Medical image automatic segmentation method based on semi-supervised learning

Abstract

The invention relates to a medical image automatic segmentation method based on semi-supervised learning, and belongs to the technical field of medical image accurate segmentation. The method comprises the following steps: S1, proposing an automatic segmentation task of a medical image, wherein the method comprises the following small steps: S11, establishing a DulecrASE-Net semi-supervised network structure; s12, an MSA-RFSEU-Net network structure is established; s13, designing an adversarial two-stage consistency training strategy; s14, designing a multi-scale attention component DyMAC based on dynamic convolution; s15, designing a combined loss function; s2, establishing a medical image data set to carry out model evaluation, wherein the model evaluation comprises the following steps: S21, selection of evaluation indexes; s22, verifying the effectiveness of the evaluation model; and S23, verifying the universality of the evaluation model. According to the method, the medical image focus area can be effectively segmented, the segmentation precision can be improved to a certain extent through the designed antagonistic two-stage consistency training strategy, and the method can be used for the precise segmentation occasion of the medical image.

Description

Technical Field

[0001] The invention relates to a medical image automatic segmentation method based on semi-supervised learning, and belongs to the technical field of accurate segmentation of medical images. Background Art

[0002] With the rapid development of medical technology, medical imaging has become one of the important reference information for doctors to understand and analyze the condition. Medical image segmentation is to segment the areas with specific meanings in medical images and extract relevant features, providing a reliable basis for clinical diagnosis and pathological research, assisting doctors to accurately analyze the pathological location and boundaries of the patient's lesion tissue, and improving the accuracy of medical diagnosis.

[0003] In recent years, deep learning technology has made rapid progress. Since Minaee et al. proposed the fully convolutional network (FCN), various models based on deep learning have been widely used in the field of medical image segmentation and have achieved remarkable success. Ronnebrge et al. proposed an end-to-end U-shaped network structure (U-Net), which uses jump connections to combine shallow and deep spatial information, and improves the segmentation accuracy to a certain extent. Gu et al. proposed a context encoder network (CE-Net) to capture more high-level semantic information and preserve spatial information, and greatly reduce the number of network parameters. Xia et al. proposed a multi-scale dilated convolution network (MD-Net) that uses dilated convolutions of different sizes to form a feature pyramid to extract semantic information in the image, and uses bilinear interpolation and repeated convolution to reduce the number of network parameters, achieving better results. Later, Ding et al. used a multi-scale strategy and attention mechanism at the same time to suppress irrelevant parts or backgrounds to highlight the spatial information of deep features, and further extract detailed information of image features.

[0004] However, deep neural networks for medical image segmentation often require a large number of original images and corresponding labeled images. Label images need to be manually annotated by clinicians, which is costly, time-consuming, and difficult to obtain. In addition, due to different clinical experiences of doctors, the segmentation results are affected by subjective factors of doctors, which reduces the reliability of segmentation. At the same time, class imbalance problems often occur in medical images because the lesion area is relatively small and significantly different from the surrounding healthy tissue. In addition, the lesion area will also have intra-class variation in morphology and size, making accurate segmentation of the lesion area challenging.

[0005] In order to reduce the burden of manual labeling, the semi-supervised learning image segmentation algorithm uses a small number of labeled images to mine feature information in a large number of unlabeled images, enhance model performance, and improve segmentation accuracy, which is of great help to actual clinical applications. In order to make better use of unlabeled data, the invention mainly adopts consistency learning and adversarial learning to improve network performance. Consistency learning refers to adding small perturbations to the input data to keep the prediction results of the model output consistent. MeanTeacher (MT) is a typical representative of the consistency regularization method, which consists of a student model and a teacher model with the same structure. The student model updates parameters by optimizing the joint loss composed of supervised learning loss and consistency regularization loss, while the parameters of the teacher model are the exponential moving average (EMA) of the student model parameters. Adversarial learning uses a generative adversarial network (GAN) in medical image segmentation, and its structure includes a generator network and a discriminator network. The generator network receives a noise input that obeys a certain distribution and generates output samples similar to real data. The goal of the generator network is to make its output samples close to the real data. Subsequently, the fake data samples output by the generator network are mixed with the real data samples and input into the discriminator network, which determines the source of these input samples. The parameters of the two networks are updated alternately and promote each other. Summary of the invention

[0006] In view of the above-mentioned defects in the prior art, the present invention proposes a medical image automatic segmentation method based on semi-supervised learning.

