the technical problem of low segmentation accuracy in the prior art.

[0006] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Obtain training sample set and test sample set:

[0008] For Z including multiple pixel categories s A 3D medical image with labels is preprocessed, and the K s N 2D medical images and their labels in the 2D medical images are used as labeled training sample sets, more than half of the remaining 2D medical images are used as unlabeled training sample sets, and the remaining 2D medical images are used as test sample sets, where Z s ≥200,K s ≥2000,

[0009] (2) Construct a medical image segmentation network model based on cross pseudo-supervision:

[0010] Construct a medical image segmentation network model O including a weak perturbation module and a first network and a second network cascaded and arranged in parallel therewith, and a cross pseudo-supervision module connected to the output ends of the two networks; wherein the first network includes a cascaded strong perturbation pool, an image-level strong perturbation segmentation subnetwork and a perturbation recovery pool; the second network includes a cascaded feature-level strong perturbation segmentation subnetwork and a perturbation redo pool; the output ends of the image-level strong perturbation segmentation subnetwork and the feature-level strong perturbation segmentation subnetwork are also connected to the cross pseudo-supervision module;

[0011] (3) Iterative training of the medical image segmentation network model:

[0012] The labeled training sample set and the unlabeled training sample set are divided into batches and used as the input of the medical image segmentation network model O for iterative training to obtain the trained medical image segmentation network model O*;

[0013] (4) Obtain the segmentation results of medical images:

[0014] The test sample set is used as the input of the image-level strong perturbation segmentation subnetwork P* in the trained medical image segmentation network model O* for forward inference to obtain the segmentation result map of each test sample.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] In the process of training the medical image segmentation network model of the present invention, the first network performs strong perturbation on the weakly perturbed samples and then performs image-level strong perturbation segmentation, while the second network performs feature-level strong perturbation segmentation on the weakly perturbated samples. One of the sub-networks can serve as a pseudo-label for another sub-network for the segmentation of unlabeled data. The cross-pseudo-supervision loss of the image-level and feature-level strong perturbation segmentation sub-networks is calculated respectively through a cross-pseudo-supervision module, thereby achieving consistency between strong and weak perturbations, avoiding the problem of accumulated training deviations caused by inaccurate weak perturbation predictions, and better learning of complex edge features in images. Compared with the prior art, the segmentation accuracy is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart for realizing the present invention;

[0018] Figure 2 It is a schematic diagram of the overall structure of the medical image segmentation network model of the present invention;

[0019] Figure 3 It is a schematic diagram of the structure of the segmentation network of the present invention;

[0020] Figure 4 Schematic diagram of the structure of the convolution module of the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0022] Reference Figure 1 , the present invention comprises the following steps:

[0023] Step 1) Obtain training sample set and test sample set:

[0024] Step 1a) Obtain Z from the ICHS intracerebral hemorrhage dataset s The dimension is D s ×H s ×W s 3D cerebral hemorrhage medical images and their corresponding segmentation labels; along each medical image and its corresponding dimension is D s ×H s ×W s The dimension D of the 3D segmentation labels s Extract the vertical s Axis slice, get K s The dimension is H s ×W s The original image is sliced ​​and the skull is removed to keep only the central brain tissue area. s =Z s ×D s In this embodiment, Z s =270,16≤D s ≤24, H s =512,W s =512,K s =4982;

[0025] Step 1b) K s The dimension is H s ×W s The original image slices after skull removal and their corresponding labels are center cropped with a dimension of H×W, and the original image slices after center cropping are normalized to obtain the preprocessed K s 2D medical images of dimension H×W and their corresponding segmentation labels In this embodiment, H=256, W=256;

[0026] Step 1c) Preprocessing K s The N 2D medical images and their labels in the set constitute a labeled training sample set The remaining M preprocessed images form an unlabeled training sample set The remaining U preprocessed target domain images form a test sample set In this embodiment, N=1170, M=2338, and U=1474.

