A semi-supervised three-dimensional medical image segmentation model training method and application

Abstract

This invention provides a training method and application for a semi-supervised 3D medical image segmentation model. It constructs a dual encoder-decoder system with identical structures and uses EMA to update parameters. Training is divided into two phases: warm-up and prototype guidance. In the warm-up phase, only low-uncertainty regions participate in pseudo-label consistency learning. In the prototype guidance phase, reliable voxel weighting is used to construct and dynamically fuse global category prototypes, expanding semantically matching voxels in high-uncertainty regions to the supervised range. Strong and weak view feature consistency constraints are applied to voxels with semantic conflicts. The model combines supervised segmentation loss and phased consistency loss for iterative optimization, fully utilizing unlabeled image information, suppressing pseudo-label noise and confirmation bias. In CT and MRI scenarios with low-label ratios, its segmentation accuracy and generalization robustness outperform existing methods.

Description

Technical Field

[0001] This invention relates to the field of computer signal processing, specifically to a training method and application of a semi-supervised three-dimensional medical image segmentation model, and particularly to a training method for a semi-supervised three-dimensional medical image segmentation model based on evidence uncertainty modeling and prototype semantic-guided region adaptive consistency learning, as well as a three-dimensional medical image segmentation method using this model. Background Technology

[0002] Medical image segmentation is a core foundational step in computer-aided diagnosis and treatment. By performing voxel-level classification and boundary delineation of target areas such as organs and lesions in 3D medical images such as CT, MRI, and ultrasound, it can provide precise quantitative evidence for disease diagnosis, surgical planning, prognostic assessment, and radiotherapy plan formulation. With the widespread use of 3D medical imaging equipment, a large amount of unlabeled image data has been generated in clinical settings. However, medical images generally have characteristics such as low contrast, blurred tissue boundaries, and large individual differences in anatomical structures. High-quality voxel-level annotation relies heavily on manual delineation by professional physicians, which is not only time-consuming, labor-intensive, and costly, but also easily influenced by the subjective experience of physicians. This results in a severe shortage of high-quality labeled data that can be used for fully supervised training, becoming a core bottleneck restricting the performance improvement and clinical application of deep learning segmentation models.

[0003] To fully utilize massive amounts of unlabeled data and alleviate labeling dependence, semi-supervised medical image segmentation technology has become a research hotspot in the field. Existing technical approaches mainly fall into two categories: pseudo-label driven and consistency regularization. Pseudo-label-based methods employ an iterative training strategy, first training an initial model with a small amount of labeled data, then using the model to generate predicted pseudo-labels from unlabeled data and retraining to gradually optimize model performance. However, this type of method is highly dependent on the prediction accuracy of the initial model; erroneous pseudo-labels accumulate and amplify during iteration, leading to severe confirmation bias problems and limiting the model's performance ceiling. Consistency regularization-based methods, based on the smoothness assumption, constrain the model's predictions for the same input by applying perturbations to the input, network, or features, and are currently the mainstream framework for semi-supervised segmentation. However, most existing consistency methods apply undifferentiated supervision constraints to all voxels without fine-grained differentiation of the reliability of the prediction results. For high-uncertainty regions such as tissue boundaries and blurred lesions, the pseudo-labels themselves have large errors; directly applying consistency constraints will introduce erroneous supervision signals into the training process, disrupting model feature learning and reducing segmentation accuracy.

[0004] Meanwhile, existing methods mostly construct constraints only at the output level, lacking guidance from global semantic structure information, making it difficult to effectively identify and utilize highly uncertain regions—either directly discarding the potential information of difficult regions, resulting in a waste of unlabeled volume data; or introducing noise due to incorrect supervision, making it impossible to achieve a balance between training stability and utilization of unlabeled volume data.

[0005] In summary, existing semi-supervised medical image segmentation techniques still suffer from drawbacks such as significant interference from pseudo-label noise, insufficient utilization of information in high-uncertainty regions, susceptibility to confirmation bias, and insufficient model generalization ability. Summary of the Invention

[0006] The purpose of this invention is to provide a training method and apparatus for a semi-supervised 3D medical image segmentation model, which can accurately assess prediction reliability, adaptively distinguish supervision strategies, and integrate global semantic information to further improve the segmentation accuracy and robustness of 3D medical images in scenarios with low annotation ratios.

[0007] Specifically, the semi-supervised 3D medical image segmentation model construction method in this scheme is based on evidence uncertainty and prototype guidance. During the training process of the semi-supervised 3D medical image segmentation model, supervised learning of labeled data and consistency learning of unlabeled data are introduced simultaneously. Pseudo-labels are generated using an exponential moving average teacher network, and reliable voxel regions are adaptively selected based on evidence uncertainty. In the later stages of training, a prototype is further constructed to refine the high-uncertainty regions with consistency masks, thereby expanding the usable pseudo-supervised region. In addition, for voxel regions with high uncertainty and divergence, strong and weak enhancement feature alignment constraints are introduced to improve the stability of consistency learning of unlabeled data, ultimately obtaining better segmentation model parameters.

[0008] To achieve the above objectives, this technical solution provides a semi-supervised training method for a 3D medical image segmentation model, comprising the following steps: S1: Obtain the training image dataset, preprocess the training image dataset to obtain the input dataset; wherein, the training image dataset includes labeled images with segmentation annotations and unlabeled images without segmentation annotations, and the input dataset includes labeled volume data generated corresponding to the labeled images and unlabeled volume data generated corresponding to all unlabeled images, and the labeled volume data and unlabeled volume data each include multiple three-dimensional medical image volume data. S2: Construct a semi-supervised 3D medical image segmentation architecture. The segmentation architecture adopts a dual-network framework consisting of a student network and a teacher network with identical structures. Both the student network and the teacher network are encoder-decoder 3D convolutional structures. The parameters of the teacher network are iteratively updated through the parameters of the student network using an exponential moving average method. S3: Perform weak augmentation and strong augmentation on each volume data in the input dataset to generate a spatially strictly aligned weakly augmented view and a strongly augmented view. S4: Based on weak and strong augmented views, perform two-stage progressive training on the semi-supervised 3D medical image segmentation architecture, including warm-up training and prototype-guided training, to obtain a semi-supervised 3D medical image segmentation model. When the number of warm-up training iterations reaches a preset iteration threshold, the prototype-guided training stage begins.