[0007] The method for automatic segmentation of medical images based on semi-supervised learning of the present invention comprises the following steps:

[0008] S1: Propose the automatic segmentation task of medical images, including the following steps:

[0009] S11: Establish the DulecrASE-Net network structure: Using semi-supervised learning MT as the basic framework, the MSA-RFSEU-Net segmentation network is used to refine the segmentation effect, and an adversarial two-level consistency training strategy and a dynamic convolutional multi-scale attention component (DyMAC) are introduced in the discriminator to solve the problem of medical image category imbalance, enhance the model feature learning ability, and reduce the overfitting problem;

[0010] S12: Establish the MSA-RFSEU-Net network structure: The U-Net network is used as the basic framework, the encoder adopts ResNet34, the decoder block includes two RFSEConv (Receptive-Field Squeeze-and-ExcitationConvolution) and an upsampling operation, and a multi-scale side attention block MSA-Block (Multi-Scale AttentionBlock) is added after the first decoder block, which is conducive to enhancing the network's multi-scale feature extraction capability, making it easier to obtain high-level features of the lesion area of ​​the medical image, solving the problem of low contrast and blurred edges of medical images, and further improving the segmentation performance;

[0011] S13: Designing an adversarial two-level consistency training strategy: Introducing a discriminator network in semi-supervised learning MT to transform image-level prediction consistency L ILPC Loss, Feature-level Attention Consistency L FLAC Loss and Discriminator L D The loss is combined to solve the problem of imbalanced categories in medical images, thereby improving the segmentation accuracy;

[0012] S14: Design of DyMAC based on dynamic convolution multi-scale attention component: Introducing MSA-Block in the discriminator can better enhance the multi-scale information representation of feature maps, making the network more robust when dealing with objects of different scales; Introducing dynamic convolution to better utilize the prior relationship between labeled data and unlabeled data, enhance the model feature learning ability, and reduce the overfitting problem;

[0013] S15: Design a combined loss function: In image segmentation problems, Dice loss is usually used as the loss function, but Dice loss has a poor effect on the segmentation task with imbalanced categories. To address this problem, the present invention introduces a binary cross entropy (BCE) loss function, and weights the Dice and BCE losses to form a joint loss, so that the student network pays more attention to the details of the feature map on the unlabeled data and generates high-quality segmentation results. By introducing image-level prediction consistency L ILPC Loss, Feature-level Attention Consistency L FLAC Loss and Discriminator L D The loss is reasonably designed to improve the image segmentation accuracy. The combined loss function redesigned by the present invention is defined as shown in formula (4);

[0014]

[0015] L D (θ) = L ad (D(x i ,f i),1)+L ad (D(M a (x p ,x q ),f(M a )),0) (3)

[0016] L total =L D +L ILPC +L FLAC (4)

[0017] Where: f(M a (x p ,x q ) represents the prediction result of the mixed image student model;

[0018] Indicates the true label of the mixed image;

[0019] M a (a' p ,a' q ) represents the mixed attention map obtained by the PolypMix operation;

[0020] a(M a (x p ,x q )) represents the attention prediction result of the mixed image student model;

[0021] M a (a' p ,a' q ) represents the real attention map of the mixed image;

[0022] x i , f i Represents label images and prediction results;

[0023] M a (x p ,x q ), f(M a ) represents unlabeled images and prediction results;

[0024] S2: Establish a medical image dataset for model evaluation, including the following steps:

[0025] S21: Selection of evaluation indicators: Seven indicators are selected: Dice coefficient, accuracy, recall rate, mean intersection ratio, Jaccard similarity coefficient, precision and average consistent distance, among which:

