A semi-supervised medical image segmentation data enhancement method and system based on cropping and splicing

Abstract

The application discloses a kind of semi-supervised medical image segmentation data enhancement methods and systems based on cutting and splicing, and relates to medical image processing technical field.The application includes: obtaining medical image data, carrying out horizontal cutting and vertical cutting to medical image data to obtain mixed picture data set;Pre-training set is input into Mean Teacher model, pre-training is carried out, and pre-trained Mean Teacher model is obtained;Training set is input into pre-trained Mean Teacher model, and after several iterations, trained Mean Teacher model is obtained;Test set is input into trained Mean Teacher model, and medical image segmentation result is evaluated.The application uses consistent learning strategy, makes full use of the complementarity of labeled and unlabeled data, further improves the segmentation effect, so that the model can show higher accuracy and reliability when facing complex medical images.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and particularly relates to a semi-supervised medical image segmentation data enhancement method and system based on clipping and splicing. BACKGROUND

[0002] In the field of medical image segmentation, semi-supervised learning has been widely applied to address the problem of insufficient labeled data. Although this approach can effectively combine labeled and unlabeled data, existing semi-supervised medical image segmentation methods still have some shortcomings, especially in how to effectively balance the use of these two types of data. Many existing methods usually use labeled data for supervised learning, while unlabeled data is processed through pseudo-labels or self-supervised learning. Although these methods improve the performance of the model to some extent, there are still some key problems. First, this kind of method often processes labeled and unlabeled data separately, failing to fully combine their potential synergies. Due to the failure to effectively integrate these two types of data, the utilization efficiency of labeled data is low, which limits the generalization ability of the model. In practical applications, the model may not be able to extract enough semantic information from the limited labeled data, while the unlabeled data also fails to fully play its role in the training process, ultimately leading to unsatisfactory segmentation results.

[0003] Although there are some strategies for fusing labeled and unlabeled data in existing methods, such as the bidirectional copy-paste strategy based on CutMix, this method alleviates the problem of distribution mismatch between labeled and unlabeled data to some extent, but its efficiency is not high, and there are still some significant defects. Specifically, the bidirectional copy-paste strategy has randomness when cutting and splicing, which may cause the key part of the image to be accidentally lost, thereby negatively affecting the accuracy of the segmentation result. At the same time, in the fused mixed image, the number of labeled and unlabeled pixels is often not equal, which makes the model unable to effectively learn the equivalent labeled and unlabeled knowledge. In this case, the model's ability is still limited when learning for unlabeled data, and the overall performance is not fully improved. In addition, the lack of knowledge transfer is particularly prominent in the case of labeled data scarcity, which makes it difficult for the model to effectively balance the advantages of labeled and unlabeled data, ultimately reducing the segmentation accuracy and consistency of the model. In existing semi-supervised methods, the quality problem of pseudo-labels cannot be ignored. In the training process of unlabeled data, the model is highly dependent on pseudo-labels, but the quality of pseudo-labels cannot be guaranteed. Especially in the case of labeled data scarcity, false pseudo-labels may be amplified by the model during training, further increasing noise and affecting the convergence and segmentation performance of the model. With the deepening of training, the accumulation of errors will further weaken the stability and accuracy of the model. In summary, although existing semi-supervised medical image segmentation methods have improved the utilization efficiency of labeled and unlabeled data to some extent, due to the fact that the existing fusion strategy has not achieved sufficient transfer of knowledge from labeled data to unlabeled data, there are still problems such as insufficient generalization ability, poor quality of pseudo-labels, and low data utilization efficiency. Therefore, how to design a more effective fusion strategy to maximize the use of labeled and unlabeled data and alleviate the distribution difference between the two is still a key problem that needs to be solved in the current semi-supervised medical image segmentation field.