[0009] Secondly, this solution provides a three-dimensional medical image segmentation method, including: Acquire 3D medical images; The semi-supervised three-dimensional medical image segmentation model, trained by the aforementioned training method, outputs the segmentation results.

[0010] Compared with existing technologies, this technical solution has the following characteristics and beneficial effects: This solution addresses the shortcomings of existing semi-supervised 3D medical image segmentation methods, such as over-reliance on pseudo-labels, susceptibility to noise interference, easy generation of confirmation bias, and inability to fully exploit effective information in high-uncertainty regions. It constructs a dual-network framework consisting of a student network and a teacher network dynamically updated by exponential moving average, and generates spatially strictly aligned strong and weak enhancement views for each 3D volumetric data set. Based on the network output non-negative evidence tensor, this invention constructs a Dirichlet distribution to complete voxel-level uncertainty quantification, obtaining an interpretable voxel reliability metric. It uses an image-level average uncertainty adaptive threshold to divide low-uncertainty and high-uncertainty regions, achieving fine-grained region segmentation learning for unlabeled volumetric data. Training is executed in two progressive phases. In the warm-up phase, only reliable voxels with low uncertainty are selected for masked pseudo-label consistency learning to avoid the amplification of erroneous pseudo-labels in the early training stage. After iterations reach a preset threshold, prototype-guided training begins. High-reliability voxels within labeled and unlabeled volumetric data are selected and weighted to generate two categories of prototypes. A global category prototype is obtained through dynamic weight fusion, introducing global semantic structure information to compensate for the limitations of relying solely on local voxel confidence. The global category prototype is used to perform semantic analysis on the high-uncertainty regions segmented in the warm-up phase. The method involves validating and selecting reliable voxels that match the prototype predictions and teacher pseudo-labels to expand the effective supervision range, fully releasing the value of unlabeled voxels. For difficult voxels in high-uncertainty regions where the prototype predictions and teacher pseudo-labels differ, the method abandons the high-risk pseudo-label supervision and instead constrains the consistency of feature representations under the enhanced view of student network strength, replacing output layer supervision with feature alignment to block the continuous propagation of erroneous supervision signals. This invention adopts a region-decoupled differentiated learning strategy, balancing training stability and unlabeled voxel data utilization, effectively alleviating the confirmation bias of semi-supervised training. At the same time, the reliability-weighted prototype combined with a dynamic fusion mechanism can gradually fuse the global semantics of labeled and unlabeled voxels with training iterations, continuously improving feature discrimination ability and representation robustness. Validated on 3D CT and MRI medical image datasets, this invention significantly improves the segmentation accuracy of blurred boundaries of organs and lesions compared to traditional methods in clinical scenarios with low annotation ratios of 5%, 10%, and 20%, exhibiting stronger model generalization performance and high clinical auxiliary diagnosis and treatment application value. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is an architecture diagram of a semi-supervised three-dimensional medical image segmentation architecture according to an embodiment of this application; Figure 2 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0012] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0013] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0014] Example 1 This solution provides a semi-supervised training method for a 3D medical image segmentation model, including the following steps: S1: Obtain the training image dataset, preprocess the training image dataset to obtain the input dataset; wherein, the training image dataset includes labeled images with segmentation annotations and unlabeled images without segmentation annotations, and the input dataset includes labeled volume data generated corresponding to the labeled images and unlabeled volume data generated corresponding to all unlabeled images, and the labeled volume data and unlabeled volume data each include multiple three-dimensional medical image volume data. S2: Construct a semi-supervised 3D medical image segmentation architecture. The segmentation architecture adopts a dual-network framework consisting of a student network and a teacher network with identical structures. Both the student network and the teacher network are encoder-decoder 3D convolutional structures. The parameters of the teacher network are iteratively updated through the parameters of the student network using an exponential moving average method. S3: Perform weak augmentation and strong augmentation on each volume data in the input dataset to generate a spatially strictly aligned weakly augmented view and a strongly augmented view. S4: Based on weak and strong augmented views, perform a two-stage progressive training on the semi-supervised 3D medical image segmentation architecture, including warm-up training and prototype-guided training, to obtain a semi-supervised 3D medical image segmentation model. The prototype-guided training phase begins after the number of warm-up training iterations reaches a preset iteration threshold. The warm-up training process is as follows: For labeled data, the weakly augmented view corresponding to the labeled data is input into the student network, and a supervised segmentation loss is constructed by combining the voxel-level real labels. For unlabeled volume data, the weakly augmented view corresponding to the unlabeled volume data is input into the teacher network to generate voxel-level pseudo-labels, and the strongly augmented view corresponding to the unlabeled volume data is input into the student network to obtain evidence output. Pseudo-label consistency constraints are applied to the evidence output based on voxel-level pseudo-labels only in the low uncertainty region covered by the voxel-level reliability mask to obtain mask consistency loss. The total loss function for the warm-up stage is the weighted sum of the supervised segmentation loss and the mask consistency loss. The prototype-guided training process is as follows: Based on reliable voxel features from labeled and unlabeled data respectively, labeled category prototypes and unlabeled category prototypes are constructed by category aggregation. The global category prototype is obtained by dynamically weighting and fusing the labeled and unlabeled category prototypes. After the weak augmented view based on unlabeled volume data is input into the teacher network, the similarity between the voxel-level feature representation output by the decoder and the global category prototype is calculated, and the prototype predicted label is calculated based on the similarity. A consistency hint is constructed based on the prototype predicted label, and the voxel-level mask is optimized based on the consistency hint to obtain an optimized consistency mask. The optimized mask consistency loss is calculated based on the optimized consistency mask. For difficult voxels where the prototype predicted label and the voxel-level pseudo label are inconsistent in the high uncertainty region, the feature representation of the student network under strong and weak augmented views is extracted, and feature consistency loss is constructed only on difficult voxels based on the feature representation. The total loss function for training is modeled as a weighted sum of supervised segmentation loss, optimized mask consistency loss, and feature consistency loss.