[0027] Step 2) Construct a medical image segmentation network model based on cross pseudo supervision:

[0028] Construct a medical image segmentation network model O including a weak perturbation module, a first network and a second network cascaded and arranged in parallel, and a cross pseudo-supervision module connected to the output ends of the two networks, and its structure is as follows: Figure 2 As shown; the first network includes a cascaded strong perturbation pool, an image-level strong perturbation segmentation subnetwork and a perturbation recovery pool; the second network includes a cascaded feature-level strong perturbation segmentation subnetwork and a perturbation redo pool; the output ends of the image-level strong perturbation segmentation subnetwork and the feature-level strong perturbation segmentation subnetwork are also connected to the cross pseudo-supervision module; the image-level strong perturbation segmentation subnetwork and the feature-level strong perturbation segmentation subnetwork both include a cascaded encoder Decoder And the output module Out, its structure is as follows Figure 3 As shown, the encoder It includes a sequentially connected convolution module Conv and a level I convolution Conv-encoding En module; a decoder The invention comprises a sequentially connected I-level convolution Conv-decoding De module; the input end of the i-th level convolution Conv-encoding En module is connected to the middle position of the I-i+1-th level convolution Conv-decoding De module, in this embodiment, I=4; the encoder in the feature-level strong perturbation segmentation subnetwork With decoder There is also a feature perturbation module loaded between

[0029] The structure of the convolution module Conv is as follows Figure 4As shown in the figure: the first convolution layer → the first batch of regularization layers → the first nonlinear function layer → Dropout layer → the second convolution layer → the second batch of regularization layers → the second nonlinear function layer. The first nonlinear function layer and the second nonlinear function layer in all Conv modules use the Leakyrelu nonlinear activation function. The loss probability of the Dropout layer is set to 0.5. The Dropout layer randomly inactivates the neurons in the network model with a probability of 0.5, that is, the output of the neuron is set to 0; the first convolution layer has the same structure as the second convolution layer; the convolution kernels in the first convolution layer and the second convolution layer in the convolution module Conv are set to 3×3, the number is 16, the step size is 1, and the surrounding padding is 1; the first-level encoding submodule En1 includes the maximum pooling layer and the first convolution module Conv1 connected in sequence, where the first convolution layer and the second convolution layer in Conv1 are The convolution kernel in the convolution layer is set to 3×3, the number is 32, the step size is 1, and the surrounding padding is 1; the second-level encoding submodule En2 includes a maximum pooling layer and a second convolution module Conv2 connected in sequence, wherein the convolution kernels in the first convolution layer and the second convolution layer in Conv2 are set to 3×3, the number is 64, the step size is 1, and the surrounding padding is 1; the third-level encoding submodule En3 includes a maximum pooling layer and a third convolution module Conv3 connected in sequence, wherein the convolution kernels in the first convolution layer and the second convolution layer in Conv3 are set to 3×3, the number is 128, the step size is 1, and the surrounding padding is 1; the fourth-level encoding submodule En4 includes a maximum pooling layer and a fourth convolution module Conv4 connected in sequence, wherein the convolution kernels in the first convolution layer and the second convolution layer in Conv4 are set to 3×3, the number is 256, the step size is 1, and the surrounding padding is 1;