[0004] Therefore, the present application provides a semi-supervised medical image segmentation data enhancement method and system based on cutting and splicing to solve the problems existing in the prior art, which is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the present application provides a semi-supervised medical image segmentation data enhancement method and system based on cutting and splicing to solve the problems existing in the prior art, which is a problem that needs to be solved by those skilled in the art.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0007] A semi-supervised medical image segmentation data enhancement method based on cropping and splicing, comprising the following steps:

[0008] S1. Obtaining data: obtaining medical image data, performing horizontal cropping and vertical cropping on the medical image data to obtain a mixed picture data set;

[0009] S2. Data preprocessing: preprocessing the obtained mixed picture data set;

[0010] S3. Data division: dividing the preprocessed mixed picture data set into a pretraining set, a training set and a test set;

[0011] S4. Model training: inputting the pretraining set into the Mean Teacher model, pretraining to obtain a pretrained Mean Teacher model, inputting the training set into the pre-trained Mean Teacher model, and obtaining a trained Mean Teacher model after several iterations;

[0012] S5. Segmentation result: inputting the test set into the trained Mean Teacher model to evaluate the medical image segmentation result.

[0013] Optionally, in S1, 50% of the obtained medical image data is labeled, the medical image data is horizontally cropped and vertically cropped, and the mixed picture data set is generated by component splicing. The mixed picture data set includes mixed labeled picture data, mixed unlabeled picture data, mixed picture data and mixed labeled picture data.

[0014] Optionally, the pretraining set in S3 is composed of mixed labeled picture data, and the training set and the test set are composed of mixed picture data.

[0015] Optionally, the Mean Teacher model in S4 is composed of a teacher network and a student network , wherein, is the parameter of the teacher network, is the parameter of the student network.

[0016] Optionally, in S4, in the pretraining phase, the data in the pretraining set is used, and the mixed label picture data is generated using the corresponding real mixed labeled picture data for supervision; in the training phase, the mixed picture data in the training set is used, and the mixed label picture data is used for supervision; the pretraining phase and the training phase are calculated by the following way:

[0017]

[0018] wherein, is the training set data, for the mixed image cropped horizontally for training the teacher network, tb for top and bottom, t for the teacher network teacher, and for the first and second mixed images for training the student network, for the mixed image cropped vertically for training the teacher network, lr for left and right, for the labeled image, for the unlabeled image, for element-wise multiplication, and are four zero-tensor masks for square, the four half-area values of up, down, left and right are set to 1 respectively;

[0019] The following formula is used to calculate the mixed label image data corresponding to the real mixed labeled image data in the pre-training stage for supervision:

[0020]

[0021] The following formula is used to calculate the mixed label image data in the training stage for supervision:

[0022]

[0023] wherein, is the pseudo label generated by the pre-trained teacher network, is and the corresponding real label and the mixed label after horizontal cropping and mixing, is and the corresponding real label and the mixed label after vertical cropping and mixing, is and the corresponding real label and the pseudo label the mixed label after horizontal cropping and mixing, is and the corresponding pseudo label and the real label the mixed label after horizontal cropping and mixing, is the corresponding real label, is the corresponding real label, is the corresponding pseudo label, is the corresponding real label, is the corresponding pseudo label, is the corresponding real label.

[0024] Optionally, in S4, in each iteration, firstly, the student network parameters are optimized using the stochastic gradient descent method, and the teacher network parameters are updated using the exponential moving average model of the student network parameters; in the pre-training stage of the teacher network, the loss function is defined as follows:

[0025]

[0026] The prediction result of the teacher network is calculated by the following formula:

[0027]

[0028] The total loss function in the pre-training stage is defined as:

[0029]

[0030] The total loss function in the training stage is defined as:

[0031]

[0032] wherein, is a linear combination of the loss and the cross-entropy loss, and is the prediction result of the teacher network, is the mixed image after horizontal cropping sent into the teacher network for training the prediction result obtained by the model prediction and the mixed label the loss function value calculated, is the mixed image after vertical cropping sent into the teacher network for training the prediction result obtained by the model prediction and the mixed label the loss function value calculated, and is and the corresponding real label and the mixed label after horizontal cropping and and the corresponding real label and the mixed label after vertical cropping and mixing, for the mixed image the loss function value of the prediction result obtained by model prediction and the corresponding mixed label after calculation, for the mixed image the loss function value, for the mixed image the loss function value, for the mixed image the loss function value.

