A cardiac CT image semi-supervised segmentation method, device and medium

Abstract

The application discloses a kind of heart CT image semi-supervised segmentation method, equipment and medium, it is related to artificial intelligence technical field, comprising: obtaining labeled and unlabeled data;The semi-supervised segmentation model consisting of shape auto-encoder, teacher network and student network is constructed, and the encoding-decoding main network with feature fusion module is arranged in the bottleneck layer shared by teacher and student network;Pre-training shape auto-encoder obtains frozen prior projector;Teacher network predicts and is corrected by projector to unlabeled sample, and confidence map is constructed according to the difference between prediction and correction;After applying strong enhancement to unlabeled sample, it is input into student network to predict, and then student projection output is obtained by projector;Weighted consistency loss between student and teacher projection is calculated based on confidence map, and the model is optimized by combining basic supervision loss and shape prior loss.The application utilizes frozen prior projector to constrain anatomical rationality, suppresses low-quality pseudo-label interference with confidence map, and realizes high-precision automatic segmentation under label limitation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a semi-supervised segmentation method, device and medium for cardiac CT images. Background Technology

[0002] Cardiac CT image segmentation is of great value for the clinical diagnosis and surgical planning of cardiovascular diseases. However, the complex anatomy of the heart and the blurred boundaries of its substructures, coupled with the fact that high-quality pixel-level annotations rely on a large number of professional personnel, result in a scarcity of annotated samples, posing a severe challenge to this task.

[0003] In existing technologies, to alleviate the reliance on large-scale labeled data, semi-supervised segmentation frameworks based on consistency regularization have been developed. The basic idea is to impose prediction consistency constraints on unlabeled data to uncover potential structural information. However, these schemes typically assign the same constraint strength to prediction outputs under different augmentation perspectives, failing to effectively distinguish the quality of pseudo-labels themselves. When the network makes low-confidence predictions for morphologically complex or vaguely defined cardiac regions, the introduced pseudo-label noise is still equally involved in model parameter updates, leading to overfitting of the model to local patterns in a limited set of labeled samples.

[0004] In clinical applications, the aforementioned defects often lead to anatomically unreasonable deformations or misjudgments of regions in the segmentation results, and such deviations often require time-consuming manual correction by experienced physicians. Summary of the Invention

[0005] This invention provides a semi-supervised segmentation method, device, and medium for cardiac CT images, aiming to solve the technical problem of: how to provide an effective solution for high-precision automatic segmentation of three-dimensional cardiac CT images under the premise of limited labeled data.

[0006] In a first aspect, the present invention provides a semi-supervised segmentation method for cardiac CT images, comprising: Obtain both labeled cardiac CT image datasets and unlabeled cardiac CT image datasets; A semi-supervised segmentation model is constructed, which includes a shape autoencoder, a teacher network, and a student network. The teacher network and the student network share the same segmentation backbone network. The segmentation backbone network is based on an encoder-decoder framework and a feature fusion module is configured at the bottleneck layer. The feature fusion module includes a first sub-module for nonlinear mapping and a second sub-module for long-range dependency capture. The shape autoencoder is pre-trained using the labeled cardiac CT image dataset to obtain a frozen shape prior projector; The unlabeled samples from the unlabeled cardiac CT image dataset are input into the teacher network to obtain the teacher network prediction results; The teacher network prediction results are input into the frozen shape prior projector to obtain shape correction results. The shape correction results are used as teacher projection outputs, and a confidence map is constructed based on the difference between the teacher network prediction results and the shape correction results. A preset strong enhancement operation is applied to the unlabeled samples to obtain strongly enhanced samples; the strongly enhanced samples are then input into the student network to obtain the student network prediction results. The student network prediction results are input into the frozen shape prior projector to obtain the student projection output; Based on the confidence map, the confidence-weighted consistency loss between the student's projected output and the teacher's projected output is calculated, and the parameters of the semi-supervised segmentation model are optimized based on the total loss. The total loss includes the confidence-weighted consistency loss, the basic supervision loss determined based on the prediction results of labeled data, and the shape prior loss determined based on shape prior constraints.

[0007] Optionally, the processing flow of the first sub-module of the feature fusion module includes: The input features are sequentially passed through the first standard convolution, the first KAN layer, the first depthwise convolution, and the first linear layer to form the first processing path; The output of the first processing path is sequentially passed through a second standard convolution, a second KAN layer, and a second depthwise convolution to form a second processing path. A bypass is branched off from the output of the first linear layer; The bypass is gated by ReLU and then added to the residual of the end result of the second processing path to obtain the output feature of the first sub-module.

[0008] Optionally, the second submodule of the feature fusion module includes multiple DVSSLayers, and the processing flow of a single DVSSLayer includes: The input features are linearly mapped to obtain the mapped features; The mapping features are decomposed into low-frequency sub-bands and multiple high-frequency sub-bands by discrete wavelet transform; After nonlinear activation, the low-frequency sub-band and each of the high-frequency sub-bands are input into a VMamba-based state space unit for sequence modeling to obtain the sequence features corresponding to each sub-band. The sequence features are then gated and fused to obtain the fused features; The fused features are restored to their original spatial resolution using inverse discrete wavelet transform to obtain the restored features; The residuals of the recovered features and the input features are added together to obtain the output features of DVSSLayer.

[0009] Optionally, pre-training the shape autoencoder using the labeled cardiac CT image dataset includes: A three-dimensional region of interest localization network was trained using the labeled cardiac CT image dataset to obtain the trained three-dimensional region of interest localization network; The original CT images from the labeled cardiac CT image dataset are input into the trained 3D region of interest localization network. The 3D region of interest localization network regresses the center heatmap, offset map and 3D scale map from the original CT images to obtain the whole heart 3D bounding box. The original CT image is cropped and resampled using the whole-center 3D bounding box to construct a normalized region of interest dataset; The shape autoencoder is trained based on labeled samples in the region of interest dataset.

