A method and system for semantic segmentation of medical images
By employing a semi-supervised semantic segmentation strategy that combines cross-domain prediction of generative prerequisite tasks, cross-angle consistency prediction, and pseudo-label knowledge transfer, the robustness and network collapse issues in automatic segmentation of head and neck CT images are addressed, resulting in more reliable segmentation of head and neck CT images suitable for portable systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2024-01-18
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies suffer from insufficient robustness, neighborhood shift, and network collapse in automatic segmentation of head and neck CT images. In particular, due to the large individual differences, unclear boundaries, and low contrast in the head and neck region, it is difficult for automatic segmentation algorithms to accurately label the data, and pseudo-labels in semi-supervised learning may lead to training instability.
A novel semi-supervised semantic segmentation strategy is adopted, including cross-domain prediction of generative premise tasks, cross-angle consistency prediction, and orthogonal pseudo-label knowledge transfer. Through the co-training of two sub-networks, a general representation of CT images is learned, which reduces the domain bias and enhances the network's perception capability.
It improves the accuracy of head and neck CT image segmentation, resolves confirmation bias issues, generates more reliable predictions, and mitigates the risk of network crashes, making it suitable for portable systems.
Smart Images
Figure CN118298168B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of medical image processing methods, and in particular to a system and method for semantic segmentation of clinical CT images. Background Technology
[0002] Head and neck cancers are highly aggressive malignant tumors, including those in the oral cavity, pharynx, vocal cords, and various anatomical sites within the head and neck. They are characterized by short disease courses, rapid progression, poor postoperative outcomes, and frequent recurrence. The five-year survival rate is generally less than 60%, but timely surgical intervention can improve this statistic to 80%. Therefore, timely identification of head and neck lesions is crucial for improving the quality of life for patients suffering from these diseases. However, due to significant individual differences in the morphology and size of computed tomography (CT) images of the head and neck region, unclear boundaries, and low contrast with adjacent anatomical tissues, manual identification of head and neck lesions based on CT images is quite challenging.
[0003] In computer-aided medical diagnosis, existing technologies have conducted extensive research on automatic segmentation of head and neck images. However, compared to abdominal organs such as the liver and kidneys, some head and neck organs are smaller in size and have irregular boundaries. Accurate annotation of head and neck organs requires highly skilled medical personnel and is time-consuming and labor-intensive. Therefore, there is a significant lack of accurate annotation data for head and neck organs in existing technologies. In addition, existing research mainly focuses on single datasets. These problems have led to insufficient robustness of current automatic head and neck image segmentation algorithms.
[0004] On the other hand, various semi-supervised learning methods in existing technologies can combine the distribution information of unlabeled data with the semantic information of labeled data, and obtain better decision boundaries or classification models through learning, which is expected to overcome the above problems. An intuitive strategy of semi-supervised learning is to use pseudo-labels, that is, to use the predictions of unlabeled data as labels and merge them into the training dataset as pseudo-labels to enhance training capabilities. However, pseudo-labels may introduce confirmation bias, leading to unstable training and affecting the performance of neural networks. Another strategy of semi-supervised learning is to perform co-training, which improves model performance by training multiple sub-networks with each other and transferring learned knowledge, and can overcome the domain bias problem. However, most co-training requires the use of manual perturbations, which cannot effectively learn heterogeneous features to prevent sub-network collapse. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a novel semantic segmentation method and system applicable to clinical medical images, particularly CT images. This method employs a novel semi-supervised semantic segmentation strategy, effectively mitigating the domain bias and network collapse problems in existing technologies, enhancing the network's perception capabilities, generating more reliable predictions, and further resolving the confirmation bias problem. It is particularly suitable for head and neck image segmentation.
[0006] The technical solution of the present invention is as follows:
[0007] A medical image semantic segmentation method, comprising:
[0008] S1 acquires several CT data including head and neck CT images;
[0009] S2 processes the obtained CT data to obtain a dataset containing only head and neck data; the data processing includes removing non-head and neck data and marking organs at risk in some CT data to form a labeled dataset, while the remaining CT data forms an unlabeled dataset, and the labeled dataset and the unlabeled dataset together form the dataset.
[0010] S3 randomly selects a certain proportion of the unlabeled dataset and the CT data from the labeled dataset to form a training set for training the semantic segmentation model;
[0011] S4 randomly selects a certain proportion of CT data from the labeled dataset to form a validation set, and validates the trained semantic segmentation model using a scoring metric.