[0033] A semi-supervised medical image segmentation data enhancement system based on cropping and splicing, applying any one of the above semi-supervised medical image segmentation data enhancement methods based on cropping and splicing, comprising: a data acquisition module, a data preprocessing module, a data division module, a model training module and a segmentation result module;

[0034] The data acquisition module is connected with the input end of the data preprocessing module, and is used for acquiring medical image data, performing horizontal cropping and vertical cropping on the medical image data to obtain a mixed picture data set;

[0035] The data preprocessing module is connected with the input end of the data division module, and is used for preprocessing the obtained mixed picture data set;

[0036] The data division module is connected with the input end of the model training module, and is used for dividing the preprocessed mixed picture data set into a pre-training set, a training set and a test set;

[0037] The model training module is connected with the input end of the segmentation result module, and is used for inputting the pre-training set into the MeanTeacher model to perform pre-training, obtaining the pre-trained MeanTeacher model, inputting the training set into the pre-trained MeanTeacher model, and obtaining the trained MeanTeacher model after several iterations;

[0038] The segmentation result module is connected with the output end of the model training module, and is used for inputting the test set into the trained MeanTeacher model to evaluate the medical image segmentation result.

[0039] According to the above technical solutions, compared with the prior art, the present application provides a semi-supervised medical image segmentation data enhancement method and system based on cropping and splicing, which has the following beneficial effects:

[0040] (1) The present application cuts the marked image and the unmarked image into the same size upper and lower and left and right components in the horizontal and vertical directions, and then carries out bidirectional mixed splicing based on pixels, effectively fusing the marked and unmarked data, significantly increasing the diversity of input data, solving the experience mismatch problem between the marked and unmarked data, and thus improving the overall performance of the model;

[0041] (2) The present application deploys the data enhancement method based on cutting and splicing on the Mean Teacher architecture with UNet or VNet as the backbone network, fully utilizes the complementarity of marked and unmarked data through consistent learning strategy, further improves the segmentation effect, and makes the model show higher precision and reliability when facing complex medical images;

[0042] (3) The cutting and splicing method proposed ensures that the amount of marked and unmarked pixels in the mixed image is equal, so that the model can learn the balanced semantic knowledge of marked and unmarked data at the same time, in addition, by converting each image into such a mixed image with equal pixel amount, it plays an important role in building a more unified data set, and significantly alleviates the distribution difference between marked and unmarked data;

[0043] (4) By effectively fusing marked and unmarked data, the present application significantly enhances the adaptability of the model to medical images, successfully alleviates the distribution mismatch between marked and unmarked data, improves the robustness of the model in handling diverse and complex scenarios, and makes it more effectively cope with the diversity and challenges in medical image segmentation. This innovation provides key support for high-performance optimization in the case of scarce labeled data, and lays a more reliable technical foundation for clinical application;

[0044] (5) The present application is widely used in image segmentation of important organs such as prostate and heart and diagnosis of related diseases, especially for prostate-related diseases (such as prostate hyperplasia, prostatitis and prostate cancer) and heart-related diseases, etc. Through data enhancement techniques such as cutting and splicing, the problem of data distribution imbalance is effectively alleviated, thereby significantly improving the segmentation performance, which makes the identification of lesions more accurate. This accurate lesion segmentation provides clearer and more accurate image analysis support for doctors, thereby enhancing the ability of early diagnosis and disease screening, and improving the accuracy of risk assessment for prostate diseases and heart diseases. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0046] Figure 1 A semi-supervised medical image segmentation data enhancement method based on cutting and splicing is provided for the present application.

[0047] Figure 2 A medical image cutting and splicing schematic diagram is provided for the present application.

[0048] Figure 3 A pre-training network structure diagram is provided for the present application.