[0010] Optionally, constructing a confidence map based on the difference between the teacher network prediction results and the shape correction results includes: Calculate the system-by-system absolute difference between the teacher network prediction results and the shape correction results; The absolute difference of each system is normalized to a preset interval to obtain the confidence map, which is used to indicate the reliability of the prediction result at the corresponding location.

[0011] Optionally, calculating the confidence-weighted consistency loss between the student's projected output and the teacher's projected output based on the confidence map includes: Calculate the Dice loss and cross-entropy loss between the student's projected output and the teacher's projected output; The confidence-weighted consistency loss is obtained by weighting and summing the Dice loss and the cross-entropy loss with the weights at the corresponding positions in the confidence graph.

[0012] Optionally, optimizing the parameters of the semi-supervised segmentation model based on the total loss includes: determining the total loss; and updating the parameters of the student network using the total loss. Determining the total loss includes: Obtain the labeled prediction results of the student network on the labeled cardiac CT image dataset, and determine the basic supervision loss based on the labeled prediction results and the corresponding real annotation information; The labeled prediction result is input into the frozen shape prior projector to obtain the labeled projection result, and the shape prior loss is determined based on the difference between the labeled prediction result and the labeled projection result. Obtain the confidence-weighted consistency loss; The sum of the basic supervision loss, the shape prior loss, and the confidence-weighted consistency loss is taken as the total loss; The step of updating the parameters of the student network using the total loss includes: updating the parameters of the student network using backpropagation with the total loss.

[0013] Optionally, the method further includes: updating the parameters of the teacher network, wherein updating the parameters of the teacher network includes: Obtain the updated parameters of the student network in the current iteration round; Based on the updated parameters and the previous round parameters of the teacher network in the previous iteration round, an exponential moving average operation is performed to obtain the current round parameters of the teacher network in the current iteration round. The current round parameters are used as the parameters for the teacher network update in this round.

[0014] Secondly, the present invention also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0015] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0016] This invention provides a semi-supervised segmentation method, device, and medium for cardiac CT images. Addressing the challenge of high-precision automatic segmentation of 3D cardiac CT images with limited labeled data, it constructs a teacher-student semi-supervised architecture that integrates anatomical shape priors. In the feature extraction stage, the feature fusion module configured in the bottleneck layer of the segmentation backbone network enhances the model's ability to model the global morphology and local boundaries of complex 3D cardiac structures through the synergistic effect of nonlinear mapping and long-range dependency capture, compensating for the limitation of limited labeled samples in supporting complex feature learning. Furthermore, in the semi-supervised learning stage, a pre-trained frozen shape prior projector is used to apply anatomical constraints to the teacher network's prediction results to generate shape correction results. The system-by-system difference between the two is transformed into a confidence map, achieving regional adaptive evaluation of the reliability of pseudo-labels for unlabeled samples. Furthermore, in the optimization stage, the basic supervised loss, shape prior loss, and weighted consistency loss weighted by the confidence map are combined to form the total loss. This not only suppresses structural errors by forcing the output to stay within the anatomical manifold through shape prior, but also reduces the interference weight of low-quality pseudo-labels by using the confidence map. This allows the model to efficiently mine structural information in unlabeled data even when labeling is limited, significantly improving segmentation accuracy and anatomical rationality. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the semi-supervised segmentation method for cardiac CT images provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an MTKSeg segmentation backbone network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a TKAN block provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an MTKAN block and a DVSBlock provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a shape autoencoder pre-training process provided in an embodiment of the present invention; Figure 6 A schematic diagram of the overall framework of a semi-supervised segmentation model for cardiac CT images based on anatomical shape priors and hybrid networks, provided for an embodiment of the present invention; Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0021] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0024] Please see Figure 1 This invention provides a semi-supervised segmentation method for cardiac CT images. The semi-supervised segmentation model is a deep learning architecture that integrates a mean teacher framework and anatomical shape prior constraints. By constructing a teacher network, a student network, and a shape autoencoder, it achieves high-precision 3D segmentation of cardiac substructures using only a small amount of labeled data and a large amount of unlabeled data. Specifically, the method includes the following steps: S1, obtain the labeled cardiac CT image dataset and the unlabeled cardiac CT image dataset.

[0025] In practice, both labeled and unlabeled cardiac CT image datasets are acquired. Specifically, the labeled cardiac CT image dataset contains a small number of 3D CT scan images and their corresponding 3D segmentation masks, each annotated voxel-by-voxel by professional physicians for the five cardiac structures (left ventricular myocardium, left atrium, left ventricle, right atrium, and right ventricle). The scarcity of labeled samples is a core bottleneck in clinical medical image segmentation. The unlabeled cardiac CT image dataset contains a large number of images that only contain raw CT scan data and lack manually labeled information.

[0026] S2, Construct a semi-supervised segmentation model, which includes a shape autoencoder, a teacher network, and a student network. The teacher network and the student network share the same segmentation backbone network. The segmentation backbone network is based on an encoder-decoder framework and a feature fusion module is configured at the bottleneck layer. The feature fusion module includes a first sub-module for nonlinear mapping and a second sub-module for long-range dependency capture.

[0027] In specific implementation, a semi-supervised segmentation model is constructed. This semi-supervised segmentation model includes a shape autoencoder, a teacher network, and a student network. The teacher network and the student network share the same segmentation backbone network. Further, the segmentation backbone network is based on an encoder-decoder framework, and a feature fusion module is configured at the bottleneck layer. Figure 2 As shown, the segmentation backbone network is the MTKSeg segmentation backbone network, which is based on a 3D U-Net architecture and includes an encoder and a decoder. The encoder extracts multi-scale features stepwise through three convolutional downsampling modules, and the decoder fuses the features of the encoder at the corresponding scale through three upsampling modules and skip connections to restore spatial resolution. A closed-loop interconnection structure consisting of multiple MTKAN blocks is configured in the bottleneck layer between the encoder and the decoder.