[0026] Dice coefficient: used to measure the similarity between the segmentation result and the actual tumor area;

[0027]

[0028] Accuracy (ACC): The ratio of the number of correctly predicted pixels in the predicted segmentation results to the total number of pixels;

[0029]

[0030] Recall (REC): The percentage of correctly predicted lesion areas in the true label;

[0031]

[0032] Mean Intersection over Union (mIoU): Calculates the average of the ratio of the intersection and union of all categories;

[0033]

[0034] Jaccard Similarity Coefficient (Jaccard): Compares the difference and similarity between the tumor prediction results and the true labels;

[0035]

[0036] Precision (Prec): measures the maximum difference between the segmentation result and the actual tumor area;

[0037]

[0038] Mean distance to agreement (MDA): measures the average distance between each pixel in the segmentation result and the true tumor area;

[0039] MDA(L,P)=mean l∈L,p∈P (||lp||) (11)

[0040] Where: TP represents the number of pixels correctly predicted as tumors;

[0041] TN represents the number of pixels correctly predicted as non-tumor;

[0042] FP represents the number of pixels incorrectly predicted as tumors;

[0043] FN represents the number of pixels incorrectly predicted as non-tumor;

[0044] PL and PP represent the real tumor area and predicted tumor area, respectively;

[0045] L and P represent the tumor region sets of the true labels and the predicted results, respectively;

[0046] S22: Verification of model effectiveness: DulecrASE-Net was compared with seven classic semi-supervised medical segmentation methods, including MT, UAMT, ICT, ADVENT, ClassMix, URPC, and PolypMixNet, and the fully supervised learning model U-Net using the Kvasir-SEG public colon polyp dataset. The semi-supervised learning model was trained using 15% of the labeled data. The comparison methods all used the same MSA-RFSEU-Net backbone and were trained and tested under the same experimental environment parameter settings. The DulecrASE-Nett model achieved the best segmentation effect, which was significantly higher than other models in terms of ACC, REC, Prec, and mIoU. The Dice coefficient also achieved good results.

[0047] S23: Verification of the universality of the evaluation model: To further verify the performance of the network, the DulecrASE-Net model was compared with several different advanced semi-supervised methods including MT, UCMT, ASE-Net and PolypMix, as well as a fully supervised method using a self-built small intestinal tumor CT image dataset. The semi-supervised learning model was trained using 15% of the labeled data, and the comparison methods all used the same U-Net backbone.

[0048] Preferably, in step S11, the DulecrASE-Net network consists of a segmentation network and a discrimination network, the segmentation network consists of a student model and a teacher model, both the student model and the teacher model use the designed MSA-RFSEU-Net segmentation network, and the discrimination network consists of a convolutional layer, DyMAC and global average pooling; the segmentation network and the discriminator are alternately trained to achieve adversarial two-level consistency learning, which can effectively utilize multi-level information from unlabeled data to solve the problem of imbalanced medical image categories, thereby improving segmentation accuracy; in terms of network structure, we use DyMAC to replace all convolutional layers except the first layer in the discriminator network to improve the feature representation ability of the network and reduce the risk of overfitting.

[0049] Preferably, in step S12, the MSA-RFSEU-Net network consists of an encoder and a decoder, and the encoder and decoder branches each consist of 5 blocks; the encoder adopts ResNet34, and the decoder block first introduces the channel attention squeeze-excitation (Squeeze and Excitation, SE) design RFSEConv to replace the ordinary convolution on the basis of the RFA (Receptive-Field Attention) module to process the global information, and adaptively reweight the features of each channel, thereby enhancing the network's ability to represent important features and improving the overall performance and adaptability of the model; then an MSA-Block is added after the first decoder block, and the MSA-Block introduces a spatial pyramid pooling structure and a batch normalization layer on the basis of the DA-Block, which effectively integrates information of multiple scales from global to local, enhances the network's ability to capture multi-scale spatial information, and maintains the integrity of fine-grained information while reducing the amount of calculation.