[0030] Decoder It includes 4 levels of decoding submodules De connected in sequence j, wherein the first-level decoding submodule De1 includes a first deconvolution layer and a fifth-level convolution module Conv5 connected in sequence, wherein the convolution kernel of the first deconvolution layer is set to 2×2, the number is 128, the step size is 2, and the surrounding padding is 0, and the convolution kernels in the first convolution layer and the second convolution layer in Conv5 are set to 3×3, the number is 128, the step size is 1, and the surrounding padding is 1; the second-level decoding submodule De2 includes a second deconvolution layer and a sixth-level convolution module Conv6 connected in sequence, wherein the convolution kernel of the second deconvolution layer is set to 2×2, the number is 64, the step size is 2, and the surrounding padding is 0, and the convolution kernels in the first convolution layer and the second convolution layer in Conv6 are set to 3×3, the number is 64, the step size is 1, and the surrounding padding is 1. is 1; the third-level decoding submodule De3 includes a third deconvolution layer and a seventh-level convolution module Conv7 connected in sequence, wherein the convolution kernel of the third deconvolution layer is set to 2×2, the number is 32, the step size is 2, and the surrounding padding is 0, the convolution kernels in the first convolution layer and the second convolution layer in Conv7 are set to 3×3, the number is 32, the step size is 1, and the surrounding padding is 1; the fourth-level decoding submodule De4 includes a fourth deconvolution layer and an eighth-level convolution module Conv8 connected in sequence, wherein the convolution kernel of the fourth deconvolution layer is set to 2×2, the number is 16, the step size is 2, and the surrounding padding is 0, the convolution kernels in the first convolution layer and the second convolution layer in Conv8 are set to 3×3, the number is 16, the step size is 1, and the surrounding padding is 1.

[0031] The output layer Out includes an output convolution layer and a softmax layer connected sequentially, where the convolution kernel in the output convolution layer is set to 3×3, the number is 2, the stride is 1, and the surrounding padding is 1.

[0032] Step 3) Iteratively train the medical image segmentation network model:

[0033] Step 3a) Initialize the number of iterations to r, the maximum number of iterations to R, R>10000, and the weight parameters of the current image-level strong perturbation segmentation subnetwork and feature-level strong perturbation segmentation subnetwork are θ respectively. Pr and θ Tr , and let r = 1, in this embodiment, R = 30000;

[0034] Step 3b) The weak perturbation module performs a perturbation on each training sample in each batch of input with probability p r The rotation operation f(θ) is performed under the condition that the rotation angle θ is randomly selected; with probability p x Flip the image up and down f(h); with probability p y Flip the image left and right f(ω); with probability p n In this embodiment, the image is not weakly disturbed. r =p x=p y =p n =0.5;

[0035] Step 3c) Strong perturbation pool Aug in the first network s (x) For a batch of even numbers B after weak perturbation L labeled training samples and an even number B U Unlabeled training sample data, first randomly generate a rectangular mask M(h,w) with a length and width of h and w respectively, and exchange the mask parts corresponding to the b1th and b2th images in the labeled data and the unlabeled data respectively, and obtain each labeled image x after strong perturbation L S , the label y corresponding to the labeled image S and unlabeled images The way to generate the rectangular mask and the calculation process of the strong perturbation pool is:

[0036] M(h,w)= ((h,w)∈[λ1H:λ2H,λ3W:λ4W])

[0037]

[0038] in, is the indicator function; λ1<λ2,|λ1-λ2|>0.15, λ3<λ4,|λ3-λ4|>0.15. This hyperparameter is a random value, and its range can be set according to the specific data set; In this embodiment, B L =4, B U =4;

[0039] Encoder in the image-level strong perturbation segmentation subnetwork right Encode and decode separately right The encoded features are decoded respectively, and the output module Out Decoding obtains the decoding features and predicts them separately to obtain labeled images Prediction probability plot of Unlabeled images Prediction probability plot of Disturbance recovery pool UnAug s (p S ) The rectangular mask M(h,w) generated by the strong perturbation pool is The predicted probability map of the b1th and b2th images in The part corresponding to the mask part of is exchanged to obtain the cross pseudo label y' corresponding to the feature-level strong perturbation segmentation subnetwork Un ; The calculation process of the disturbance recovery pool is:

[0040]

[0041] Step 3d) Encoder in the feature-level strong perturbation segmentation subnetwork in the second network Encode the weakly perturbed labeled training samples and unlabeled training samples separately; Feature perturbation module Two-dimensional Dropout is used to strongly perturb the bottleneck features after passing through the encoder, and the bottleneck features after strong perturbation are connected to the original bottleneck features through channels, and they pass through the encoder together to obtain the overall prediction result. Then the feature dimension is divided equally to obtain the prediction result without strong feature disturbance And the prediction results after strong feature perturbation Output module Out Decode the features to make predictions and get the predicted probability map of the labeled image And the predicted probability map of the unlabeled image Disturb the redo pool ReAug s (y' W ) The rectangular mask M(h,w) generated by the strong perturbation pool is The parts corresponding to the mask parts of the b1th and b2th predicted probability maps in the image are exchanged to obtain the cross pseudo-label y' corresponding to the feature-level strong perturbation segmentation subnetwork Re ;