[0028] Furthermore, the encoder is used to extract multi-scale spatial and semantic features layer by layer from the original image, and the decoder gradually recovers the spatial resolution and outputs a segmentation probability map through upsampling and skip connections. Furthermore, the feature fusion module configured at the bottleneck layer is designed to further interact with and enhance the highly abstract deep features. Furthermore, the feature fusion module includes a first submodule for nonlinear mapping and a second submodule for long-range dependency capture.

[0029] like Figure 6As shown, the semi-supervised segmentation model is constructed based on the mean-based teacher framework, configured with a teacher network and a student network sharing the same segmentation backbone network. The prediction results of the student network are input into the frozen shape autoencoder to obtain the student projection output; the prediction results of the teacher network are input into the shape autoencoder to obtain the shape correction result, which is then used as the teacher projection output. Simultaneously, a confidence map is constructed based on the difference between the teacher network's prediction results and the shape correction result. The confidence map, the correction function of the shape autoencoder, and the teacher-student collaborative structure together constitute the generation process of the confidence-weighted consistency loss.

[0030] In some preferred embodiments, the processing flow of the first sub-module of the feature fusion module includes: passing the input features sequentially through a first standard convolution, a first KAN layer, a first depthwise convolution, and a first linear layer to form a first processing path; passing the output of the first processing path sequentially through a second standard convolution, a second KAN layer, and a second depthwise convolution to form a second processing path; branching off a bypass from the output of the first linear layer; and adding the residual of the bypass after ReLU gate and the end result of the second processing path to obtain the output features of the first sub-module.

[0031] In specific implementation, the first submodule is a TKAN block, which enhances the network's ability to nonlinearly fit complex features by fusing the KAN layer with the residual structure. The specific implementation is as follows: the input features sequentially pass through a first standard convolution, a first KAN layer, a first depthwise convolution, and a first linear layer to form a first processing path. The first standard convolution uses a 3x3 spatial convolution kernel to perform preliminary local receptive field feature extraction on the input features. The first KAN layer performs nonlinear transformation on the features based on learnable B-spline basis functions. The KAN layer replaces the fixed activation function in the traditional multilayer perceptron with a learnable spline function. The first depthwise convolution, while maintaining the independence of each feature channel, is responsible for capturing the spatial context information between adjacent voxels. The first linear layer is used for linear remapping of the channel dimensions.

[0032] Furthermore, the output of the first processing path is sequentially passed through a second standard convolution, a second KAN layer, and a second depthwise convolution to form a second processing path. This repetitive combination structure aims to perform a deep nonlinear mapping of features, enhancing the network's ability to express the complex morphology of the boundaries of various substructures of the heart through cascaded double-layer KAN and convolution operations. To ensure the smooth execution of residual connections, the network internally matches the number of channels and spatial dimensions of the feature maps layer by layer to ensure the alignment of the output dimensions at each level.

[0033] Furthermore, a bypass is branched off from the output of the first linear layer. This bypass directly extracts the features mapped by the first layer in the first processing path. The bypass is then gated by ReLU and the residual is added to the final result of the second processing path to obtain the output features of the first sub-module. The ReLU activation function provides non-linear gating for the bypass, enabling the network to adaptively decide whether to use this intermediate feature to correct the final output. This residual addition operation effectively alleviates the vanishing gradient phenomenon common in deep networks.

[0034] For example, such as Figure 3 As shown, the first submodule is a TKAN block. The TKAN block contains a first standard convolution, a first KAN layer, a first depthwise convolution, and a first linear layer, which are sequentially connected. The output of this sequential path continues to pass through a second standard convolution, a second KAN layer, and a second depthwise convolution. At the same time, a bypass is branched off from the output of the first linear layer. The bypass is gated by ReLU and then added to the residual of the output of the second depthwise convolution.

[0035] This embodiment achieves multi-scale nonlinear mapping of input features through a cascaded structure of two-layer standard convolution, KAN layers, and depthwise convolution. The introduction of the KAN layer enables the network to adaptively learn complex spline function shapes, overcoming the limitations of traditional activation functions with fixed forms. Furthermore, the residuals of the bypass features after ReLU gating are summed, which not only preserves the detailed information of shallow features but also provides stable gradient flow for deep networks through skip connections.

[0036] In some preferred embodiments, the second submodule of the feature fusion module includes multiple DVSSLayers. The processing flow of a single DVSSLayer includes: linearly mapping the input features to obtain mapped features; decomposing the mapped features into low-frequency subbands and multiple high-frequency subbands through discrete wavelet transform; performing sequence modeling on the low-frequency subbands and each of the high-frequency subbands after nonlinear activation, and inputting them into a VMamba-based state space unit to obtain sequence features corresponding to each subband; performing gated fusion on each of the sequence features to obtain fused features; restoring the fused features to the original spatial resolution through inverse discrete wavelet transform to obtain restored features; and adding the residuals of the restored features and the input features to obtain the output features of the DVSSLayer.

[0037] In specific implementation, the second submodule contains multiple DVSSLayers. Each DVSSLayer, by integrating discrete wavelet transform and VMamba's state-space modeling capabilities, achieves multi-frequency domain decomposition and global context modeling of the three-dimensional heart structure. The specific implementation method is as follows.

[0038] For a single DVSSLayer, the input features are first linearly mapped to obtain the mapped features. The linear mapping adjusts the number of input feature channels to a dimension suitable for discrete wavelet transform decomposition.