[0050] Preferably, in step S22, in order to compare the differences between the model and other semi-supervised models and fully supervised models, the DulecrASE-Net network is compared and verified on the Kvasir-SEG public colon polyp dataset, and the adversarial two-level consistency training strategy and the multi-scale attention component of dynamic convolution can be used to effectively utilize multi-level information from unlabeled data, solve the problem of imbalanced medical image categories, and thus improve segmentation accuracy.

[0051] Preferably, in step S23, in order to verify the universality of the DulecrASE-Net network, a comparative verification is performed on a self-built small intestine tumor CT data set, and it can be found that the DulecrASE-Net model achieves the best segmentation effect.

[0052] The beneficial effects of the present invention are:

[0053] (1) This paper proposes an adversarial two-level consistent semi-supervised segmentation network DulecrASE-Net for the automatic segmentation task of semi-supervised medical images;

[0054] (2) The combined loss function is used in the model training process of the present invention, and the segmentation effect of the combined loss function is tested. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is the overall structure diagram of the DulecrASE-Net network of the present invention;

[0056] Figure 2 It is a structural diagram of the DyMAC model of the present invention;

[0057] Figure 3 It is a structural diagram of the RFSEConv model of the present invention;

[0058] Figure 4 is a structural diagram of the MSA-Block module of the present invention;

[0059] Figure 5 It is a visualization comparison diagram of the segmentation results of the DulecrASE-Net model of the present invention on the Kvasir-SEG dataset;

[0060] Figure 6 This is a visualization comparison diagram of the segmentation results of the DulecrASE-Net model of the present invention in the small intestine tumor CT image dataset; DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] Embodiment 1:

[0063] Figures 1 to 6 The figure shows the semi-supervised automatic medical image segmentation method based on adversarial two-level consistency regularization described in the present invention. The present invention aims to establish a deep learning model for automatic segmentation of medical images with a small amount of labeled data and a large amount of unlabeled data. Since MT aims to use less labeled data while improving the accuracy of prediction, the present invention uses MT as the basic framework, first refines the segmentation effect through the MSA-RFSEU-Net segmentation network; then introduces the adversarial two-level consistency training strategy and DyMAC to solve the problem of imbalanced medical image categories, enhance the model feature learning ability, and reduce the overfitting problem.

[0064] DulecrASE-Net overall network structure of the present invention Figure 1As shown in the figure: the network consists of a segmentation network and a discrimination network. The segmentation network consists of a student model and a teacher model. Both the student model and the teacher model use the designed MSA-RFSEU-Net segmentation network. The discrimination network consists of a convolutional layer, DyMAC and global average pooling. The segmentation network and the discriminator are trained alternately to achieve adversarial two-level consistency learning, which can effectively utilize multi-level information from unlabeled data to solve the problem of imbalanced medical image categories, thereby improving segmentation accuracy. In terms of network structure, we use DyMAC to replace all convolutional layers except the first layer in the discriminator network to improve the feature representation ability of the network and reduce the risk of overfitting.

[0065] The MSA-RFSEU-Net network consists of two parts: an encoder and a decoder. The encoder and decoder branches each consist of 5 blocks. The encoder uses ResNet34. The decoder block first processes global information by replacing ordinary convolution with SE-designed RFSEConv convolution based on the RFA module, and adaptively reweights the features of each channel, thereby enhancing the network's ability to represent important features and improving the overall performance and adaptability of the model. Then, MSA-Block is added after the first decoder block. Based on DA-Block, MSA-Block introduces a spatial pyramid pooling structure and a batch normalization layer, which effectively integrates information of multiple scales from global to local, enhances the network's ability to capture multi-scale spatial information, and maintains the integrity of fine-grained information while reducing the amount of computation.

[0066] It should be noted that in the DulecrASE-Net network, the input of the discriminator network is the concatenation of the segmentation result after softmax and the original image. In this way, the quality of the segmentation result can be further evaluated based on the original image, and the matching relationship between the segmentation result and the reference image can be determined. In addition, the image-level prediction consistency loss and feature-level attention consistency loss are for unlabeled data. When the network input is unlabeled data, the teacher model first uses the PolypMix strategy to perform multi-scale attention on two unlabeled images. p and a q To guide the adaptive generation of the mixed image M a (x p ,x q ) and mixed labels Secondly, PolypMix is ​​applied at the feature level to generate a mixed attention image M a (a' p ,a' q ), as the label for mixed image attention prediction.