[0049] Figure 4 A training network structure diagram is provided for the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0051] Referring to Figure 1 The present application discloses a semi-supervised medical image segmentation data enhancement method based on cutting and splicing, comprising the following steps:

[0052] S1. Data acquisition: acquiring medical image data, performing horizontal cutting and vertical cutting on the medical image data to obtain a mixed picture data set;

[0053] S2. Data preprocessing: preprocessing the obtained mixed picture data set;

[0054] S3. Data division: dividing the preprocessed mixed picture data set into a pre-training set, a training set and a test set;

[0055] S4. Model training: inputting the pre-training set into a Mean Teacher model, pre-training to obtain a pre-trained Mean Teacher model, inputting the training set into the pre-trained Mean Teacher model, and obtaining a trained Mean Teacher model after several iterations;

[0056] S5. Segmentation result: input the test set into the trained Mean Teacher model to evaluate the medical image segmentation result.

[0057] Further, 50% of the acquired medical image data in S1 is labeled, the medical image data is horizontally cropped and vertically cropped, and a mixed picture data set is generated by component splicing, the mixed picture data set including mixed labeled picture data, mixed unlabeled picture data, mixed picture data and mixed label picture data.

[0058] Further, the pre-training set in S3 is composed of mixed labeled picture data, and the training set and the test set are composed of mixed picture data.

[0059] Further, the Mean Teacher model in S4 is composed of a teacher network and a student network , wherein, is the parameter of the teacher network, is the parameter of the student network.

[0060] Further, in S4, in the pre-training stage, the data in the pre-training set is used, and the mixed label picture data is generated using the corresponding real mixed labeled picture data for supervision; in the training stage, the mixed picture data in the training set is used, and the mixed label picture data is used for supervision; the pre-training stage and the training stage are calculated by the following way:

[0061]

[0062] wherein, is the training set data, is the mixed image after horizontal cropping for training the teacher network, tb is top and bottom, t is the teacher network teacher, and is the first mixed image and the second mixed image for training the student network, is the mixed image after vertical cropping for training the teacher network, lr is left and right, is the labeled image, is the unlabeled image, is the element-wise multiplication, and is four zero tensor masks, which are square, and the upper, lower, left and right four half area values are respectively set to 1;

[0063] The following formula is used to calculate the mixed label picture data generated by the corresponding real mixed label picture data in the pre-training stage:

[0064]

[0065] The following formula is used to calculate the mixed label picture data generated by the corresponding real mixed label picture data in the pre-training stage:

[0066]

[0067] wherein, is the pseudo label generated by the pre-trained teacher network, is and the corresponding real label and the mixed label after horizontal cropping mixing, is and the corresponding real label and the mixed label after vertical cropping mixing, is and the corresponding real label and the pseudo label the mixed label after horizontal cropping mixing, is and the corresponding pseudo label and the real label the mixed label after horizontal cropping mixing, is the corresponding real label, is the corresponding real label, is the corresponding pseudo label, is the corresponding real label, is the corresponding pseudo label, is the corresponding real label.

[0068] Further, in S4, in each iteration, first, the student network parameters are optimized using the stochastic gradient descent method, and the teacher network parameters are updated using the exponential moving average model of the student network parameters. In the pre-training stage of the teacher network, the loss function is defined as follows:

[0069]

[0070] The prediction result of the teacher network is calculated by the following formula:

[0071]

[0072] The total loss function of the pre-training stage is defined as:

[0073]

[0074] The total loss function of the training stage is defined as:

[0075]

[0076] wherein, is a linear combination of the loss and the cross-entropy loss, and is the prediction result of the teacher network, is the mixed image after horizontal cropping sent to the teacher network for training the prediction result obtained by the model prediction and the mixed label the loss function value calculated, is the mixed image after vertical cropping sent to the teacher network for training the prediction result obtained by the model prediction and the mixed label the loss function value calculated, and is and the corresponding real label and the mixed label after horizontal cropping mixing and and the corresponding real label and the mixed label after vertical cropping mixing, is the mixed image the prediction result obtained by the model prediction and the loss function value calculated by the corresponding mixed label, is the loss function value of the mixed image , is the loss function value of the mixed image , is the loss function value of the mixed image .