[0039] Furthermore, the mapped features are decomposed into low-frequency sub-bands and multiple high-frequency sub-bands using discrete wavelet transform. The low-frequency sub-bands represent the coarse outline and overall morphological distribution information of the heart structure, while the high-frequency sub-bands correspond to the fine structural textures of valves, blood vessel walls, and ventricular boundaries. Through discrete wavelet decomposition, the originally overlapping global morphological information and local high-frequency noise are isolated into independent sub-bands.

[0040] Furthermore, the low-frequency subbands and each of the high-frequency subbands are nonlinearly activated and then input into a VMamba-based state-space unit for sequence modeling to obtain the sequence features corresponding to each subband. VMamba, based on a state-space model, captures long-range dependencies through parallel computational state-space modeling. In the sequence modeling stage, the three-dimensional feature map corresponding to each subband is treated as a one-dimensional sequence unfolded along the spatial dimension and input into the VMamba module, thereby establishing global associations between voxels at different frequency scales.

[0041] Further, the sequence features are gated and fused to obtain fused features. The gated fusion employs a learnable weighting mechanism to adaptively adjust the contribution ratio of different sub-band features in subsequent reconstruction. The fused features are restored to their original spatial resolution using inverse discrete wavelet transform (IDR), yielding restored features. IDR transforms the frequency domain representation back to the spatial domain, maintaining the same spatial dimension as the input features. The restored features are then residually added to the input features to obtain the output features of the DVSSLayer. This residual connection ensures that the integrity of the underlying basic features is not lost due to multi-layer nonlinear processing. For example, the MTKAN module consists of multiple MTKAN blocks, such as... Figure 4 As shown, each MTKAN block consists of a TKAN block, a convolutional layer, a DVSBlock, and another convolutional layer connected in series. The DVSBlock is composed of multiple DVSSLayers stacked together.

[0042] This embodiment decomposes the 3D feature map into multi-frequency sub-bands using discrete wavelet transform, and then performs sequence modeling using VMamba state space units, achieving multi-scale, multi-frequency feature decomposition and reconstruction. Furthermore, global modeling of the low-frequency sub-bands ensures the consistency of the macroscopic morphology of the heart structure, while independent processing of the high-frequency sub-bands significantly improves boundary sharpness and texture detail resolution. Furthermore, gated fusion and inverse discrete wavelet transform ensure accurate synthesis of multi-frequency information and precise restoration of spatial resolution.

[0043] S3, the shape autoencoder is pre-trained using the labeled cardiac CT image dataset to obtain a frozen shape prior projector.

[0044] In specific implementation, the shape autoencoder is pre-trained using the labeled cardiac CT image dataset to obtain a frozen shape prior projector. During the pre-training phase, a real 3D segmentation mask from the labeled dataset is input into the shape autoencoder, and the joint shape distribution of five cardiac structures is learned through an encoder-decoder structure. Further, after training, the network parameters of the shape autoencoder are fixed, transforming it into a decoupled module for projecting prediction results, i.e., the frozen shape prior projector. This projector can map any input segmentation prediction result to a manifold space containing only anatomically reasonable shapes, thereby normalizing and constraining the predicted structure.

[0045] In some preferred embodiments, pre-training the shape autoencoder using the labeled cardiac CT image dataset includes: training a three-dimensional region of interest (ROI) localization network using the labeled cardiac CT image dataset to obtain a trained three-dimensional ROI localization network; inputting the original CT images from the labeled cardiac CT image dataset into the trained three-dimensional ROI localization network, and regressing the center heatmap, offset map, and three-dimensional scale map from the original CT images using the three-dimensional ROI localization network to obtain a whole-heart three-dimensional bounding box; cropping and resampling the original CT images using the whole-heart three-dimensional bounding box to construct a normalized ROI dataset; and training the shape autoencoder based on the labeled samples in the ROI dataset.

[0046] In specific implementation, a 3D region of interest (ROI) localization network is used to achieve high-precision localization and cropping of the heart region. A standardized ROI dataset is constructed before training a shape autoencoder. The specific implementation method is as follows: First, the 3D ROI localization network is trained using the labeled cardiac CT image dataset, resulting in the trained 3D ROI localization network. The 3D ROI localization network is based on a 3D-UNet architecture and outputs three independent branches: a center heatmap for locating the geometric center coordinates of the heart (through network regression of heatmap values ​​based on a Gaussian distribution), an offset map for correcting the relative position of each voxel from the center, and a 3D scale map for predicting the bounding box size of the entire heart in three dimensions. During training, a multi-task loss function is used to jointly supervise the output of the three branches, ensuring that the localization network can quickly grasp the heart localization patterns based on a small amount of labeled data.

[0047] Furthermore, the original CT images from the labeled cardiac CT image dataset are input into the trained 3D region of interest (ROI) localization network. The ROI localization network then regresses the center heatmap, offset map, and 3D scale map from the original CT images to obtain the whole-heart 3D bounding box. The bounding box is the smallest cuboid region that can tightly enclose the five substructures of the entire heart.

[0048] Furthermore, the original CT image is cropped and resampled using the whole-heart 3D bounding box to construct a standardized region of interest (ROI) dataset. The cropping operation extracts local regions from the image containing only the heart structure, eliminating interference from irrelevant anatomical structures such as the lungs, ribs, and spine on shape learning. The resampling operation uniformly adjusts the cropped image to a preset fixed spatial resolution (e.g., 128x128x96 voxels), eliminating spatial scale variability caused by different scanning devices and different subjects. Finally, the shape autoencoder is trained based on labeled samples in the ROI dataset. The training objective of the shape autoencoder is to construct a low-dimensional anatomical shape manifold in the latent space, enabling the input segmentation mask to be reconstructed into an anatomically plausible shape after encoding and decoding. Since the input has had irrelevant background noise removed, the shape autoencoder can more accurately learn the joint distribution patterns of heart substructures, thus providing more accurate shape prior constraints in the subsequent semi-supervised training stage.