[0015] As mentioned earlier, the training method for the semi-supervised 3D medical image segmentation model provided in this solution addresses the shortcomings of existing semi-supervised 3D medical image segmentation methods, such as susceptibility to pseudo-label noise, ineffective utilization of information in high-uncertainty regions, and susceptibility to confirmation bias. It constructs a dual-network architecture consisting of a structurally consistent student network and an EMA-dynamically updated teacher network, and generates spatially aligned strong and weak augmented views for each voxel's data. A two-stage progressive region adaptive consistency learning strategy is employed. In the warm-up stage, a Dirichlet distribution is constructed based on network output evidence to quantify voxel-level uncertainty, adaptively dividing regions of high and low uncertainty and generating reliability masks. Furthermore, pseudo-label consistency constraints are applied only to reliable voxels with low uncertainty, stabilizing the basic training process of the model and suppressing the accumulation of erroneous supervision signals. After training reaches a preset iteration threshold, the prototype guidance phase begins. A dynamically fused global category prototype is generated by weighted aggregation of reliable voxels with low uncertainty and accurate predictions from both labeled and unlabeled data. This global prototype is used to perform semantic verification on high-uncertainty regions defined in the warm-up phase. Voxels matching the prototype predictions and teacher pseudo-labels are expanded to the effective supervision region to improve the utilization of unlabeled data. Simultaneously, for difficult voxels where prototype and pseudo-label predictions conflict, pseudo-label supervision is abandoned, and instead, feature consistency constraints under the weak view are applied to enhance the robustness of the difficult region representation. The overall training employs backpropagation using multi-loss weighted functions corresponding to each stage to optimize the student network, while the teacher network is continuously updated smoothly through exponential moving average. This invention can fully exploit the effective semantic information of unlabeled data in low-annotation-ratio 3D CT / MRI medical imaging scenarios, effectively alleviating confirmation bias and improving the model's segmentation accuracy and generalization ability for ambiguous boundaries and lesion regions.

[0016] Regarding step S1 of this plan: The semi-supervised 3D medical image segmentation model trained by this scheme is suitable for segmenting 3D medical images. Therefore, the corresponding training images include 3D medical images such as CT and MRI.

[0017] In some embodiments, the labeled images in the training images are labeled images with voxel-level ground truth labels drawn manually by a physician, while the unlabeled images correspond to unlabeled images without segmentation annotations. That is, the segmentation result is the voxel-level ground truth label in the current training image.

[0018] Furthermore, this scheme divides the acquired 3D medical images into a training image dataset and a test image dataset. Each training image dataset must include labeled images and unlabeled images, with the number of unlabeled images exceeding the number of labeled images.

[0019] In the embodiments of this scheme, 5%, 10%, and 20% of the three-dimensional images in the training image dataset are retained as labeled images with voxel-level segmentation ground values ​​manually drawn by physicians, and the remaining images without manual segmentation annotations are regarded as unlabeled images. This simulates application scenarios with different degrees of annotation scarcity in clinical practice and is used to verify the segmentation performance of this semi-supervised segmentation method with a small amount of labeled data.

[0020] In some embodiments, the number of unlabeled body data in the input dataset is greater than the number of labeled body data.

[0021] Because the volume data of the original training images has inconsistent dimensions and high resolution, it cannot be directly fed into the network. Therefore, this solution requires image preprocessing of the training image dataset to obtain the input dataset. During image preprocessing, each training image is read sequentially according to the index, and the training images are randomly cropped into 3D image blocks of uniform standard size. The 3D image blocks of the cropped images corresponding to the labeled segmentation results are organized into labeled volume data, and the 3D image blocks of the unlabeled images are organized into unlabeled volume data. The input dataset obtained after preprocessing contains labeled and unlabeled volume data, and both labeled and unlabeled volume data are composed of multiple volumes of 3D image blocks.

[0022] Regarding step S2 of this plan: The semi-supervised 3D medical image segmentation architecture of this solution is as follows: Figure 1 As shown, this semi-supervised 3D medical image segmentation architecture uses a student network and a teacher network with identical structures to form a teacher-student dual-network framework. The teacher network is updated by the parameters of the student network through an exponential moving average method, which is used to stably generate pseudo-labels and uncertain information of unlabeled volume data during training.

[0023] As mentioned earlier, the student network and teacher network of this scheme adopt an encoder-decoder three-dimensional convolutional structure. The encoder extracts multi-scale features of each volume data through multi-level three-dimensional convolutional blocks and step-by-step downsampling. The decoder restores the resolution through step-by-step upsampling of the corresponding level. The decoded features of the decoder and the multi-scale features of the encoder at the same scale are added and fused element-wise through skip connections. This enables more complete feature transfer and fusion in the process of restoring resolution, thereby improving the ability to restore the details and boundaries of the segmented target.

[0024] In some embodiments, the decoder outputs evidence for each category and constrains the evidence to be non-negative using a non-negative activation function softplus. That is, the decoder outputs non-negative evidence for each category to characterize the model’s support strength for each category, while also outputting the decoder’s feature representation for subsequent consistency constraints.

[0025] It should be noted that the non-negative evidence output from the decoder of the student network refers to the non-negative evidence values ​​for each category output at the decoder end after the student network performs forward propagation on the input image (including strongly augmented views corresponding to labeled and unlabeled data). This non-negative evidence is used to calculate training objectives such as supervised loss, evidence consistency constraints, and feature consistency constraints. The non-negative evidence output from the decoder of the teacher network refers to the non-negative evidence values ​​for each category output at the decoder end after the teacher network performs forward propagation on the input image (weakly augmented views corresponding to unlabeled data). This non-negative evidence further generates voxel-level class probabilities, pseudo-labels, and voxel-level uncertainty information to guide the training of the student network.