[0042] For the input weakly perturbed image, the overall calculation process of the feature-level strong perturbation segmentation model is:

[0043]

[0044] Among them, e W represents the original bottleneck feature after the encoder, Indicates the bottleneck feature after the feature is strongly perturbed. In this chapter, the feature strong perturbation module Feat(·) uses the two-dimensional Dropout in the verification mode. concate(·,·) means concatenating two features, and chunk(·) means dividing the feature into equal parts;

[0045] The specific calculation process of the disturbance redo pool is:

[0046]

[0047] Step 3e) The cross pseudo-supervision module uses consistency regularization and entropy minimization methods to obtain the unlabeled image after strong perturbation. Prediction probability plot of and cross pseudo-label y' Re Calculate the cross pseudo-supervision loss L of the image-level strongly perturbed segmentation subnetwork unsup1 , while passing the predicted probability map of the unlabeled image and cross pseudo-label y' Un Calculate the cross pseudo-supervision loss L of the feature-level strongly perturbed segmentation subnetwork unsup2 , the calculation process is:

[0048]

[0049]

[0050] Among them, λ cons and λ ent They represent the weighted coefficients for balancing the mean square error loss and the cross entropy loss, respectively. Onehot(·) represents the operation of converting features into one-hot encoding. L cons , L ent Respectively represent the mean square error loss function and the cross entropy loss function, B u represents the number of unlabeled images in each batch, p and p' respectively represent the feature map output by the encoder and the probability map obtained by the activation function for the data after different perturbations.

[0051] Step 3f) Segment all predicted probability maps p corresponding to the sub-network through image-level strong perturbation S , all predicted probability maps p corresponding to the feature-level strong perturbation segmentation sub-network W , the true labels y of all labeled training samples in each batch S ,by

[0052] and cross pseudo-label y' Un and y' Re ; Calculate the current loss value d of O r :

[0053] d r =L sup +λL unsup

[0054] Where L sup , L unsup Respectively represent supervised loss function and unsupervised loss function, λ represents L sup With L unsup Balance coefficient; supervised loss function, unsupervised loss function L sup , L unsup , whose expressions are:

[0055] L sup =λ sup1 L sup1 +λ sup2 L sup2

[0056]

[0057] L unsup =λ un (L unsup1 +L unsup2 )+λ fp L fp

[0058]

[0059] Among them, B L Indicates the number of labeled images in each batch, B u represents the number of unlabeled images in each batch, p and p' respectively represent the feature map output by the encoder and the probability map obtained by the activation function for the data after different perturbations;

[0060] Step 3g) Calculate the loss value d respectively r For θ Pr and θ Tr The partial derivative g Pr and g Tr ; Using the stochastic gradient descent method, through the partial derivative g Pr and g Tr For θ Pr and θ Tr Update, the update formula is:

[0061]

[0062] in, Represent weight parameters and The update result of α r represents the learning rate. In this embodiment, α1=0.01, α r =α r-1 (1.0-r / R) 0.9 ; r>1, and the semi-supervised medical image segmentation model O of this iteration is obtained r ;

[0063] Step 3h) Determine whether r=1 is true. If so, obtain the trained medical image segmentation network model O*. Otherwise, set r=r+1, O r =O, and execute step (3b).

[0064] Step 4) Obtain the segmentation results of the medical image:

[0065] The test sample set is used as the input of the image-level strong perturbation segmentation subnetwork P* in the trained medical image segmentation network model O* for forward inference to obtain the segmentation result map of each test sample.