[0012] S5 performs head and neck CT image segmentation using a validated semantic segmentation model;
[0013] The semantic segmentation model includes two sub-networks, namely the first sub-network and the second sub-network. The first sub-network and the second sub-network have the same structure, both of which are composed of UNet based on ResNet18 encoder. The training is carried out under the generative premise task cross-domain prediction strategy, cross-angle consistency prediction strategy and orthogonal pseudo-label knowledge transfer strategy.
[0014] The cross-domain prediction strategy for generative prerequisite tasks includes:
[0015] The CT data in the training set are adjusted in Henle units according to the window width and window level of the tissue domain and the bone domain, respectively, to generate tissue domain image data and bone domain image data accordingly.
[0016] The tissue domain image data and the bone domain image data are respectively fed into two sub-networks of the semantic segmentation model for semantic segmentation to obtain their respective segmentation results. The model is then optimized using a cross-domain mean squared error loss function. During this process, the two sub-networks do not share neuron parameters.
[0017] The cross-angle consistency prediction strategy includes: after completing the cross-domain prediction strategy of the generative prerequisite task, the encoded features obtained by the encoder of one sub-network are spatially mapped to the feature space of another sub-network, and the mapped features are used as the encoded features of the other sub-network. The segmentation results are then obtained by the decoders of the two sub-networks respectively, and the model is optimized by the first difference loss function and the segmentation loss function.
[0018] The orthogonal-based pseudo-label knowledge transfer strategy includes: after completing the cross-angle consistency prediction strategy, both sub-networks are trained using the segmentation results output by the other's decoder as labels. During this process, the model is optimized using an orthogonal loss function and a second difference loss function.
[0019] According to some preferred embodiments of the present invention, the organs at risk include: brainstem, parotid gland, mandible, optic nerve and optic chiasm.
[0020] According to some preferred embodiments of the present invention, the data processing further includes: unifying the names of the same organ in all CT data in the unlabeled dataset and the labeled dataset; and adjusting all CT data and their labels in the unlabeled dataset and the labeled dataset to the same size.
[0021] According to some preferred embodiments of the present invention, the cross-domain mean square error loss function The settings are as follows:
[0022]
[0023] Where N represents the number of images, n represents the ordinal number of the image in the input subnetwork, H represents the height of the image, W represents the width of the image, m represents the ordinal number of the pixels in the image, and i represents the subnetwork ordinal number, i = 1 or 2. Representing image pairs respectively and in, Represents a skeletal domain image. Represents an organization domain image, () nm，i This indicates that m elements in the nth image pass through the i-th subnetwork, h i () indicates a decoding operation. This indicates an encoding operation.
[0024] According to some preferred embodiments of the present invention, the first difference loss function The settings are as follows:
[0025]
[0026] in, The features are extracted by the encoder of the first sub-network, namely encoder 1. The second subnetwork is obtained by mapping the output of encoder 1 to the feature space of the second subnetwork. The encoder 2 extracts the features of the second subnetwork. |||| represents the L2 norm.
[0027] According to some preferred embodiments of the present invention, the second difference loss function as follows:
[0028]
[0029] in, The segmentation features are obtained by the decoder of the first sub-network, namely decoder 1. The segmentation features obtained by decoder 2 of the second sub-network are obtained after mapping the segmentation features of decoder 1 to the labels of the second sub-network.
[0030] According to some preferred embodiments of the present invention, the mapping is implemented through a mapping layer, which includes a convolutional layer, a batch normalization layer, and a ReLU activation function having the same output and input dimensions.
[0031] According to some preferred embodiments of the present invention, the segmentation loss function The settings are as follows:
[0032]
[0033] in, and Let l represent the predicted value and label of the m-th pixel in the n-th image, respectively. ce This represents the cross-entropy loss.
[0034] According to some preferred embodiments of the present invention, the orthogonal loss function The settings are as follows:
[0035]
[0036] Among them, w c =2 represents the weighting coefficient used for stable orthogonal confidence prediction, δ nm,i This represents the gating coefficient, used to control whether the predictions of the two sub-networks are orthogonal. This represents a pseudo tag, where express The c-th prediction dimension, c = {1, ..., Y} represents the 1 to Y categories.