[0026] Furthermore, the teacher network parameters in this scheme are iteratively updated using an exponential moving average method based on the student network parameters, as follows: ; in This represents the number of training iterations, and α is the smoothing coefficient. For the first The parameters of the teacher network in the next iteration. For the first Parameters of the teacher network in -1 iteration. For the first The parameters of the student network in the next iteration.

[0027] Regarding step S3: This scheme generates spatially strictly aligned weakly and strongly augmented views for each volume data set in the input dataset. On the one hand, it constructs a strong-weak pairing input consisting of weakly and strongly augmented views through differential perturbation, building a consistency regularization learning framework dedicated to the teacher-student network. It relies on spatial alignment to ensure accurate and effective calculation of voxel-level consistency loss and feature alignment loss. On the other hand, it introduces grayscale and noise perturbations through strong augmentation to improve the model's robustness to changes in the intensity of 3D medical images. At the same time, the weakly augmented view provides the teacher network with stable and reliable pseudo-labels, reducing the accumulation of training noise and alleviating the confirmation bias defect of semi-supervised learning.

[0028] Furthermore, this scheme utilizes a dual-network framework consisting of a teacher network and a student network to perform consistency learning from weakly augmented views to strongly augmented views on unlabeled volume data. Specifically, the weakly augmented view of the unlabeled volume data is input into the teacher network, which generates relatively stable voxel-level pseudo-labels using an exponentially moving average-updated teacher network. The strongly augmented view, aligned with the weakly augmented view space, is input into the student network, and the voxel-level prediction results obtained by the student network under strong perturbations are constrained to maintain consistency with the voxel-level pseudo-labels generated by the teacher network. This improves the model's robustness to input perturbations and reduces the impact of unstable pseudo-labels on the training process.

[0029] Specifically, each volume data is randomly rotated, randomly flipped, and randomly cropped to obtain a weakly enhanced view. The weakly enhanced view is then further subjected to random brightness and contrast adjustments and Gaussian noise processing to obtain a strongly enhanced view.

[0030] The following is a schematic diagram illustrating the handling of strongly enhanced and weakly enhanced views; The tagged body data is set as follows: ; ; ; in This represents the total number of body data in the labeled body data. For the i-th individual data in the labeled data, The data is labeled at the voxel level for the i-th individual data, i.e., the segmentation labeling result; H, W and D represent the height, width and depth of the volume data, respectively.

[0031] Unlabeled body data is represented as follows: ; ; in This represents the total number of unlabeled data items. This refers to the i-th individual data point in the unlabeled data set. Weak augmentation is performed on each volume data to obtain a weakly augmented view: ; in This indicates a weakly enhanced view.

[0032] Strong augmentation is performed on each individual's data to obtain a strongly augmented view: (exist Additional intensity perturbations are applied on top of the existing setup, including random brightness, contrast adjustment, and Gaussian noise. in This indicates a strongly enhanced view.

[0033] Regarding step S4 of this plan: First, let's introduce the prediction training phase. In the warm-up phase of this scheme, only low-uncertainty voxels with low confidence noise risk are adaptively selected based on the voxel-level uncertainty of the teacher network output to construct a reliability mask. Mask-based consistency supervision is performed only based on reliable voxels. On the one hand, this can avoid the confirmation bias problem caused by the back accumulation and amplification of erroneous pseudo-labels in high-uncertainty regions when the model's representation ability is weak in the early stage of training. This ensures the smooth completion of the model's basic semantic feature learning and avoids a large amount of noisy supervision signals interfering with the initial parameter update of the network. On the other hand, the supervised loss of labeled data provides the model with basic anatomical and lesion semantic discrimination capabilities. Combined with the consistency regularization framework of teacher-student dual networks and strong and weak enhancement dual views, a stable basic segmentation representation is gradually established. This provides reliable network weights and uncertainty measurement benchmarks for subsequent prototype-guided training to mine information in high-uncertainty and difficult regions, effectively reducing the iterative fluctuations of the overall two-stage training and improving the training robustness of the 3D medical image segmentation model in low-labeling scenarios.

[0034] The loss function for this preheating stage is as follows: ; in This represents the supervised segmentation loss on labeled data. This indicates the loss of mask consistency. The weight represents the mask consistency loss and is used to control the relative contribution of the mask consistency loss.

[0035] In some embodiments, the supervised segmentation loss for labeled data combines the Dice loss and the evidence learning loss under the Dirichlet framework, where the evidence learning loss under the Dirichlet framework corresponds to the Bayesian risk with a KL regularization term. The supervised segmentation loss is expressed as: ; in, Number of labeled data items; coefficient Used to perform an equally weighted combination of Dice loss and evidence learning loss; The Dice loss is calculated based on the expected class probability and the true voxel-level labels; The evidence learning loss is calculated based on the nonnegative evidence tensor and the voxel-level true labels; For the first After a weakly augmented view with labeled body data is input into the student network, the expected class probability is calculated based on the non-negative evidence output by the student network. The nonnegative evidence tensor output by the student network for the weakly augmented view; This corresponds to the voxel-level real labels.

[0036] In some embodiments, the formula for mask consistency loss is as follows: ; ; ; in, This indicates the number of unlabeled data points participating in the current training. An index representing untagged body data; Indicates the first A voxel grid consisting of all voxels contained in an unlabeled volume data. Indicates the first The total number of voxels contained in an unlabeled volume data set; Represents a voxel-level reliability mask; when When, it indicates voxels Located in the low-uncertainty region covered by the voxel-level reliability mask and incorporated into the pseudo-label consistency constraint; When, it indicates voxels Excluded from pseudo-label consistency constraints; The strongly augmented view of the j-th unlabeled volume data is represented by the evidence output through the student network. This represents the evidence output from the student network for the strongly augmented view of the j-th unlabeled volume data. The voxel-level pseudo-labels for the weakly augmented view of the j-th unlabeled volume data after passing through the teacher network; Indicates the first After a weakly augmented view of unlabeled volume data is input into the teacher network, the voxels are calculated based on the non-negative evidence output by the teacher network. Belongs to the Expected class probabilities for each category.

