Total heart CT image segmentation method and system coping with contrast agent induction domain offset

Abstract

The invention provides a full-heart CT image segmentation method and system for coping with contrast agent induction domain offset. The method comprises the following steps: S1, acquiring a to-be-processed heart CT angiography three-dimensional image and preprocessing the to-be-processed heart CT angiography three-dimensional image; s2, inputting the preprocessed image into a coding path of a segmentation model to extract an intermediate feature, executing a space consistency compensation operation on the intermediate feature, and inputting the feature after space consistency compensation into a decoding path of the segmentation model to obtain an initial segmentation prediction result and a corresponding deep feature; s3, introducing an uncertainty modeling mechanism based on an evidence theory to construct a reliability mask for an initial segmentation prediction result; and S4, under the constraint of a reliability mask, carrying out sample-level online calibration by using the deep features output in the step S2 to obtain a sample-level calibration prediction result, and carrying out weighted fusion on the sample-level calibration prediction result and the initial segmentation prediction result to obtain a final segmentation result. According to the method, stable segmentation of the non-contrast mode can be realized under the condition of no target domain labeling.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, specifically relating to a method and system for segmenting whole-cardiac CT images to address contrast agent-induced domain shift. Background Technology

[0002] Currently, whole-heart CT image segmentation technology targets CTA (CT angiography) images, employing 3D fully convolutional networks and their variants (such as 3D U-Net, nnU-Net, and improved network architectures) to segment various anatomical structures, including the left atrium, right atrium, left ventricle, right ventricle, myocardium, ascending aorta, and pulmonary artery, through end-to-end learning. To improve the cross-center generalization ability of segmentation methods, published methods typically improve upon them in the following ways: 1. Data / style layer enhancement: Simulating domain differences through normalization, intensity perturbation, style transfer, etc.; 2. Geometric topological constraints: Introducing shape priors, topological consistency, or geometric similarity learning to alleviate instability caused by the complexity and blurred boundaries of the heart structure; 3. Generative modeling: Utilizing a diffusion generation framework to jointly model the image and mask distribution to adapt to different appearances; 4. Test-time adaptation: During the inference phase, using unlabeled test data for entropy minimization, self-supervised tasks, or batch normalization (BN) statistical calibration to update and adapt to the target domain.

[0003] However, the aforementioned techniques still have significant shortcomings in multi-center CTA scenarios with "contrast agent-induced domain shift": Shortcoming 1: Lack of explicit modeling of complex enhancement patterns caused by differences in "contrast agent injection protocol / pass timing / scanning equipment". In actual cross-center scenarios, significant gray-scale statistical drift and local enhancement heterogeneity can occur in the same anatomical structure, resulting in prediction bias. Style perturbation alone is often insufficient to cover the statistical fluctuations caused by real enhancement dynamics. Shortcoming 2: 3D cardiac segmentation is often limited by GPU memory, resulting in small training batch sizes. Existing feature enhancement and normalization methods that rely on current small-batch statistics are prone to bias and fail to reflect global domain statistics, leading to high difficulty in cross-domain generalization. Shortcoming 3: Local structural boundaries are easily affected by contrast agent heterogeneity, resulting in "blurring and breakage". Methods that rely solely on global style alignment or geometric constraints often struggle to maintain good structural continuity and clear boundary details simultaneously in small vessels and thin-walled structures. Shortcoming 4: Existing test-time adaptive methods typically rely on backpropagation to update model parameters or are based on a single assumption (such as high-confidence prediction or BN calibrability). In medical image segmentation tasks, such methods are prone to instability, high computational cost, or negative transfer, and it is difficult to achieve reliable region adaptation for a single sample.

[0004] In view of this, the present invention proposes a whole-cardiac CT image segmentation method and system to address contrast agent-induced domain shift. Summary of the Invention

[0005] The purpose of this invention is to propose a whole-cardiac CT image segmentation method and system to address contrast agent-induced domain shift. Following the overall technical approach of "generalization during training and adaptive collaboration during testing", it achieves stable segmentation of unseen contrast patterns under conditions where no target domain is labeled.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A whole-cardiac CT image segmentation method to address contrast agent-induced domain shift, specifically including the following steps:

[0008] S1. Acquire the three-dimensional cardiac CT angiography image to be processed and perform preprocessing operations;

[0009] S2. Input the preprocessed image into the encoding path of the segmentation model based on the three-dimensional convolutional neural network, extract the intermediate features that represent the spatial and semantic information of the heart's anatomical structure, perform spatial consistency compensation operation on the intermediate features, and input the spatial consistency compensated features into the decoding path of the segmentation model to obtain the initial segmentation prediction results and the corresponding deep features.

[0010] During the training phase, the segmentation model performs online domain statistical coordination operations on the features after spatial consistency compensation, and uses the features after online domain statistical coordination as the input to the segmentation model's decoding path.

[0011] S3. For the initial segmentation prediction results, an uncertainty modeling mechanism based on evidence theory is introduced to construct a reliability mask;

[0012] S4. Under the constraint of the reliability mask, the deep features output in step S2 are used to perform sample-level online calibration to obtain sample-level calibration prediction results. The sample-level calibration prediction results are then weighted and fused with the initial segmentation prediction results to obtain the final segmentation results.

[0013] Preferably, the preprocessing operations include resampling the image, intensity normalization, and heart region cropping.