[0037] According to some preferred embodiments of the present invention, in the training of the semantic segmentation model, a total loss function is used for model optimization, and the total loss function is set as follows:
[0038]
[0039] Where λ1, λ2, λ3, and λ4 are the set trade-off parameters. Let L be the total cross-domain mean squared error loss function. dis The total difference loss function, The total segmentation loss function is... This is the total orthogonal loss function;
[0040] in:
[0041] in, This represents the cross-domain mean squared error loss function of the first subnetwork. This represents the cross-domain mean squared error loss function of the second subnetwork;
[0042] in, This represents the first difference loss function. This represents the second difference loss function;
[0043] in, This represents the segmentation loss function of the first subnetwork. This represents the segmentation loss function of the second subnetwork;
[0044] in, Denotes the orthogonal loss function of the first subnetwork. Let represent the orthogonal loss function of the second subnetwork.
[0045] According to some preferred embodiments of the present invention, the scoring index includes one or more of the following: Dice similarity coefficient, Jaccard coefficient, average surface distance, and Hausdorff distance.
[0046] The present invention further provides a semantic segmentation system for clinical medical images, which includes a data and CT image display module for displaying CT images and input data, a data input module for data input, an organ selection and segmentation module for selecting and segmenting organs, a segmentation result display module for displaying segmentation results, and an evaluation index calculation and display module for calculating and displaying scoring indicators. The present invention applies the above-mentioned semantic segmentation method to segment CT images.
[0047] The present invention has the following beneficial effects:
[0048] This invention introduces a novel generative pre-task strategy for cross-domain prediction in semi-supervised learning, namely the generative pre-task cross-domain prediction strategy. It can learn a general representation of CT images through mutual prediction between the tissue domain and the skeletal domain, which has good rationality and universality.
[0049] This invention introduces a cross-angle consistency prediction strategy in semi-supervised learning, which can minimize the similarity between the features extracted by two sub-networks, thereby encouraging the two sub-networks to extract different features, enabling them to obtain reliable predictions from two unrelated perspectives and promoting collaborative training.
[0050] This invention also introduces an orthogonal pseudo-label knowledge transfer strategy to transfer acquired knowledge from one sub-network to another, thereby enhancing the network's perception capabilities.
[0051] The semantic segmentation method of this invention adopts an innovative semi-supervised semantic segmentation strategy. It can learn the inherent attributes of CT images through a generative premise task cross-domain prediction strategy. Then, through a cross-angle consistency prediction strategy, it uses the acquired prior knowledge to promote the two sub-networks to independently learn different features from the same input. Furthermore, through an orthogonal pseudo-label knowledge transfer strategy, it enhances the perceptual ability of the sub-networks, successfully reduces the domain bias, prevents network collapse, generates more reliable predictions, and further solves the confirmation bias problem.
[0052] The semantic segmentation method or system of the present invention can be applied to CT multi-organ segmentation of patients with head and neck cancer, with excellent accuracy, and can be deployed in a portable system, increasing the application scenarios. Attached Figure Description
[0053] Figure 1 This is a flowchart illustrating the medical image semantic segmentation method in Example 1.
[0054] Figure 2 This is a schematic diagram of the data processing and model training process in Example 1.
[0055] Figure 3 This is a model training diagram under the cross-domain prediction strategy for generative prerequisite tasks described in Example 1.
[0056] Figure 4 This is the model training diagram under the cross-angle consistency prediction strategy described in Example 1.
[0057] Figure 5 This is the model training graph under the orthogonal pseudo-label knowledge transfer strategy described in Example 1.
[0058] Figure 6 This is a user interface diagram of the portable system in Example 1. Detailed Implementation
[0059] The present invention will now be described in detail with reference to embodiments and accompanying drawings. However, it should be understood that the embodiments and drawings are for illustrative purposes only and do not constitute any limitation on the scope of protection of the present invention. All reasonable modifications and combinations included within the inventive spirit of the present invention fall within the scope of protection of the present invention.
[0060] Example 1
[0061] See attached document Figure 1-2 The medical image semantic segmentation method of the present invention includes:
[0062] S1 collects CT data
[0063] In this embodiment, the CT data comes from public datasets from nine different institutions, including five public datasets from non-local institutions and four public datasets from local institutions.
[0064] The CT data collected from five non-local institutions' public datasets include: 46 CT image samples from the Head-NeckCetuximab collection; 92 CT image samples from the Jewish General Hospital of Montreal, Canada; 102 CT image samples from the University of Sherbrooke Hospital Centre, Canada; 41 CT image samples from the Rosemont Hospital of Montreal, Canada; and 65 CT image samples from the University of Montreal Hospital Centre, Canada. The CT data collected from four local institutions' public datasets include: 60 CT image samples from Hubei Cancer Hospital; 50 CT image samples from the Affiliated Hospital of Nanchang University; 50 CT image samples from Bengbu Hospital; and 80 CT image samples from Anhui Provincial People's Hospital.