[0077] Specifically, the loss value of the pseudo-supervised part is adjusted by multiplying it by a factor. The value in the default case is set to 0.5. The loss function of each mixed image is defined as follows:

[0078]

[0079] and are the prediction results of the student network, which can be calculated by the following formula:

[0080]

[0081] wherein, is the student network, is the parameter of the student network.

[0082] In a specific embodiment, the following contents are included:

[0083] Four medical images are selected, two labeled images and two unlabeled images The labeled and unlabeled images are combined in pairs and Through horizontal and vertical cropping in Figure 2 , mixed and spliced to form four mixed pictures, obtaining . The four mixed images are used to train the network, and the entire model framework is composed of the teacher network and the student network , wherein, is the parameter of the teacher network, is the parameter of the student network. Figure 3 and Figure 4 correspond to the pre-training and training stages of the entire model, respectively.

[0084] The present application is compared with various state-of-the-art semi-supervised segmentation methods on PROMISE 12 and LA datasets, and semi-supervised experiments are conducted for different labeling ratios. Four evaluation indicators are selected, including region-sensitive indicators: Dice coefficient (%) and Jaccard similarity coefficient (%), and edge-sensitive indicators: 95% Hausdorff distance (95HD) and average surface distance (ASD). The present application achieves the best performance on all four evaluation indicators, significantly outperforming other competitors. The Dice score of the present application on the PROMISE 12 dataset at two labeling ratios (20% and 40%) is increased by 3.70% and 4.08%, respectively. The Dice score on the LA dataset at two labeling ratios (10% and 20%) is increased by 1.04% and 0.57%, respectively.

[0085] From the perspective of non-image level, the application increases the diversity of data by merging labeled and unlabeled data in the form of patches, and encourages the model to learn the shared semantics of the two data types better. Extensive experiments on PROMISE 12 and LA datasets show the effectiveness of the method. The outstanding segmentation performance shows that the labeled data is used more effectively.

[0086] With Figure 1 Corresponding to the method, the embodiment of the application also provides a semi-supervised medical image segmentation data enhancement system based on cropping and splicing, which specifically comprises: a data acquisition module, a data preprocessing module, a data division module, a model training module and a segmentation result module.

[0087] The data acquisition module is connected with the input end of the data preprocessing module, and is used to acquire medical image data, and perform horizontal cropping and vertical cropping on the medical image data to obtain a mixed picture data set.

[0088] The data preprocessing module is connected with the input end of the data division module, and is used to preprocess the obtained mixed picture data set.

[0089] The data division module is connected with the input end of the model training module, and is used to divide the preprocessed mixed picture data set into a pretraining set, a training set and a test set.

[0090] The model training module is connected with the input end of the segmentation result module, and is used to input the pretraining set into the MeanTeacher model for pretraining to obtain a pretrained MeanTeacher model, and input the training set into the pre-trained MeanTeacher model to obtain a trained MeanTeacher model after several iterations.

[0091] The segmentation result module is connected with the output end of the model training module, and is used to input the test set into the trained MeanTeacher model to evaluate the medical image segmentation result.

[0092] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed by the embodiment, since it corresponds to the method disclosed by the embodiment, the description is relatively simple, and the related parts can be referred to the method part.

[0093] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A semi-supervised medical image segmentation data augmentation method based on cropping and stitching, characterized in that, The method comprises the following steps: S1. Data acquisition: acquire medical image data, perform horizontal and vertical cropping on the medical image data to obtain a mixed picture data set; S2. Data preprocessing: preprocessing the obtained mixed picture data set; S3. Data division: dividing the preprocessed mixed picture data set into a pretraining set, a training set and a test set; S4. Model training: inputting the pretraining set into the Mean Teacher model, pretraining to obtain a pretrained Mean Teacher model, inputting the training set into the pre-trained Mean Teacher model, and obtaining a trained Mean Teacher model after several iterations; S5. Segmentation result: inputting the test set into the trained Mean Teacher model to evaluate the medical image segmentation result; In the pretraining stage, the data in the pretraining set is used, and the mixed label picture data generated by the corresponding real mixed label picture data is used for supervision; In the training stage, the mixed picture data in the training set is used, and the mixed label picture data is used for supervision; the pretraining stage and the training stage are calculated by the following method: wherein, is the training set data, is the mixed image after horizontal cropping sent into the teacher network for training, tb is top and bottom, t is the teacher network teacher, and is the first mixed image and the second mixed image sent into the student network for training, is the mixed image after vertical cropping sent into the teacher network for training, lr is left and right, is the labeled image, is the unlabeled image, is the element-wise multiplication, and is the four zero-tensor masks, which are square, and the four half-area values of top, bottom, left and right are set to 1; The following formula is used to calculate the supervision of the mixed label picture data generated by the corresponding real mixed label picture data in the pretraining stage: The following formula is used to calculate the supervision of the mixed label picture data in the training stage: wherein, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network, is a pseudo label generated by a pre-trained teacher network.