[0067] Figure 2The multi-scale attention component DyMAC based on dynamic convolution described in the present invention makes full use of prior knowledge by adaptively adjusting the parameters of each sample, reduces the risk of overfitting and improves the feature representation ability; for a given input x in ∈R C×H×W First, MSA-Block is applied to enhance the multi-scale information representation of the feature map, which solves the problem of low contrast and blurred edges in medical images, making the network more robust when dealing with objects of different scales; then the spatial attention module is used to reduce the dimension through 1×1 convolution, and then normalized through the sigmoid activation function, and the obtained weights are multiplied element by element with the input feature map to obtain the feature map x1; ​​finally, x1 is converted to x2∈R through global average pooling C×1×1 Then, 1×1 convolution is used for dimensionality reduction, and softmax activation is used to obtain the dynamic convolution kernel p∈R N×1×1 The weight of N represents the number of convolution kernels, the coefficient p is multiplied by N convolution kernels, and they are weighted to generate a dynamic convolution kernel.

[0068] Embodiment 2:

[0069] The effectiveness of the present invention is evaluated below with reference to specific cases.

[0070] This paper tests the effectiveness of our proposed method on the Kvasir-SEG public colon polyp segmentation dataset and the small intestine tumor CT image dataset.

[0071] The Kvasir-SEG dataset contains 1,000 polyp images with resolutions ranging from 332×487 to 1920×1070 pixels. Since the Kvasir-SEG dataset is not divided into training, validation, and test sets, in order to ensure the reliability of the experiment, 60% of the Kvasir-SEG images are randomly divided for training, 20% for validation, and 20% for testing. The small intestinal tumor CT image dataset is derived from CT images of 79 patients diagnosed with small intestinal tumors in the Affiliated Hospital of Qingdao University, including one arterial phase image and one venous phase image, for a total of 158 images. The resolution of all images is uniformly adjusted to 256×256. These images are desensitized and the small intestinal tumor lesion areas are annotated by three experienced doctors. In order to improve the robustness of the model and reduce the risk of overfitting, this paper randomly selects 20 original CT image data of different patients as test set images, and expands the remaining images to 822 images as training sets by rotation, flipping, etc.

[0072] In order to verify the effectiveness of the network proposed in the present invention, the present invention first uses 15% labeled data on the Kvasir-SEG public colon polyp segmentation dataset to compare and verify the DulecrASE-Net network with MT, UAMT, ICT, ADVENT, URPC and PolypMixNet semi-supervised medical segmentation methods. In order to prove the effectiveness of semi-supervision, the model of the present invention also compares the performance difference with the classic fully supervised segmentation network U-Net.

[0073] Table 1 Test results of different models on the Kvasir-SEG dataset

[0074]

[0075] Table 1 shows the comparison between the DulecrASE-Net network structure proposed in the present invention and other network segmentation indicators. It can be seen from the table that compared with the fully supervised U-Net method, the performance of DulecrASE-Net has been significantly improved, among which the Dice index has been improved by 6.15% and the mIoU has been improved by 7.58%. In addition, other evaluation indicators have also been improved. This shows that DulecrASE-Net can make full use of unlabeled data and effectively improve segmentation accuracy. Compared with other advanced semi-supervised learning methods, DulecrASE-Net has achieved the best in all five evaluation indicators, specifically Dice is 90.05% and mIoU is 92.45%. Compared with the suboptimal model PolypMixNet, DulecrASE-Net has obvious improvements in all evaluation indicators. This fully demonstrates the effectiveness of the DulecrASE-Net network structure proposed in the present invention in the task of colon polyp segmentation image segmentation.