[0065] S2 performs CT data processing to build a dataset.
[0066] It further includes:
[0067] S21 uses CT data collected from public datasets of five non-local institutions as unlabeled data to create an unlabeled dataset, and uses CT data collected from public datasets of four local institutions to label organs at risk as labeled data to create a labeled dataset. Organs at risk include: brainstem, parotid gland, mandible, optic nerve and optic chiasm.
[0068] S22 removes chest data from all CT data in both the unlabeled and labeled datasets, retaining only the head and neck data;
[0069] S23 unifies the names of the same organ in all CT data of the unlabeled dataset and labeled dataset obtained after processing by S22;
[0070] S24 adjusts all CT data and their labels in the unlabeled dataset and labeled dataset obtained after processing by S23 to a size of (512, 512) to obtain the training dataset, which is used to prepare for model training.
[0071] S3 trains the semantic segmentation model based on the dataset established in S2.
[0072] It further includes:
[0073] S31 establishes a semantic segmentation model that includes two sub-networks, namely the first sub-network and the second sub-network. The first sub-network and the second sub-network have the same structure and are both composed of UNet based on the ResNet18 encoder.
[0074] Based on the dataset obtained from S2, S32 trains the semantic segmentation model according to the cross-domain prediction strategy, the cross-angle consistency prediction strategy, and the orthogonal pseudo-label knowledge transfer strategy for the generative premise task. During training, the model is optimized by the cross-domain mean squared error loss function, the difference loss and the segmentation loss function, and the orthogonal loss and the difference loss function corresponding to the above strategies, respectively, to obtain the trained semantic segmentation model.
[0075] S32 further includes:
[0076] S321 adjusts all data in the training dataset according to the window width and window level of the tissue domain and the bone domain respectively using Henle units, and generates corresponding tissue domain image data and bone domain image data.
[0077] S322 (Refer to Appendix) Figure 3 A generative prerequisite task cross-domain prediction strategy is implemented, which includes: feeding the obtained tissue domain image data and bone domain image data into two sub-networks of the semantic segmentation model for semantic segmentation. During this process, the two sub-networks do not share neuron parameters and the model is optimized through a cross-domain mean squared error loss function.
[0078] S323 (Refer to Appendix) Figure 4The cross-angle consistency prediction strategy is implemented. The cross-angle consistency prediction strategy includes: after completing S322, the encoded features obtained by the encoder of one sub-network are spatially mapped to the feature space of another sub-network. The mapped features are used as the encoded features of the other sub-network. Then, the segmentation results are obtained by the decoders of the two sub-networks respectively. In this process, the two sub-networks can encode and extract different features of the same target from different angles. After being sent into their respective decoders, segmentation results from different angles can be obtained. The model is optimized by the first difference loss function and the segmentation loss function.
[0079] S324 (Refer to Appendix) Figure 5 An orthogonal pseudo-label knowledge transfer strategy is implemented, which includes: after completing S323, both sub-networks are trained using the segmentation results output by the other's decoder as labels. During this process, the model is optimized using orthogonal loss and second difference loss functions.
[0080] Wherein, the cross-domain mean square error loss function This can be used to calculate the error between the prediction and the label in two sub-networks, with the following settings:
[0081]
[0082] Where N represents the number of images, n represents the ordinal number of the image in the input subnetwork, H represents the height of the image, W represents the width of the image, m represents the ordinal number of the pixels in the image, and i represents the subnetwork ordinal number, i = 1 or 2. Representing image pairs respectively and in, Represents a skeletal domain image. Represents an organization domain image, () nm，i This indicates that m elements in the nth image pass through the i-th subnetwork, h i () indicates a decoding operation. This indicates an encoding operation.
[0083] The first difference loss function The settings are as follows:
[0084]
[0085] in, The features are extracted by the encoder of the first sub-network, namely encoder 1. The second subnetwork is obtained by mapping the output of encoder 1 to the feature space of the second subnetwork. The encoder 2 extracts the features of the second subnetwork. |||| represents the L2 norm.
[0086] The segmentation loss function The settings are as follows:
[0087]
[0088] in, and Let l represent the predicted value and label of the m-th pixel in the n-th image, respectively. ce This represents the cross-entropy loss.