Claims

1. A medical image segmentation method based on cross pseudo supervision, characterized in that: The steps include: (1) Obtain training sample set and test sample set: For Z including multiple pixel categories s A 3D medical image with labels is preprocessed, and the K s N 2D medical images and their labels in the 2D medical images are used as labeled training sample sets, more than half of the remaining 2D medical images are used as unlabeled training sample sets, and the remaining 2D medical images are used as test sample sets, where Z s ≥200,K s ≥2000, (2) Construct a medical image segmentation network model based on cross pseudo-supervision: Construct a medical image segmentation network model O including a weak perturbation module and a first network and a second network cascaded and arranged in parallel therewith, and a cross pseudo-supervision module connected to the output ends of the two networks; wherein the first network includes a cascaded strong perturbation pool, an image-level strong perturbation segmentation subnetwork and a perturbation recovery pool; the second network includes a cascaded feature-level strong perturbation segmentation subnetwork and a perturbation redo pool; the output ends of the image-level strong perturbation segmentation subnetwork and the feature-level strong perturbation segmentation subnetwork are also connected to the cross pseudo-supervision module; (3) Iterative training of the medical image segmentation network model: The labeled training sample set and the unlabeled training sample set are divided into batches and used as the input of the medical image segmentation network model O for iterative training to obtain the trained medical image segmentation network model O*; (4) Obtain the segmentation results of medical images: The test sample set is used as the input of the image-level strong perturbation segmentation subnetwork P* in the trained medical image segmentation network model O* for forward inference to obtain the segmentation result map of each test sample.

2. The method according to claim 1, characterized in that The pair Z of the pixels in step (1) includes multiple pixel categories. s The preprocessing steps of the labeled 3D medical image are as follows: Each labeled 3D medical image is sliced ​​along the depth direction, and each 2D medical image slice after slicing is center-cropped, and then the center-cropped image slices are normalized to obtain the preprocessed K s 2D medical image.

3. The method according to claim 1, characterized in that The medical image segmentation network model O described in step (2), wherein: Both the image-level strong perturbation segmentation subnetwork and the feature-level strong perturbation segmentation subnetwork include cascaded encoders Decoder And the output module Out, the encoder in the feature-level strong perturbation segmentation subnetwork With decoder There is also a feature perturbation module loaded between The encoder It includes a cascaded convolution module Conv and a level I convolution Conv-encoding En module; the convolution module Conv includes multiple stacked convolution layers, normalization layers and activation function layers; the Conv-encoding En module includes multiple stacked maximum pooling layers and multiple convolution layers; the decoder It includes a cascaded I-level convolution Conv-decoding De module; the output end of the i-th level convolution Conv-encoding En module is connected to the input end of the I-i+1-th level convolution Conv-decoding De module.