[0076] Figure 5The test results of the eight models in Table 1 under different samples are shown. The eight columns correspond to the sample test results of U-Net, MT, UAMT, ICT, ADVENT, URPC, PolypMixNet and the DulecrASE-Net network structure proposed in the present invention. As can be seen from No. 1, the traditional fully supervised method is easily affected by background noise during specific segmentation, resulting in blurred boundaries in the segmentation results. For sample No. 2 where the lesion area is not obvious, most methods find it difficult to accurately identify polyps, resulting in inaccurate predictions. For samples No. 3 and No. 4 with blurred boundaries, the network of the present invention can provide more accurate boundary delineation, divide relatively complete polyp areas, and achieve superior segmentation predictions. As can be seen from the above images, DulecrASE-Net can accurately segment small-sized polyps and polyps with unclear boundaries, and for polyps with irregular shapes, most models have problems of over-segmentation and under-segmentation. Compared with other methods, the model proposed in the present invention has a complete segmentation area, clear edges, and better segmentation accuracy and stability.

[0077] In addition, in order to continue to explore the universality of the model, this chapter further designed a comparative experimental scheme. Using the small intestinal tumor CT image dataset provided by the Affiliated Hospital of Qingdao University, the DulecrASE-Net model was compared and verified with four different advanced semi-supervised models: MT, UCMT, ASE-Net, and PolypMix. All semi-supervised methods used 15% labeled data. In addition, DulecrASE-Net was also compared and verified with the fully supervised U-Net.

[0078] Table 2 Test results of different models on the small intestine tumor CT image dataset

[0079]

[0080] Table 2 shows the comparison of the DulecrASE-Net network structure proposed in the present invention with other network segmentation indicators on the small intestinal tumor CT dataset. It can be seen from the table that the DulecrASE-Net proposed in the present invention has a significant improvement in the Dice indicator, even far exceeding the Dice indicator of fully supervised learning. Compared with the PolypMix semi-supervised method, the model of the present invention has improved by 3.79% in Dice and 10.92% in Jaccard, and other indicators have reached the best in the comparison model. This shows that the DulecrASE-Net model has more advantages in the segmentation task of small intestinal tumor CT images.

[0081] Figure 6The test results of the six models in Table 2 under different samples are shown. The six columns correspond to the sample test results of U-Net, MT, UCMT, ASE-Net, PolypMixNet and the DulecrASE-Net network structure proposed in this invention. From the comparison chart, it can be seen that the segmentation results of the proposed method are more similar to the real label edge, and the segmentation results retain more fine details.

[0082] Aiming at the problem of lack of sufficient manually annotated data sets due to the difficulty of annotating medical images, the present invention proposes an adversarial two-level consistency semi-supervised medical image segmentation network DulecrASE-Net, which realizes the automatic segmentation task of medical images when there is only a small amount of labeled data. The model solves the problem of imbalanced medical image categories through an adversarial two-level consistency training strategy and a dynamic convolution-based multi-scale attention component DyMAC, enhances the model feature learning ability, and reduces the overfitting problem; and uses the MSA-RFSEU-Net segmentation network to refine the segmentation effect; uses a combined loss function to further improve the segmentation accuracy, and conducts comparative experiments on the Kvasir-SEG public colon polyp segmentation dataset and the small intestine tumor CT image dataset. The experiments all use 15% labeled data and achieve good results, verifying the effectiveness and versatility of the method of the present invention, and showing the superiority and effectiveness of the segmentation performance of the network model proposed by the present invention. It can use a large amount of unlabeled data to achieve accurate segmentation of medical images under the condition of limited labeled data.

[0083] This paper proposes a new method for automatic medical image segmentation by combining a small amount of labeled data and a large amount of unlabeled data. The main contents of this paper can be summarized as follows: 1) A network model DulecrASE-Net that integrates an adversarial two-level consistency training strategy and a multi-scale attention component based on dynamic convolution is proposed for the task of automatic medical image segmentation; 2) A small dataset of small intestinal tumor CT image segmentation is constructed, and the versatility of the method proposed in this paper is verified on this dataset.