[0089] Meanwhile, since both subnetworks require supervision using real labels, the overall first segmentation loss function... The settings are as follows:
[0090]
[0091] The orthogonal loss function The settings are as follows:
[0092]
[0093] Among them, w c =2 represents the weighting coefficient used for stable orthogonal confidence prediction, δ nm,i This represents the gating coefficient, used to control whether the predictions of two sub-networks are orthogonal. This represents a pseudo tag, where express The c-th prediction dimension, c = {1, ..., Y} represents the 1 to Y categories.
[0094] It is important to note that the orthogonal-based pseudo-label knowledge transfer strategy also encourages two sub-networks to reason about the same input image from different perspectives. Therefore, the second difference loss function in the orthogonal-based pseudo-label knowledge transfer strategy... as follows:
[0095]
[0096] in, The segmentation features are obtained by the decoder of the first sub-network, namely decoder 1. The segmentation features obtained by decoder 2 of the second subnetwork are obtained by mapping the segmentation features of decoder 1 to the labels of the second subnetwork. |||| represents the L2 norm.
[0097] Preferably, the first difference loss function Second difference loss function The mappings in the model are all implemented through a mapping layer, which includes a convolutional layer with the same output and input dimensions, a batch normalization layer, and a ReLU activation function.
[0098] Preferably, during training, the semantic segmentation model is optimized using a total loss function L, which is set as follows:
[0099]
[0100] Where λ1, λ2, λ3, and λ4 are the set trade-off parameters.
[0101]
[0102] Based on the above training process, in the first stage of model training under the generative premise task cross-domain prediction strategy, the generative premise task makes the two sub-networks sensitive to low-level features of CT images, enabling them to identify correlations between tissue and bone domains through cross-domain interaction.
[0103] In the second stage of training under the cross-angle consistency prediction strategy, the two sub-networks are forced to output different feature distributions. By mapping the output of the encoder of one sub-network to another feature space through the mapping layer, the two sub-networks can extract different features, thereby making consistent predictions from different angles.
[0104] In the third stage of training, model training based on orthogonal pseudo-label knowledge transfer strategy, one sub-network can learn more useful semantic information from another sub-network. The output of the encoder of one sub-network is mapped to another feature space through a mapping layer. This mapping is then compared with the output of the other sub-network, thereby enabling the two sub-networks to learn from each other through orthogonal predictions. This ensures that the two sub-networks can make consistent and reliable predictions and carry out stable training.
[0105] The mapping layers used in both the cross-perspective consistency prediction strategy and the orthogonal-based pseudo-label knowledge transfer strategy include convolutional layers, batch normalization layers, and ReLU activation functions with the same output and input dimensions.
[0106] In this embodiment, Adam is used to train each sub-network with a maximum number of iterations of 10,000, an initial learning rate of 1e-3, and an exponential decay rate of 0.99. The batch size and period are set to 16 and 80, respectively. The loss function's tradeoff parameters λ1, λ2, λ3, and λ4 are set to 1.0, 1.0, 1.0, and 1.0, respectively.
[0107] The training set used includes 25% or 75% of the CT data in the dataset, and includes both labeled and unlabeled data.
[0108] S4 validates the trained semantic segmentation model using scoring metrics based on CT data from the labeled dataset.
[0109] The scoring metrics used in this embodiment include the Dice similarity coefficient (DSC), Jaccard coefficient, average surface distance, and Hausdorff distance (HD). 25% or 75% of the CT data in the labeled dataset were used as the validation set.
[0110] In comparison, this embodiment also uses the PC2Seg method for model training and compares the differences between the two in scoring metrics. The comparison shows that on the Hubei Provincial Cancer Hospital dataset, the present invention outperforms the PC2Seg method by 1.5% and 1.7% in DSC, respectively; by 2.3% and 46% in Average Surface Distance (ASD), respectively; and by 5.4% and 29% in Hausdorff distance, respectively. On the Nanchang University Affiliated Hospital dataset, the present invention outperforms the PC2Seg method by 8.6% and 0.98% in DSC, and by 25% and 44% in Average Surface Distance (ASD), respectively; and by 31% and 3.1% in Hausdorff distance, respectively. On the Bengbu Hospital dataset… On the dataset, the method of this invention improves DSC by 4.9% and 2.3%, respectively; reduces mean surface distance (ASD) by 4.2% and 13%; improves Jaccard coefficient by 6.7% and 3.9%; and reduces Hausdorff distance by 16% and 8.1%, all of which are superior to the PC2Seg method. On the Anhui Provincial People's Hospital dataset, the method of this invention surpasses the PC2Seg method by 2.6% and 1.3% in DSC, reduces mean surface distance (ASD) by 3.2% and 27%, improves Jaccard by 5.2% and 3.34%, and maintains a similar level in Hausdorff distance.