[0037] Specifically, the low uncertainty region in this scheme refers to the region where the total evidence output by the teacher network for the corresponding 3D volume data is relatively sufficient, and the voxel-level uncertainty calculated based on the total evidence is relatively low.

[0038] In some embodiments, a Dirichlet distribution is constructed based on each non-negative evidence tensor output by the teacher network. The voxel-level uncertainty value corresponding to each volume data is calculated based on the Dirichlet distribution. The image-level average uncertainty of a single 3D volume data is statistically analyzed as an adaptive threshold. The region consisting of all voxels whose voxel-level uncertainty values ​​are less than or equal to the adaptive threshold is taken as the low uncertainty region.

[0039] Specifically, the formula for constructing the Dirichlet distribution based on the non-negative evidence tensors output by the teacher network is as follows: ; ; in Here are the Dirichlet distribution parameters, and C is the number of categories. Let c be the non-negative evidence tensor of the c-th category. The Dirichlet intensity is the Dirichlet distribution.

[0040] The formula for calculating the voxel-level uncertainty value for each volume data point based on the Dirichlet distribution is as follows: ; in Let C be the Dirichlet intensity of the Dirichlet distribution, and C be the number of classes. The value represents the uncertainty at the voxel level.

[0041] Furthermore, the expected class probabilities obtained based on the Dirichlet distribution are as follows; ; in For Dirichlet distribution parameters, The Dirichlet intensity of the Dirichlet distribution. Let be the expected class probability.

[0042] The image-level average uncertainty of a single 3D volume data point is calculated and used as the adaptive threshold formula as follows:

[0043] in, The teacher network is targeting the first The image-level average uncertainty is calculated from each unlabeled volume data and used as the adaptive uncertainty threshold for that volume data. Indicates a teacher network; An index representing untagged body data; Indicates the first A voxel grid consisting of all voxels contained in an unlabeled volume data; This represents the total number of voxels in the voxel grid; Indicates voxel index; The teacher network is targeting the first voxels in unlabeled data The calculated voxel-level uncertainty value; Indicates the first The summation of the voxel-level uncertainty values ​​of all voxels in an unlabeled dataset.

[0044] The formula for defining the region of low uncertainty as the area comprised of all voxels whose voxel-level uncertainty values ​​are less than or equal to the adaptive threshold is expressed as follows: , in [] indicates an indicator function, which takes the value 1 when the condition inside the brackets is true, and 0 otherwise; The teacher network is targeting the first voxels in unlabeled data The calculated voxel-level uncertainty value, Indicates the first The average value of the voxel-level uncertainty of all voxels in an unlabeled volume data is used as the adaptive uncertainty threshold for that volume data.

[0045] The region comprising all voxels whose voxel-level uncertainty values ​​exceed the adaptive threshold is defined as the high uncertainty region, and its formula is expressed as follows: , in As a reliability indicator in unsupervised consistency learning The initial value is used to filter uncertain voxels to participate in pseudo-label supervision.

[0046] To ensure that voxels with low uncertainty and high reliability contribute more to the subsequent class prototype construction process, a voxel-level reliability weight map is constructed: ; in As a reliability indicator in unsupervised consistency learning The initial value, i.e., all voxels whose corresponding voxel-level uncertainty values ​​are less than or equal to this adaptive threshold, is used to filter voxels with uncertainty to participate in pseudo-label supervision. Indicates the first The average value of the voxel-level uncertainty of all voxels in an unlabeled volume data is used as the adaptive uncertainty threshold for that volume data.

[0047] Furthermore, both the student network and the teacher network output non-negative evidence tensors for each category. Based on the evidence, a Dirichlet distribution is constructed, and voxel-level uncertainty and category probability are calculated. Specifically, the non-negative evidence tensor output by the student network under a strongly augmented view is used to calculate the consistency loss, while the non-negative evidence tensor output by the teacher network under a weakly augmented view is used to generate pseudo-labels and voxel-level uncertainty. Both networks use the exact same Dirichlet distribution calculation method.

[0048] In some embodiments, prototype-guided training is performed after the number of warm-up training iterations has been reached. An introduction to prototype-guided training is as follows: Unlike warm-up training, which only performs pseudo-label supervision on low-uncertainty regions, this scheme introduces a prototype guidance mechanism in prototype-guided training to optimize the consistency mask.

[0049] In some embodiments, for labeled and unlabeled volume data, voxel feature representations under weak enhancement perspective are extracted, respectively, and denoted as... and Voxels whose voxel feature weight is greater than the corresponding voxel-level reliability weight are taken as reliable voxel features, and labeled category prototypes and unlabeled category prototypes are constructed according to the category, as follows: ; ; in, For prototypes with labeled categories, For unlabeled category prototypes, For the voxel-level reliability weights corresponding to labeled data, For voxel-level reliability weights corresponding to unlabeled volume data, This indicates the number of labeled data points in the current training batch; This indicates the number of unlabeled data points in the current training batch; This indicates that there is a tag body data index; This indicates an index of untagged body data; Indicates the first A set of all voxels contained in a labeled body data; Indicates the first The complete set of voxels contained in an unlabeled volume data; Represents a voxel index. Indicates the first Voxels in labeled data The true category label; The first number obtained from the teacher network prediction voxels in unlabeled data pseudo-tags; This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. This is a preset minimum positive number to prevent the denominator from being zero and to ensure the stability of numerical calculations.

[0050] The process of dynamically weighting and fusing labeled and unlabeled category prototypes to obtain the global category prototype is represented as follows: ; in For global category prototype, For prototypes with labeled categories, For unlabeled category prototypes, These are dynamic weights that change with training iterations.