[0049] For example, such as Figure 5 As shown, a three-dimensional region of interest (ROI) localization network is used to locate labeled cardiac CT images, regress the center heatmap, offset map, and three-dimensional scale map to obtain the whole-heart three-dimensional bounding box, and then the original image is cropped and resampled based on the bounding box to construct a normalized ROI dataset; finally, the labeled segmentation mask in the ROI dataset is input into the shape autoencoder for training.

[0050] This embodiment outputs a center heatmap, offset map, and scale using a three-dimensional region of interest localization network. Figure 3 The preprocessing process, consisting of several branches, achieves coarse segmentation and high-precision localization of the cardiac region. Furthermore, this preprocessing step narrows down the spatial scope to be processed, eliminating interference from irrelevant anatomical structures. Additionally, normalized pruning and resampling operations ensure that different samples have a unified geometric standard, providing a high-quality data foundation for the shape autoencoder to learn the joint shape distribution of the heart.

[0051] S4, input the unlabeled samples from the unlabeled cardiac CT image dataset into the teacher network to obtain the teacher network prediction results.

[0052] In practice, a semi-supervised learning phase is initiated. Unlabeled samples from the unlabeled cardiac CT image dataset are input into the teacher network to obtain the teacher network prediction results. These prediction results represent the probability distribution of each voxel in the cardiac region of the unlabeled sample belonging to a different substructure.

[0053] S5, input the teacher network prediction result into the frozen shape prior projector to obtain the shape correction result, output the shape correction result as the teacher projection, and construct a confidence map based on the difference between the teacher network prediction result and the shape correction result.

[0054] In practice, the teacher network prediction results are input into the frozen shape prior projector to obtain shape correction results. These correction results are adjusted versions of the teacher network prediction results after being constrained by the anatomical shape prior space. Since the frozen shape prior projector is trained on a healthy anatomical shape distribution, if the teacher prediction shows structural fractures, abnormal adhesions, or perforations, the frozen shape prior projector will forcibly pull it back to a shape conforming to anatomical principles. Furthermore, the shape correction results are used as the teacher projection output. Simultaneously, a confidence map is constructed based on the difference between the teacher network prediction results and the shape correction results. This difference reflects the degree to which the teacher prediction deviates from anatomical rationality.

[0055] In some preferred embodiments, constructing a confidence map based on the difference between the teacher network prediction results and the shape correction results includes: calculating the system-by-system absolute difference between the teacher network prediction results and the shape correction results; normalizing the system-by-system absolute difference to a preset interval to obtain the confidence map, which is used to indicate the reliability of the prediction results at the corresponding location.

[0056] In practice, the confidence graph is the key intermediate data for calculating the adaptive weighted consistency loss, and the specific implementation method is as follows: First, the system-by-system absolute difference between the teacher network prediction result and the shape correction result is calculated. The system-by-system absolute difference is the L1 norm distance between the teacher network prediction result and the shape autoencoder correction result at each 3D voxel location. This operation requires traversing every spatial coordinate point of the entire 3D probability map and obtaining the difference between the probability values ​​of the two at that coordinate.

[0057] Further, the system-by-system absolute differences are normalized to a preset interval to obtain the confidence map. The normalization is performed as follows: calculate the maximum value of the system-by-system absolute differences for all voxel positions in the current sample's three-dimensional space, then divide the absolute difference for each position by this maximum value to obtain a deviation score. Subtract this deviation score from 1 to obtain the final confidence value. This formula transforms the original dimensionless difference values ​​into a standardized confidence score between 0 and 1. The confidence map indicates the reliability of the prediction results at the corresponding position. A value closer to 1 indicates a high degree of consistency between the teacher network's prediction and the anatomical shape correction results at that position; a value closer to 0 indicates a significant deviation, suggesting potential anatomically unreasonable predictions such as structural fractures or abnormal adhesions in that area.

[0058] This embodiment constructs a confidence map representing the reliability of pseudo-labels at each voxel location through system-by-system L1 distance and its normalization operation. This map can quantitatively identify regions in the teacher network prediction results that deviate from the anatomical shape manifold, providing an integrated confidence assessment based on prior knowledge for weighted loss calculation. Furthermore, by shielding the interference of low-confidence regions, the semi-supervised segmentation model can avoid being misled by unreasonable pseudo-labels during training.

[0059] S6, apply a preset strong enhancement operation to the unlabeled samples to obtain strongly enhanced samples; input the strongly enhanced samples into the student network to obtain the student network prediction results.

[0060] In specific implementation, a preset strong enhancement operation is applied to the unlabeled samples to obtain strongly enhanced samples. The strong enhancement operation includes composite data enhancement such as random scaling, rotation, translation, elastic deformation, and gamma contrast adjustment of the original CT image; this invention is not specifically limited to these operations. The strongly enhanced samples are then input into the student network to obtain the student network prediction results.

[0061] S7. Input the student network prediction results into the frozen shape prior projector to obtain the student projection output.

[0062] In practice, the student network prediction results are input into the frozen shape prior projector to obtain the student projection output.

[0063] S8. Based on the confidence map, calculate the confidence-weighted consistency loss between the student's projected output and the teacher's projected output, and optimize the parameters of the semi-supervised segmentation model based on the total loss. The total loss includes the confidence-weighted consistency loss, the basic supervision loss determined based on the prediction results of labeled data, and the shape prior loss determined based on shape prior constraints.