[0084] The present invention can be widely used in the precise segmentation of medical images.

[0085] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0086] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic segmentation of medical images based on semi-supervised learning, characterized in that: The steps include: S1: Propose the automatic segmentation task of medical images, including the following steps: S11: Establish the DulecrASE-Net network structure: Using the semi-supervised learning Mean Teacher as the basic framework, the MSA-RFSEU-Net segmentation network is used to refine the segmentation effect, and the adversarial two-level consistency training strategy and the dynamic convolutional multi-scale attention component (DyMAC) are introduced in the discriminator to solve the problem of medical image category imbalance, enhance the model feature learning ability, and reduce the overfitting problem; S12: Establish the MSA-RFSEU-Net network structure: The U-Net network is used as the basic framework, the encoder adopts ResNet34, the decoder block includes two RFSEConv (Receptive-Field Squeeze-and-Excitation Convolution) and an upsampling operation, and a multi-scale side attention block MSA-Block (Multi-Scale AttentionBlock) is added after the first decoder block, which is conducive to enhancing the network's multi-scale feature extraction capability, making it easier to obtain high-level features of the lesion area of ​​the medical image, solving the problem of low contrast and blurred edges of medical images, and further improving the segmentation performance; S13: Designing an adversarial two-level consistency training strategy: Introducing a discriminator network in semi-supervised learning MT to transform image-level prediction consistency L ILPC Loss, Feature-level Attention Consistency L FLAC Loss and Discriminator L D The loss is combined to solve the problem of imbalanced categories in medical images, thereby improving the segmentation accuracy; S14: Design of DyMAC based on dynamic convolution multi-scale attention component: Introducing MSA-Block in the discriminator can better enhance the multi-scale information representation of feature maps, making the network more robust when dealing with objects of different scales; Introducing dynamic convolution to better utilize the prior relationship between labeled data and unlabeled data, enhance the model feature learning ability, and reduce the overfitting problem; S15: Design a combined loss function: In image segmentation problems, Dice loss is usually used as the loss function, but Dice loss has a poor effect on the segmentation task with imbalanced categories. To address this problem, the present invention introduces a binary cross entropy (BCE) loss function, and weights the Dice and BCE losses to form a joint loss, so that the student network pays more attention to the details of the feature map on the unlabeled data and generates high-quality segmentation results. By introducing image-level prediction consistency L ILPC Loss, Feature-level Attention Consistency L FLAC Loss and Discriminator L D The loss is reasonably designed to improve the image segmentation accuracy. The combined loss function redesigned by the present invention is defined as shown in formula (4); L D (θ)=L ad (D(x i ,f i ),1)+L ad (D(M a (x p ,x q ),f(M a )),0) (3) L total =L D +L ILPC +L FLAC (4) Where: f(M a (x p ,x q ) represents the prediction result of the mixed image student model; represents the true label of the mixed image; M a (a' p ,a' q ) represents the mixed attention map obtained by the PolypMix operation; a(M a (x p ,x q )) represents the attention prediction result of the mixed image student model; M a (a' p ,a' q ) represents the real attention map of the mixed image; x i , f i Represents label images and prediction results; M a (x p ,x q ), f(M a ) represents unlabeled images and prediction results; S2: Establish a medical image dataset for model evaluation, including the following steps: S21: Selection of evaluation indicators: Seven indicators are selected: Dice coefficient, accuracy, recall rate, mean intersection ratio, Jaccard similarity coefficient, precision and average consistent distance, among which: Dice coefficient: used to measure the similarity between the segmentation result and the actual tumor area; Accuracy (ACC): The ratio of the number of correctly predicted pixels in the predicted segmentation results to the total number of pixels; Recall (REC): The percentage of correctly predicted lesion areas in the true label; Mean Intersection over Union (mIoU): Calculates the average of the ratio of the intersection and union of all categories; Jaccard Similarity Coefficient (Jaccard): Compares the difference and similarity between the tumor prediction results and the true labels; Precision (Prec): measures the maximum difference between the segmentation result and the actual tumor area; Mean distance to agreement (MDA): measures the average distance between each pixel in the segmentation result and the true tumor area; MDA(L,P)=mean l∈L,p∈P (||l-p||) (11) Where: TP represents the number of pixels correctly predicted as tumors; TN represents the number of pixels correctly predicted as non-tumor; FP represents the number of pixels incorrectly predicted as tumors; FN represents the number of pixels incorrectly predicted as non-tumor; PL and PP represent the real tumor area and predicted tumor area, respectively; L and P represent the tumor region sets of the true labels and the predicted results, respectively; S22: Verification of model effectiveness: DulecrASE-Net was compared with seven classic semi-supervised medical segmentation methods, including MT, UAMT, ICT, ADVENT, ClassMix, URPC, and PolypMixNet, and the fully supervised learning model U-Net using the Kvasir-SEG public colon polyp dataset. The semi-supervised learning model was trained using 15% of the labeled data. The comparison methods all used the same MSA-RFSEU-Net backbone and were trained and tested under the same experimental environment parameter settings. The DulecrASE-Nett model achieved the best segmentation effect, which was significantly higher than other models in terms of ACC, REC, Prec, and mIoU. The Dice coefficient also achieved good results. S23: Verification of the universality of the evaluation model: To further verify the performance of the network, the DulecrASE-Net model was compared with several different advanced semi-supervised methods including MT, UCMT, ASE-Net and PolypMix, as well as a fully supervised method using a self-built small intestinal tumor CT image dataset. The semi-supervised learning model was trained using 15% of the labeled data, and the comparison methods all used the same U-Net backbone.