2. The semi-supervised medical image segmentation data enhancement method based on cropping and splicing according to claim 1, wherein, In S1, 50% of the acquired medical image data is labeled, the medical image data is horizontally and vertically cropped, and the mixed picture data set is generated by component splicing, the mixed picture data set comprising mixed labeled picture data, mixed unlabeled picture data, mixed picture data and mixed label picture data.

3. The semi-supervised medical image segmentation data enhancement method based on cropping and splicing according to claim 1 or 2, wherein, The pretraining set in S3 is composed of mixed labeled picture data, and the training set and the test set are composed of mixed picture data.

4. The semi-supervised medical image segmentation data enhancement method based on cropping and splicing according to claim 1, wherein, The Mean Teacher model in S4 is composed of a teacher network and a student network wherein, is a parameter of the teacher network, is a parameter of the student network.

5. The semi-supervised medical image segmentation data enhancement method based on cropping and splicing according to claim 1, wherein, In S4, in each iteration, first, the student network parameters are optimized using the stochastic gradient descent method, and the teacher network parameters are updated using the exponential moving average model of the student network parameters. In the pretraining stage of the teacher network, the loss function is defined as follows: The prediction result of the teacher network is calculated by the following formula: The total loss function of the pretraining stage is defined as: The total loss function of the training stage is defined as: wherein, is a linear combination of the loss and the cross-entropy loss, and is the prediction result of the teacher network, is the mixed image after horizontal cropping sent into the teacher network for training is the prediction result obtained by the model prediction is the mixed label is the loss function value obtained by calculation, is the mixed image after vertical cropping sent into the teacher network for training is the prediction result obtained by the model prediction is the mixed label is the loss function value obtained by calculation, and is and is the corresponding real label and is the mixed label after horizontal cropping mixing, and and is the corresponding real label and is the mixed label after vertical cropping mixing, is the mixed image is the loss function value obtained by calculation of the prediction result obtained by the model prediction and the corresponding mixed label, is the loss function value of the mixed image , is the loss function value of the mixed image , is the loss function value of the mixed image .

6. A semi-supervised medical image segmentation data augmentation system based on cropping and pasting, characterized in that, The semi-supervised medical image segmentation data enhancement method based on cropping and splicing according to any one of claims 1-5 comprises: a data acquisition module, a data preprocessing module, a data division module, a model training module and a segmentation result module. The acquisition data module is connected with the input end of the data preprocessing module, and is used for acquiring medical image data, performing horizontal cutting and vertical cutting on the medical image data, and obtaining a mixed picture data set; The data preprocessing module is connected with the input end of the data division module, and is used for preprocessing the obtained mixed picture data set; The data division module is connected with the input end of the model training module, and is used for dividing the preprocessed mixed picture data set into a pre-training set, a training set and a test set; The model training module is connected with the input end of the segmentation result module, and is used for inputting the pre-training set into a Mean Teacher model, pre-training, obtaining a pre-trained Mean Teacher model, inputting the training set into the pre-trained Mean Teacher model, and obtaining a trained Mean Teacher model after a plurality of iterations; The segmentation result module is connected with the output end of the model training module, and is used for inputting the test set into the trained Mean Teacher model, and evaluating a medical image segmentation result.

Images