[0064] In practice, based on the confidence graph, the confidence-weighted consistency loss between the student's projected output and the teacher's projected output is calculated. The formula for calculating the confidence-weighted consistency loss is as follows: Where i represents the position of all three-dimensional voxels in the image; C i This represents the confidence weight at the i-th position, with its value normalized to the interval [0,1]. A larger value indicates a higher confidence level at that position; D i B represents the value of the prediction result of the student network at the i-th voxel position on the strongly augmented sample; i This represents the value of the shape correction result obtained after the teacher network prediction result is projected by the frozen shape prior projector at the i-th voxel position; The underlying difference metric loss function is represented in this embodiment as a combination of Dice loss and cross-entropy loss. Dice loss measures the overlap between the predicted mask and the reference mask, and is naturally insensitive to class imbalance problems common in medical images; cross-entropy loss, based on voxel-level probability distribution differences, provides a smooth gradient optimization signal. Confidence graph C i This plays a key adjusting role in the weighting calculation: for positions with high confidence, C... i When the value approaches 1, the consistency loss at that position is included in the total loss with a high weight; for positions with low confidence, C... i When the value approaches 0, the loss contribution at that position is actively suppressed.

[0065] Furthermore, in this embodiment, the parameters of the semi-supervised segmentation model are optimized based on the total loss. The total loss L... total Including the confidence-weighted consistency loss L confidence The basic supervised loss L is determined based on the prediction results of labeled data. label and the shape prior loss L determined based on shape prior constraints shape The total loss calculation formula is: L = L label + L shape + L confidence Among them, the basic monitoring loss L label The loss is a combination of Dice and cross-entropy based on the prediction results of the student network on labeled data and the corresponding ground truth labels; shape prior loss L... shape This refers to the difference between the labeled predictions of the student network and the labeled projections obtained through the shape projector. The synergistic optimization of these three elements ensures that the model fully utilizes unlabeled data while remaining supervised by the constraint of anatomical rationality.

[0066] This embodiment constructs a collaborative optimization framework that balances the reliability of labeled samples and the utilization rate of unlabeled sample information by embedding a frozen shape prior projector into the mean teacher framework and combining it with an adaptive weighting mechanism based on the confidence map. Furthermore, the confidence map utilizes the per-system bias between the teacher network prediction results and their shape correction results to assign reliable weight values ​​to the voxel positions of each unlabeled sample. Furthermore, by combining the basic supervision loss and the shape prior loss, the model can obtain high-precision segmentation boundaries under conditions of scarce labeled data, while avoiding anatomical structural deformations. In some preferred embodiments, the step of calculating the confidence-weighted consistency loss between the student projection output and the teacher projection output based on the confidence map includes: calculating the Dice loss and cross-entropy loss between the student projection output and the teacher projection output; and weighting and summing the Dice loss and the cross-entropy loss with the weights at the corresponding positions in the confidence map to obtain the confidence-weighted consistency loss.

[0067] In practice, firstly, the Dice loss and cross-entropy loss between the student's projected output and the teacher's projected output are calculated. The Dice loss is calculated by measuring the regional overlap between the predicted mask and the reference mask, and it has good robustness to the common background-lesion class imbalance problem in cardiac CT segmentation. Furthermore, the cross-entropy loss is calculated based on the difference in probability distribution at the voxel level, providing a smooth gradient signal for network optimization. The two losses are linearly superimposed to construct a comprehensive basic loss term that balances overall regional consistency and pixel-level classification accuracy.

[0068] Furthermore, the Dice loss and the cross-entropy loss are weighted and summed with the weights at the corresponding positions on the confidence map to obtain the confidence-weighted consistency loss. Specifically, when calculating the cross-entropy loss, the loss value of each voxel is multiplied by the confidence weight at that position; when calculating the Dice loss, the confidence map is used for weighted counting, so that high-confidence regions occupy a higher proportion in the loss calculation. By weighting the loss of each voxel, the model's learning focus is concentrated on regions with reasonable anatomical structures and reliable prediction results, while ignoring noise interference that may be caused by blurred edges or structurally abnormal regions.

[0069] This embodiment achieves differentiated learning of regions with different confidence levels in unlabeled data by incorporating the confidence map into the weighted summation process of Dice loss and cross-entropy loss. Furthermore, the loss contribution of high-confidence regions is given higher weight, driving the model to converge toward the correct anatomical structure; furthermore, the loss contribution of low-confidence regions is actively suppressed, effectively avoiding interference from low-quality pseudo-labels on model parameter updates.

[0070] In some preferred embodiments, optimizing the parameters of the semi-supervised segmentation model based on the total loss includes: determining the total loss; and updating the parameters of the student network using the total loss; determining the total loss includes: obtaining the labeled prediction results of the student network on the labeled cardiac CT image dataset, and determining the basic supervised loss based on the labeled prediction results and corresponding ground truth annotations; inputting the labeled prediction results into the frozen shape prior projector to obtain labeled projection results, and determining the shape prior loss based on the difference between the labeled prediction results and the labeled projection results; obtaining the confidence-weighted consistency loss; and using the sum of the basic supervised loss, the shape prior loss, and the confidence-weighted consistency loss as the total loss; updating the parameters of the student network using the total loss includes: updating the parameters of the student network using backpropagation based on the total loss.

[0071] In specific implementation, this embodiment constructs the basic supervised loss, shape prior loss, and confidence-weighted consistency loss step by step, and uses the joint loss obtained by summing the three to complete a single iterative update of the student network parameters. The specific implementation method is as follows: First, determine the total loss. Determining the total loss includes the following three parallel or serial acquisition processes.