2. The method for automatic segmentation of medical images based on semi-supervised learning according to claim 1, characterized in that: In the step S11, the DulecrASE-Net network consists of a segmentation network and a discrimination network, the segmentation network consists of a student model and a teacher model, both the student model and the teacher model use the designed MSA-RFSEU-Net segmentation network, and the discrimination network consists of a convolutional layer, DyMAC and global average pooling; the segmentation network and the discriminator are alternately trained to achieve adversarial two-level consistency learning, which can effectively utilize multi-level information from unlabeled data to solve the problem of imbalanced medical image categories, thereby improving segmentation accuracy; in terms of network structure, we use DyMAC to replace all convolutional layers except the first layer in the discriminator network to improve the feature representation ability of the network and reduce the risk of overfitting.

3. The method for automatic segmentation of medical images based on semi-supervised learning according to claim 1, characterized in that: In step S12, the MSA-RFSEU-Net network consists of an encoder and a decoder, and the encoder and decoder branches each consist of 5 blocks; The encoder uses ResNet34, and the decoder block first introduces the channel attention squeeze and excitation (SE) design RFSEConv to replace the ordinary convolution on the basis of the RFA (Receptive-FieldAttention) module to process the global information and adaptively reweight the features of each channel, thereby enhancing the network's ability to represent important features and improving the overall performance and adaptability of the model; then an MSA-Block is added after the first decoder block. MSA-Block introduces a spatial pyramid pooling structure and a batch normalization layer based on DA-Block, which effectively integrates information of multiple scales from global to local, enhances the network's ability to capture multi-scale spatial information, and maintains the integrity of fine-grained information while reducing the amount of computation.

4. The method for automatic segmentation of medical images based on semi-supervised learning according to claim 1, characterized in that: In step S22, in order to compare the differences between the model and other semi-supervised models and fully supervised models, the DulecrASE-Net network is compared and verified on the Kvasir-SEG public colon polyp dataset. It can be found that the adversarial two-level consistency training strategy and the multi-scale attention component of dynamic convolution can effectively utilize multi-level information from unlabeled data to solve the problem of imbalanced medical image categories, thereby improving segmentation accuracy.

5. The method for automatic segmentation of medical images based on semi-supervised learning according to claim 1, characterized in that: In step S23, in order to verify the universality of the DulecrASE-Net network, a comparative verification was performed on a self-built small intestinal tumor CT data set, and it was found that the DulecrASE-Net model achieved the best segmentation effect.

Images