[0051] Regarding the mask acquisition for the mask consistency loss in this scheme: After a weakly augmented view based on unlabeled volumetric data is input into the teacher network, the formula for calculating the similarity between the voxel-level feature representation output by the decoder and the global category prototype is as follows: ; ; in Let be the similarity of the c-th category. For global category prototype, For the corresponding unlabeled body data, the first Voxel-level pseudo-labels for individual data For cosine similarity, For the number of categories, Voxel representation The predicted probability of the prototype belonging to class c; Voxel representation The similarity between the teacher characteristics and the fusion prototype of type c; This represents an exponential function used to map similarity to positive numbers; Indicates the total number of segmentation categories; Indicates a category summation index; Voxel representation Teacher characteristics and the first Similarity between class-fusion prototypes; Voxel representation The sum of the exponential similarities of all class prototypes is used as the normalized denominator.

[0052] The formula for predicting labels based on similarity is as follows: ; in Predict labels for prototypes.

[0053] The formula for constructing consistency hints using prototype prediction labels is as follows: ; in It is a voxel-level pseudo-label. [] is an indicator function, which is 1 when the condition is met.

[0054] The formula for optimizing the voxel-level mask based on consistency hints is as follows: ; in This is a region of low uncertainty. This is a region of high uncertainty. For consistency indication, For consistency mask.

[0055] That is, the union of the low-uncertainty region and the high-uncertainty region of the corresponding data and the consistency hint is taken as the consistency mask.

[0056] The formula for calculating the optimized mask consistency loss based on the optimized consistency mask is as follows: ; in, This indicates the number of unlabeled data points participating in the current training. This indicates an index of untagged body data; Indicates the first The set of voxels consisting of all voxels in an unlabeled volume data. This indicates the total number of voxels contained in the voxel set; Indicates voxel index; Indicates the first voxels in unlabeled data The corresponding optimized consistency mask, when When, it indicates voxels When included in pseudo-label consistency constraints, When, it indicates voxels They are excluded from the pseudo-label consistency constraint, and thus feature consistency loss is used for constraint. The teacher network is targeting the first voxels in unlabeled data Generated voxel-level pseudo-labels; Indicates the first After inputting a strongly augmented view of unlabeled volumetric data into the student network, the student network targets the voxels. The output is non-negative evidence.

[0057] Regarding feature consistency loss: The high uncertainty region of this scheme is the area where the teacher network has low confidence in segmenting and predicting the corresponding individual data and a high probability of prediction error.

[0058] In some embodiments, a Dirichlet distribution is constructed based on each non-negative evidence tensor output by the teacher network. The voxel-level uncertainty value corresponding to each volume data is calculated based on the Dirichlet distribution. The image-level average uncertainty of a single 3D volume data is statistically analyzed as an adaptive threshold. The region consisting of all voxels whose voxel-level uncertainty values ​​are greater than the adaptive threshold is taken as the high uncertainty region.

[0059] This scheme does not employ pseudo-label supervision for difficult voxels in regions of high uncertainty where the prototype prediction is inconsistent with the teacher's pseudo-label. Instead, it constrains the difficult voxels based on feature representation and then calculates the feature consistency loss.

[0060] The difficult voxel representation for inconsistencies between prototype predicted labels and voxel-level pseudo-labels in regions of high uncertainty is as follows: ; in This is a region of high uncertainty. Predict labels for prototypes. It is a voxel-level pseudo-label. [] is an indicator function that returns 1 when the condition is met. The voxels are considered difficult voxels.

[0061] The feature representations extracted from the student network's strongly augmented and weakly augmented views are as follows: ; ; in Feature representation of a weakly augmented view of a student network. Feature representation of a strongly enhanced view of the student network.

[0062] Only in difficult voxels (i.e. We construct a feature consistency loss on the voxels, and the feature consistency loss is expressed as follows: ; in This represents the feature consistency loss between strongly and weakly augmented views. This indicates an index of untagged body data; Indicates voxel index; Indicates the first The complete set of voxels contained in an unlabeled volume data; Indicates the first voxels in unlabeled data The corresponding hard voxel indicator value is determined by the optimized consistency mask. Inverting the result yields the voxel. The value is 1 if the preset difficult voxel condition is met, and 0 otherwise. Represents a predefined, extremely small positive number; Indicates that the student network is targeting the first The voxels output by a strongly augmented view of unlabeled volume data Characteristic representation; Indicates that the student network is targeting the first The voxels output by a weakly augmented view of unlabeled volume data Characteristic representation; This represents the number of channels used to represent the features.

[0063] It should be noted that, in order to reduce computational overhead, this scheme selects at most 128 voxels from the difficult voxels for calculation in each iteration. If the number of difficult voxels does not exceed 128, then all difficult voxels are used for calculation.

[0064] After obtaining the optimized mask consistency loss and feature consistency loss, this scheme uses the weighted sum of the supervised segmentation loss, the optimized mask consistency loss, and the feature consistency loss as a prototype to guide the training of the total loss function, which is expressed as follows: ; in To monitor the division of losses, For the optimized mask consistency loss, For feature consistency loss, For feature alignment loss weights, This represents the weighting coefficient.

[0065] As mentioned above, this solution provides a training method for the semi-supervised 3D medical image segmentation model. The segmentation results of the trained semi-supervised medical image segmentation model are tested using a test set, and an evaluation index is calculated. If the evaluation index meets the requirements, the optimal parameters of the convolutional neural network model and data augmentation in the semi-supervised medical image segmentation model are obtained; otherwise, the parameters of the convolutional neural network model and data augmentation are adjusted, and the semi-supervised medical image segmentation model is retrained and tested.

[0066] Example 2 Based on the semi-supervised 3D medical image segmentation model constructed in Example 1, this scheme provides a 3D medical image segmentation method, including: Acquire 3D medical images; The three-dimensional medical images are input into a trained semi-supervised three-dimensional medical image segmentation model, which outputs the segmentation results.

[0067] In some embodiments, the segmentation result refers to the lesion content in a three-dimensional medical image.