4. The method according to claim 1, characterized in that: The iterative training of the medical image segmentation network model O described in step (3) is implemented as follows: (3a) The number of initial iterations is r, the maximum number of iterations is R, R>10000, and the weight parameters of the current image-level strong perturbation segmentation subnetwork and feature-level strong perturbation segmentation subnetwork are θ respectively. Pr and θ Tr , and let r = 1; (3b) The weak perturbation module rotates, flips upside down, or flips left to right each labeled training sample and unlabeled training sample in each batch according to a preset probability to obtain the weakly perturbed labeled training samples and unlabeled training samples; (3c) The strong perturbation pool in the first network performs strong perturbations on the labeled training samples and unlabeled training samples after weak perturbations respectively; the image-level strong perturbation segmentation subnetwork performs strong perturbation on each labeled image obtained and each unlabeled image Perform feature learning and prediction to obtain Prediction probability plot of and Prediction probability plot of Disturbance recovery pool vs. predicted probability map Perform perturbation recovery to obtain the cross pseudo-label y' corresponding to the feature-level strong perturbation segmentation sub-network Un ; (3d) The feature-level strong perturbation segmentation subnetwork in the second network performs feature learning and prediction on the weakly perturbed labeled training samples and unlabeled training samples, respectively, and obtains the predicted probability map of the labeled image And the predicted probability map of the unlabeled image Disturb the redo pool Perform perturbation and redo to obtain the cross pseudo-label y' corresponding to the image-level strong perturbation segmentation subnetwork Re ; (3e) The cross-pseudo-supervision module uses consistency regularization and entropy minimization methods to obtain the unlabeled image after strong perturbation. Prediction probability plot of and cross pseudo-label y' Re Calculate the cross pseudo-supervision loss L of the image-level strongly perturbed segmentation subnetwork unsup1 , while passing the predicted probability map of the unlabeled image and cross pseudo-label y' Un Calculate the cross pseudo-supervision loss L of the feature-level strongly perturbed segmentation subnetwork unsup2 ; (3f) All predicted probability maps p corresponding to the image-level strong perturbation segmentation sub-network S , all predicted probability maps p corresponding to the feature-level strong perturbation segmentation sub-network W , the true and cross-pseudo labels y' for all labeled training samples in each batch Un and y' Re ; Calculate the current loss value d of O r , and use the stochastic gradient descent method, through d r The weight parameters θ are Pr and θ Tr Update to obtain the semi-supervised medical image segmentation model O of this iteration r ; (3g) Determine whether r = 1. If so, obtain the trained medical image segmentation network model O*. Otherwise, let r = r + 1, O r =O, and execute step (3b).

5. The method according to claim 4, characterized in that The step (3c) Prediction probability plot of and Prediction probability plot of The steps to achieve the acquisition process are: (3c1) The strong perturbation pool randomly generates a rectangular mask M(h,w) with a length and width of h and w respectively for a batch of data containing an even number of labeled training samples and an even number of unlabeled training samples after weak perturbation, and exchanges the mask parts corresponding to the b1th and b2th images in the labeled data and the unlabeled data respectively, to obtain each labeled image after strong perturbation. and each unlabeled image (3c2) Encoder in the image-level strong perturbation segmentation subnetwork right Encode and decode separately right The encoded features are decoded respectively, and the output module Out Decoding obtains the decoding features and predicts them separately to obtain labeled images Prediction probability plot of Unlabeled images Prediction probability plot of 6. The method according to claim 4, characterized in that The feature-level strong perturbation segmentation subnetwork described in step (3d) performs feature learning and prediction on the weakly perturbed labeled training samples and unlabeled training samples respectively, and the implementation steps are as follows: Encoder Encode the weakly perturbed labeled training samples and unlabeled training samples separately; Feature perturbation module Two-dimensional Dropout is used to strongly perturb the bottleneck features obtained by encoding; the decoder The strongly disturbed features are concatenated with the unperturbed features and then decoded; the output module Out Decode the features to make predictions and get the predicted probability map of the labeled image And the predicted probability map of the unlabeled image 7. The method according to claim 4, characterized in that The cross pseudo-supervision loss L described in step (3e) unsup1 , and the cross pseudo-supervision loss L unsup2 , the calculation formulas are: Among them, λ cons and λ ent They represent the weighted coefficients for balancing the mean square error loss and the cross entropy loss, respectively. Onehot(·) represents the operation of converting features into one-hot encoding. L cons , L ent Respectively represent the mean square error loss function and the cross entropy loss function, B u represents the number of unlabeled images in each batch, p and p' respectively represent the feature map output by the encoder and the probability map obtained by the activation function for the data after different perturbations.

8. The method according to claim 4, characterized in that The current loss value d of O described in step (3f) r , the calculation formula is: d r =L sup +λL unsup Where L sup , L unsup Respectively represent supervised loss function and unsupervised loss function, λ represents L sup With L unsup The balance coefficient.

9. The method according to claim 4, characterized in that The step (3b) described in θ Pr and θ Tr To update, the update formulas are: in, Represent the weight parameters θ Pr and θ Tr The update result, α r represents the learning rate, g Pr , g Tr Respectively represent d r For θ Pr and θ Tr The partial derivative of .

Images