[0072] The first step involves obtaining the labeled prediction results of the student network on the labeled cardiac CT image dataset, and determining the basic supervision loss based on the labeled prediction results and the corresponding ground truth annotations. Specifically, the labeled samples are input into the student network to perform forward propagation computation, resulting in a 3D probability prediction map for each voxel in the labeled samples belonging to different cardiac substructures. Then, the region overlap between the 3D probability prediction map and the physician-annotated ground truth 3D mask is calculated and used as the Dice loss; simultaneously, the probability distribution difference of each voxel point between the 3D probability prediction map and the ground truth 3D mask is calculated and used as the cross-entropy loss. The Dice loss and the cross-entropy loss are linearly weighted and summed to obtain the basic supervision loss. Since the Dice loss optimizes the overall segmentation accuracy by measuring the overlap area between the predicted region and the ground truth region, while the cross-entropy loss optimizes boundary details by measuring pixel-by-pixel classification error, the combination of the two gives the basic supervision loss both global capture capability and local edge sensitivity.

[0073] The second step involves inputting the labeled prediction results into the frozen shape prior projector to obtain labeled projection results. Based on the difference between the labeled prediction results and the labeled projection results, the shape prior loss is determined. Specifically, the frozen shape prior projector first encodes the labeled prediction results, compressing them into a low-dimensional anatomical shape latent space. Then, a decoder reconstructs this latent vector into a shape correction map with the same dimension as the input space. The voxel-by-voxel error (e.g., mean squared error or absolute error) between the labeled prediction results and the shape correction map is calculated and used as the shape prior loss. The shape prior loss numerically quantifies the degree to which the student network's output deviates from the anatomically reasonable shape manifold space. By backpropagating this value to the student network, the update direction of the network parameters is forced to always point towards the anatomically reasonable morphological region.

[0074] The third step is to obtain the confidence-weighted consistency loss. As described in Example 6, this loss provides consistency constraints on unlabeled data, and its weights depend on a confidence map constructed from the differences between the teacher network prediction results and the shape correction results.

[0075] Furthermore, the sum of the basic supervision loss, the shape prior loss, and the confidence-weighted consistency loss is taken as the total loss. That is, the total loss L = L label + L shape + L confidence The three sub-losses complement each other: L label Provides basic target identification capabilities; L shape To prevent anatomical errors in network prediction; L confidence This drives the network to mine the potential semantic information in unlabeled data.

[0076] Furthermore, the parameters of the student network are updated using the total loss. Specifically, the parameters of the student network are updated using backpropagation with the total loss. The calculated total loss is backpropagated, and the gradient values ​​of the learnable parameters of each layer are calculated according to the chain rule. Then, an optimizer (such as Adam or SGD) is used to update all weights and biases in the student network based on the gradient values.

[0077] This embodiment achieves collaborative multi-constraint optimization of the segmentation model by systematically constructing and aggregating three types of loss functions. The basic supervised loss ensures fundamental segmentation accuracy, the shape prior loss strictly limits the anatomical rationality of the output structure, and the confidence-weighted consistency loss avoids the introduction of label noise into unlabeled data. Furthermore, this three-level loss combination enables the model to achieve high-quality segmentation results comparable to fully supervised training using only a small amount of labeled data and a large amount of unlabeled data, while simultaneously ensuring the morphological robustness of the 3D segmentation results.

[0078] In some preferred embodiments, the method further includes: updating the parameters of the teacher network, including: obtaining the updated parameters of the student network in the current iteration round; performing an exponential moving average operation based on the updated parameters and the previous round parameters of the teacher network in the previous iteration round to obtain the current round parameters of the teacher network in the current iteration round; and using the current round parameters as the parameters for updating the teacher network in this round.

[0079] In specific implementation, the parameters of the teacher network are dynamically updated according to the definition of the mean teacher framework through an exponential moving average mechanism. The specific implementation method is as follows: First, the updated parameters of the student network in the current iteration are obtained. After the student network completes a backpropagation update based on the total loss, its parameters undergo a slight change. In this embodiment, the updated parameters are used as the input source for the moving average.

[0080] Furthermore, based on the updated parameters and the previous round parameters of the teacher network in the previous iteration, an exponential moving average operation is performed to obtain the current round parameters of the teacher network in the current iteration. The exponential moving average operation smoothly weights and merges the current round student network parameters with the historical teacher network parameters. Specifically, the fusion mechanism uses a preset momentum coefficient (preferably 0.99) to weight the historical parameters, and then adds the current student network update amount with a smaller weight (1 minus the momentum coefficient, i.e., 0.01). This operation ensures that the teacher network parameters smoothly follow the improvement pace of the student network during iteration, without being drastically disturbed by stochastic gradient noise in a single iteration, thus effectively suppressing parameter oscillations.

[0081] Furthermore, the parameters of the current round are used as the parameters for the teacher network's update in this round. In subsequent inference processes within this round, the teacher network will use these parameters to predict unlabeled data. Since the teacher network parameters are obtained through long-term smoothing, their output is more stable than the short-term fluctuations in the student network's output after a single iteration, thus providing higher-quality pseudo-label guidance for semi-supervised learning.

[0082] This embodiment updates the parameters of the teacher network using an exponential moving average mechanism, achieving a slow-following strategy for knowledge transfer from the student network. This mechanism effectively suppresses parameter jitter caused by stochastic gradient descent in a single iteration, ensuring that the pseudo-labels output by the teacher network remain smooth and of high quality throughout the training process, thus playing a stable guiding role within the mean-based teacher framework. Furthermore, this dynamic parameter update strategy provides an important guarantee for the successful training of the semi-supervised segmentation model with limited labeled data, significantly improving the stability of model convergence and the accuracy of the final prediction.

[0083] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0084] Please see Figure 7 , Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0085] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0086] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it causes the processor 502 to perform a semi-supervised segmentation method for cardiac CT images.

[0087] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0088] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a semi-supervised segmentation method for cardiac CT images.