[0068] Example 3 This embodiment also provides an electronic device, see reference. Figure 2 It includes a memory 402 and a processor 401. The memory 402 stores a computer program, and the processor 401 is configured to run the computer program to perform the steps of any of the above-described semi-supervised three-dimensional medical image segmentation model training methods or segmentation method embodiments.

[0069] Specifically, the processor 401 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0070] The memory 402 may include a large-capacity memory 402 for data or instructions. The memory 402 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 401.

[0071] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any of the training methods or segmentation methods for the semi-supervised three-dimensional medical image segmentation model in the above embodiments.

[0072] Optionally, the electronic device may further include a transmission device 403 and an input / output device 404, wherein the transmission device 403 is connected to the processor 401 and the input / output device 404 is connected to the processor 401.

[0073] The transmission device 403 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 403 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0074] The input / output device 404 is used to input or output information. In this embodiment, the input information may be a training image dataset, etc., and the output information may be a trained semi-supervised 3D image segmentation architecture, etc.

[0075] Optionally, in this embodiment, the processor 401 can be configured to perform the following steps via a computer program: S1: Obtain the training image dataset, preprocess the training image dataset to obtain the input dataset; wherein, the training image dataset includes labeled images with segmentation annotations and unlabeled images without segmentation annotations, and the input dataset includes labeled volume data generated corresponding to the labeled images and unlabeled volume data generated corresponding to all unlabeled images, and the labeled volume data and unlabeled volume data each include multiple three-dimensional medical image volume data. S2: Construct a semi-supervised 3D medical image segmentation architecture. The segmentation architecture adopts a dual-network framework consisting of a student network and a teacher network with identical structures. Both the student network and the teacher network are encoder-decoder 3D convolutional structures. The parameters of the teacher network are iteratively updated through the parameters of the student network using an exponential moving average method. S3: Perform weak augmentation and strong augmentation on each volume data in the input dataset to generate a spatially strictly aligned weakly augmented view and a strongly augmented view. S4: Based on weak and strong augmented views, perform a two-stage progressive training on the semi-supervised 3D medical image segmentation architecture, including warm-up training and prototype-guided training, to obtain a semi-supervised 3D medical image segmentation model. The prototype-guided training stage begins after the warm-up training reaches a preset iteration threshold. The warm-up training process is as follows: For labeled data, the weakly augmented view corresponding to the labeled data is input into the student network, and a supervised segmentation loss is constructed by combining the voxel-level real labels. For unlabeled volume data, the weakly augmented view corresponding to the unlabeled volume data is input into the teacher network to generate voxel-level pseudo-labels, and the strongly augmented view corresponding to the unlabeled volume data is input into the student network to obtain evidence output. Pseudo-label consistency constraints are applied to the evidence output based on voxel-level pseudo-labels only in the low uncertainty region covered by the voxel-level reliability mask to obtain mask consistency loss. The total loss function for the warm-up stage is the weighted sum of the supervised segmentation loss and the mask consistency loss. The prototype-guided training process is as follows: Based on reliable voxel features from labeled and unlabeled data respectively, labeled category prototypes and unlabeled category prototypes are constructed by category aggregation. The global category prototype is obtained by dynamically weighting and fusing the labeled and unlabeled category prototypes. Similarity is calculated based on voxel-level pseudo-labels and global category prototypes, and prototype predicted labels are calculated based on similarity. Consistency hints are constructed based on prototype predicted labels. The voxel-level mask is optimized based on consistency hints to obtain an optimized consistency mask. The optimized mask consistency loss is calculated based on the optimized consistency mask. For difficult voxels where the prototype predicted label and the voxel-level pseudo label are inconsistent in the high uncertainty region, the feature representation of the student network under strong and weak augmented views is extracted, and feature consistency loss is constructed only on difficult voxels based on the feature representation. The total loss function for training is modeled as a weighted sum of supervised segmentation loss, optimized mask consistency loss, and feature consistency loss.

[0076] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0077] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0078] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. One or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted that any block in the logical flow of the figures may represent a program step, or interconnected logical circuitry, blocks and functions, or a combination of program steps and logical circuitry, blocks and functions. The software may be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0079] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A training method for a semi-supervised three-dimensional medical image segmentation model, characterized in that, Includes the following steps: S1: Obtain the training image dataset, preprocess the training image dataset to obtain the input dataset; wherein, the training image dataset includes labeled images with segmentation annotations and unlabeled images without segmentation annotations, and the input dataset includes labeled volume data generated corresponding to the labeled images and unlabeled volume data generated corresponding to all unlabeled images, and the labeled volume data and unlabeled volume data each include multiple three-dimensional medical image volume data. S2: Construct a semi-supervised 3D medical image segmentation architecture. The segmentation architecture adopts a dual-network framework consisting of a student network and a teacher network with identical structures. Both the student network and the teacher network are encoder-decoder 3D convolutional structures. The parameters of the teacher network are iteratively updated through the parameters of the student network using an exponential moving average method. S3: Perform weak augmentation and strong augmentation on each volume data in the input dataset to generate a spatially strictly aligned weakly augmented view and a strongly augmented view. S4: Based on weak and strong augmented views, perform two-stage progressive training on the semi-supervised 3D medical image segmentation architecture, including warm-up training and prototype-guided training, to obtain a semi-supervised 3D medical image segmentation model. When the number of warm-up training iterations reaches a preset iteration threshold, the prototype-guided training stage is entered. The prototype-guided training process is as follows: Based on reliable voxel features from labeled and unlabeled data respectively, labeled category prototypes and unlabeled category prototypes are constructed by category aggregation. The global category prototype is obtained by dynamically weighting and fusing the labeled and unlabeled category prototypes. After the weakly augmented view based on unlabeled volume data is input into the teacher network, the similarity between the voxel-level feature representation output by the decoder and the global category prototype is calculated, and the prototype predicted label is calculated based on the similarity. A consistency hint is constructed based on the prototype predicted label, and the voxel-level reliability mask is optimized based on the consistency hint to obtain an optimized consistency mask. The optimized mask consistency loss is calculated based on the optimized consistency mask. For difficult voxels where the prototype predicted label and the voxel-level pseudo label are inconsistent in the high uncertainty region, the feature representation of the student network under strong and weak augmented views is extracted, and feature consistency loss is constructed only on difficult voxels based on the feature representation. The total loss function for prototype-guided training is the weighted sum of supervised segmentation loss, optimized mask consistency loss, and feature consistency loss.