[0111] Furthermore, this embodiment also conducted ablation experiments using 25% of the CT data in the dataset, including both labeled and unlabeled data. The experimental results showed that the various metrics of the model in the ablation experiment, such as the Dice similarity coefficient (DSC), Hausdorff distance (HD), and Jaccard coefficient, were not as good as the metrics obtained by applying the total loss function of this invention.
[0112] The above demonstrates that the model and training method of the present invention are effective and superior to existing methods.
[0113] S5 deploys the validated semantic segmentation model to portable systems.
[0114] In this embodiment, the portable system includes a data and CT image display module for displaying CT images and input data, a data input module for data input, an organ selection and segmentation module for selecting and segmenting organs, a segmentation result display module for displaying segmentation results, and an evaluation index calculation and display module for calculating and displaying scoring indicators. Its user interface is shown in the attached figure. Figure 6 As shown.
[0115] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A method for semantic segmentation of medical images, characterized in that, It includes: S1 acquires several CT data including head and neck CT images; S2 processes the obtained CT data to obtain a dataset containing only head and neck data; the data processing includes removing non-head and neck data and marking organs at risk in some CT data to form a labeled dataset, while the remaining CT data forms an unlabeled dataset, and the labeled dataset and the unlabeled dataset together form the dataset. S3 Randomly select a certain proportion of the unlabeled dataset and the CT data in the labeled dataset to form a training set to train the semantic segmentation model; S4 Randomly select a certain proportion of CT data from the labeled dataset to form a validation set, and validate the trained semantic segmentation model using a scoring metric; S5 performs head and neck CT image segmentation using the validated semantic segmentation model; The semantic segmentation model includes two sub-networks, namely the first sub-network and the second sub-network. The first sub-network and the second sub-network have the same structure, both of which are composed of UNet based on ResNet18 encoder. The training is carried out under the generative premise task cross-domain prediction strategy, cross-angle consistency prediction strategy and orthogonal pseudo-label knowledge transfer strategy. The cross-domain prediction strategy for generative prerequisite tasks includes: The CT data in the training set are adjusted in Henle units according to the window width and window level of the tissue domain and the bone domain, respectively, to generate tissue domain image data and bone domain image data accordingly. The tissue domain image data and the bone domain image data are fed into two sub-networks of the semantic segmentation model for semantic segmentation to obtain their respective segmentation results. The semantic segmentation model is then optimized using a cross-domain mean squared error loss function. During this process, the two sub-networks do not share neuron parameters. The cross-angle consistency prediction strategy includes: after completing the cross-domain prediction strategy of the generative prerequisite task, the encoded features obtained by the encoder of one sub-network are spatially mapped to the feature space of another sub-network, and the mapped features are used as the encoded features of the other sub-network. The segmentation results are then obtained by the decoders of the two sub-networks respectively, and the model is optimized by the first difference loss function and the segmentation loss function. The orthogonal-based pseudo-label knowledge transfer strategy includes: after completing the cross-angle consistency prediction strategy, both sub-networks are trained using the segmentation results output by the other's decoder as labels. During this process, the model is optimized using an orthogonal loss function and a second difference loss function.
2. The medical image semantic segmentation method according to claim 1, characterized in that, The organs at risk include: brainstem, parotid gland, mandible, optic nerve, and optic chiasm.
3. The medical image semantic segmentation method according to claim 1, characterized in that, The data processing also includes: unifying the names of the same organ in all CT data in the unlabeled dataset and the labeled dataset; and adjusting all CT data and their labels in the unlabeled dataset and the labeled dataset to the same size.
4. The medical image semantic segmentation method according to claim 1, characterized in that, The mapping is implemented through a mapping layer, which includes a convolutional layer with the same output and input dimensions, a batch normalization layer, and a ReLU activation function.
5. The medical image semantic segmentation method according to claim 1, characterized in that, The scoring metrics include one or more of the following: Dice similarity coefficient, Jaccard coefficient, average surface distance, and Hausdorff distance.
6. A medical image semantic segmentation system, comprising a data and CT image display module for displaying CT images and input data, a data input module for data input, an organ selection and segmentation module for selecting and segmenting organs, a segmentation result display module for displaying segmentation results, and an evaluation index calculation and display module for calculating and displaying scoring indicators, and applying the semantic segmentation method of any one of claims 1-5 to segment CT images.