[0089] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0090] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of a semi-supervised segmentation method for cardiac CT images provided in any of the above method embodiments.

[0091] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0092] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0093] Therefore, the present invention also provides a storage medium. This storage medium may be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of the semi-supervised segmentation method for cardiac CT images in any of the above method embodiments.

[0094] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0095] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0096] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0097] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0098] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0099] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, this invention is also intended to include these modifications and variations as long as they fall within the scope of the claims of this invention and their equivalents.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A semi-supervised segmentation method for cardiac CT images, characterized in that, include: Obtain both labeled cardiac CT image datasets and unlabeled cardiac CT image datasets; A semi-supervised segmentation model is constructed, which includes a shape autoencoder, a teacher network, and a student network. The teacher network and the student network share the same segmentation backbone network. The segmentation backbone network is based on an encoder-decoder framework and a feature fusion module is configured at the bottleneck layer. The feature fusion module includes a first sub-module for nonlinear mapping and a second sub-module for long-range dependency capture. The shape autoencoder is pre-trained using the labeled cardiac CT image dataset to obtain a frozen shape prior projector; The unlabeled samples from the unlabeled cardiac CT image dataset are input into the teacher network to obtain the teacher network prediction results; The teacher network prediction results are input into the frozen shape prior projector to obtain shape correction results. The shape correction results are used as teacher projection outputs, and a confidence map is constructed based on the difference between the teacher network prediction results and the shape correction results. A preset strong enhancement operation is applied to the unlabeled samples to obtain strongly enhanced samples; The strongly enhanced samples are input into the student network to obtain the student network prediction results; The student network prediction results are input into the frozen shape prior projector to obtain the student projection output; Based on the confidence map, the confidence-weighted consistency loss between the student's projected output and the teacher's projected output is calculated, and the parameters of the semi-supervised segmentation model are optimized based on the total loss. The total loss includes the confidence-weighted consistency loss, the basic supervision loss determined based on the prediction results of labeled data, and the shape prior loss determined based on shape prior constraints.

2. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The processing flow of the first sub-module of the feature fusion module includes: The input features are sequentially passed through the first standard convolution, the first KAN layer, the first depthwise convolution, and the first linear layer to form the first processing path; The output of the first processing path is sequentially passed through a second standard convolution, a second KAN layer, and a second depthwise convolution to form a second processing path. A bypass is branched off from the output of the first linear layer; The bypass is gated by ReLU and then added to the residual of the end result of the second processing path to obtain the output feature of the first sub-module.

3. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The second submodule of the feature fusion module contains multiple DVSSLayers, and the processing flow of a single DVSSLayer includes: The input features are linearly mapped to obtain the mapped features; The mapping features are decomposed into low-frequency sub-bands and multiple high-frequency sub-bands by discrete wavelet transform; After nonlinear activation, the low-frequency sub-band and each of the high-frequency sub-bands are input into a VMamba-based state space unit for sequence modeling to obtain the sequence features corresponding to each sub-band. The sequence features are then gated and fused to obtain the fused features; The fused features are restored to their original spatial resolution using inverse discrete wavelet transform to obtain the restored features; The residuals of the recovered features and the input features are added together to obtain the output features of DVSSLayer.

4. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The pre-training of the shape autoencoder using the labeled cardiac CT image dataset includes: A three-dimensional region of interest localization network was trained using the labeled cardiac CT image dataset to obtain the trained three-dimensional region of interest localization network; The original CT images from the labeled cardiac CT image dataset are input into the trained 3D region of interest localization network. The 3D region of interest localization network regresses the center heatmap, offset map and 3D scale map from the original CT images to obtain the whole heart 3D bounding box. The original CT image is cropped and resampled using the whole-center 3D bounding box to construct a normalized region of interest dataset; The shape autoencoder is trained based on labeled samples in the region of interest dataset.

5. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The construction of a confidence map based on the difference between the teacher network prediction results and the shape correction results includes: Calculate the system-by-system absolute difference between the teacher network prediction results and the shape correction results; The absolute difference of each system is normalized to a preset interval to obtain the confidence map, which is used to indicate the reliability of the prediction result at the corresponding location.

6. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The step of calculating the confidence-weighted consistency loss between the student's projected output and the teacher's projected output based on the confidence map includes: Calculate the Dice loss and cross-entropy loss between the student's projected output and the teacher's projected output; The confidence-weighted consistency loss is obtained by weighting and summing the Dice loss and the cross-entropy loss with the weights at the corresponding positions in the confidence graph.

7. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The optimization of the parameters of the semi-supervised segmentation model based on the total loss includes: determining the total loss; and updating the parameters of the student network using the total loss. Determining the total loss includes: Obtain the labeled prediction results of the student network on the labeled cardiac CT image dataset, and determine the basic supervision loss based on the labeled prediction results and the corresponding real annotation information; The labeled prediction result is input into the frozen shape prior projector to obtain the labeled projection result, and the shape prior loss is determined based on the difference between the labeled prediction result and the labeled projection result. Obtain the confidence-weighted consistency loss; The sum of the basic supervision loss, the shape prior loss, and the confidence-weighted consistency loss is taken as the total loss; The step of updating the parameters of the student network using the total loss includes: updating the parameters of the student network using backpropagation with the total loss.

8. The semi-supervised segmentation method for cardiac CT images according to claim 1, characterized in that, The method further includes: updating the parameters of the teacher network, wherein updating the parameters of the teacher network includes: Obtain the updated parameters of the student network in the current iteration round; Based on the updated parameters and the previous round parameters of the teacher network in the previous iteration round, an exponential moving average operation is performed to obtain the current round parameters of the teacher network in the current iteration round. The current round parameters are used as the parameters for the teacher network update in this round.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images