2. The training method for the semi-supervised three-dimensional medical image segmentation model according to claim 1, characterized in that, The warm-up training process is as follows: For labeled data, the weakly augmented view corresponding to the labeled data is input into the student network, and a supervised segmentation loss is constructed by combining the voxel-level real labels. For unlabeled volume data, the weakly augmented view corresponding to the unlabeled volume data is input into the teacher network to generate voxel-level pseudo-labels, and the strongly augmented view corresponding to the unlabeled volume data is input into the student network to obtain evidence output. Pseudo-label consistency constraints are applied to the evidence output based on voxel-level pseudo-labels only in the low uncertainty region covered by the voxel-level reliability mask, resulting in mask consistency loss. The total loss function for the warm-up phase is the weighted sum of the supervised segmentation loss and the mask consistency loss.

3. The training method for the semi-supervised three-dimensional medical image segmentation model according to claim 1, characterized in that, Each training image dataset includes labeled and unlabeled images, with more unlabeled images than labeled images.

4. The training method for the semi-supervised three-dimensional medical image segmentation model according to claim 1, characterized in that, Each volume data is randomly rotated, randomly flipped, and randomly cropped to obtain a weakly enhanced view. The weakly enhanced view is then further subjected to random brightness and contrast adjustments and Gaussian noise processing to obtain a strongly enhanced view.

5. The training method for the semi-supervised three-dimensional medical image segmentation model according to claim 2, characterized in that, The supervised segmentation loss is expressed as: The supervised segmentation loss is expressed as: ； in, Number of labeled data items; coefficient Used to perform an equally weighted combination of Dice loss and evidence learning loss; The Dice loss is calculated based on the expected class probability and the true voxel-level labels; The evidence learning loss is calculated based on the nonnegative evidence tensor and the voxel-level true labels; For the first After a weakly augmented view with labeled body data is input into the student network, the expected class probability is calculated based on the non-negative evidence output by the student network. The nonnegative evidence tensor output by the student network for the weakly augmented view; For the corresponding voxel-level real labels; The formula for mask consistency loss is as follows: ； ； ； in, This indicates the number of unlabeled data points participating in the current training. An index representing untagged body data; Indicates the first A voxel grid consisting of all voxels contained in an unlabeled volume data. Indicates the first The total number of voxels contained in an unlabeled volume data set; Represents a voxel-level reliability mask; when When, it indicates voxels Located in the low-uncertainty region covered by the voxel-level reliability mask and incorporated into the pseudo-label consistency constraint; When, it indicates voxels Excluded from pseudo-label consistency constraints; The strongly augmented view of the j-th unlabeled volume data is represented by the evidence output through the student network. This represents the evidence output from the student network for the strongly augmented view of the j-th unlabeled volume data. The voxel-level pseudo-labels for the weakly augmented view of the j-th unlabeled volume data after passing through the teacher network; Indicates the first After a weakly augmented view of unlabeled volume data is input into the teacher network, the voxels are calculated based on the non-negative evidence output by the teacher network. Belongs to the Expected class probabilities for each category.

6. The training method for the semi-supervised three-dimensional medical image segmentation model according to claim 2, characterized in that, Dirichlet distributions are constructed based on the non-negative evidence tensors output by the teacher network. The voxel-level uncertainty values ​​corresponding to each volume data are calculated based on the Dirichlet distributions. The image-level average uncertainty of a single 3D volume data is used as an adaptive threshold. The region consisting of all voxels whose voxel-level uncertainty values ​​are less than or equal to the adaptive threshold is taken as the low uncertainty region.

7. The training method for the semi-supervised three-dimensional medical image segmentation model according to claim 1, characterized in that, The optimized formula for mask consistency loss is as follows: ； in, This indicates the number of unlabeled data points participating in the current training. This indicates an index of untagged body data; Indicates the first The set of voxels consisting of all voxels in an unlabeled volume data. This indicates the total number of voxels contained in the voxel set; Indicates voxel index; Indicates the first voxels in unlabeled data The corresponding optimized consistency mask, when When, it indicates voxels When included in pseudo-label consistency constraints, When, it indicates voxels They are excluded from the pseudo-label consistency constraint, and thus feature consistency loss is used for constraint. The teacher network is targeting the first voxels in unlabeled data Generated voxel-level pseudo-labels; Indicates the first After inputting a strongly augmented view of unlabeled volumetric data into the student network, the student network targets the voxels. Output non-negative evidence; Feature consistency loss is expressed as follows: ； in This represents the feature consistency loss between strongly and weakly augmented views. This indicates an index of untagged body data; Indicates voxel index; Indicates the first The complete set of voxels contained in an unlabeled volume data; Indicates the first voxels in unlabeled data The corresponding hard voxel indicator value is determined by the optimized consistency mask. Inverting the result yields the voxel. The value is 1 if the preset difficult voxel condition is met, and 0 otherwise. Represents a predefined, extremely small positive number; Indicates that the student network is targeting the first The voxels output by a strongly augmented view of unlabeled volume data Characteristic representation; Indicates that the student network is targeting the first The voxels output by a weakly augmented view of unlabeled volume data Characteristic representation; This represents the number of channels used to represent the features.

8. A three-dimensional medical image segmentation method, characterized in that, include: Acquire three-dimensional medical images; input the three-dimensional medical images into the semi-supervised three-dimensional medical image segmentation model trained by the training method of any one of claims 1 to 7, and output the segmentation results.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to execute the training method of the semi-supervised three-dimensional medical image segmentation model according to any one of claims 1 to 7 or the three-dimensional medical image segmentation method according to claim 8.

Images