[0014] Preferably, the spatial consistency compensation operation performed on intermediate features includes parallel modeling of local detail enhancement branches and global guidance branches, as well as residual backinjection, as follows:

[0015]

[0016]

[0017]

[0018]

[0019] in, Indicates intermediate features, Represents the convolution kernel parameters. Represents a three-dimensional convolution operation. This indicates the instance normalization operation. Represents a non-linear activation function. , , These are the intermediate processing results. Features after spatial consistency compensation

[0020] Preferably, the online domain statistical coordination operation is performed on the features after spatial consistency compensation, specifically as follows:

[0021] Features of the current sample after spatial consistency compensation Calculate the instance mean corresponding to the current sample based on the channel dimension. Standard deviation of the instance The global statistics are maintained online using an exponential moving average method to obtain the global mean and variance. and global variance ;

[0022] Introducing random perturbations to generate feature reparameterization coefficients:

[0023]

[0024] in, For random noise that follows a standard normal distribution; β represents the feature reparameterization bias term introduced during the online domain statistical coordination process, used to adjust the mean position of the features; The feature reparameterization scaling term introduced during online domain statistical coordination is used to adjust the scale amplitude of features. and They represent the corresponding variances. , The statistical fluctuation range obtained by taking the square root;

[0025] Features Performing normalization and affine transformation yields the output features after statistical harmonization in the online domain:

[0026]

[0027] in, This represents the characteristics after online domain statistical reconciliation. A smoothing constant introduced to prevent numerical instability.

[0028] Preferably, the global statistics are maintained online using an exponential moving average method to obtain the global mean and variance. and global variance The details are as follows:

[0029] The instance statistics of the characteristic means are maintained using an exponential moving average method to maintain their first-order statistics. With second-order statistics :

[0030]

[0031]

[0032] in The coefficient of the exponential moving average;

[0033] Calculate the global mean and variance based on the maintained first- and second-order statistics:

[0034]

[0035] The instance statistic of the characteristic standard deviation is maintained using an exponential moving average method to maintain its first-order statistic. With second-order statistics And calculate the global variance:

[0036] .

[0037] Preferably, for the initial segmentation prediction result, a reliability mask is constructed by introducing an uncertainty modeling mechanism based on evidence theory, as follows:

[0038] For the initial segmentation prediction results ,set up Represents any voxel In the The initial predicted logit value for each segmentation category is then defined as the evidence quantity:

[0039]

[0040] in, Voxel representation In the Predict the strength of evidence for each segmentation category. This is an evidence mapping adjustment parameter used to control the degree of numerical scaling when mapping the predicted logit to the evidence quantity; The number of segmentation categories;

[0041] Constructing Dirichlet distribution parameters based on evidence quantity :

[0042]

[0043] Define voxel-level uncertainty measure Among them, a single voxel Uncertainty Represented as:

[0044]

[0045] By combining the predicted category information, a reliability mask is constructed:

[0046]

[0047] in, This represents the category index in the segmentation task; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Indicates the first Reliability masks corresponding to each segmentation category.

[0048] Preferably, under the constraint of the reliability mask, the deep features output in step S2 are used to perform sample-level online calibration to obtain sample-level calibration prediction results, as follows:

[0049] Deep features under reliability mask constraints Perform weighted aggregation to construct the first The dynamic classifier weight vector corresponding to each segmentation category :

[0050]

[0051] in, Indicates the first The dynamic classifier weight vector for each segmentation category; Indicates voxel index; Indicates deep features in the th The feature vector at the location of an individual element; Indicates the first The reliability mask corresponding to the segmentation category is in the first... The value at the position of the individual element;

[0052] Based on the dynamic classifier weight vector, the deep features Classification is performed to obtain sample-level calibration prediction results. :

[0053] .

[0054] A whole-cardiac CT image segmentation system for addressing contrast agent-induced domain shift includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the above-described whole-cardiac CT image segmentation method.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] This invention effectively mitigates the domain shift problem caused by contrast agent differences in multicenter CTA data. It stably improves the continuity of fine structures, the clarity of segmentation boundaries, and cross-center generalization performance without target domain annotation or manual intervention. This method is applicable to the automatic segmentation of multiple anatomical structures throughout the heart, enabling rapid and accurate extraction of anatomical structures such as cardiac chambers, myocardium, and major blood vessels. It provides a reliable three-dimensional quantitative analysis basis for the auxiliary diagnosis, risk assessment, and treatment planning of cardiovascular diseases, and supports standardized and batch applications in clinical and research settings. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0058] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0059] Currently, whole-cardiac CT image segmentation methods primarily target coronary angiography (CTA) images, generally employing an end-to-end segmentation framework based on a 3D fully convolutional neural network to automatically segment various cardiac anatomical structures, including the left atrium, right atrium, left ventricle, right ventricle, myocardium, ascending aorta, and pulmonary artery. To improve the model's generalization ability on multi-center data, existing techniques typically employ data augmentation, feature normalization, geometric or topological constraints, and test-time adaptation. However, in practical multi-center CTA applications, significant differences exist among medical institutions in contrast agent injection protocols, timing of contrast agent passage, and scanning equipment parameters. Consequently, the same anatomical structure often exhibits markedly different grayscale statistical distributions and local enhancement patterns in data from different centers, resulting in the "contrast agent-induced domain shift" problem. Existing techniques still have the following shortcomings in addressing this issue:

[0060] (1) Lack of explicit modeling ability for contrast agent-induced complex enhancement patterns: Existing methods often simplify cross-center differences to style changes or statistical perturbations, which are difficult to cover gray-scale statistical drift and local enhancement unevenness caused by contrast agent kinetic differences, resulting in a significant decrease in the segmentation performance of the model in the absence of contrast agent.

[0061] (2) Statistical estimation is unstable under small batch training conditions: Three-dimensional cardiac CT segmentation tasks are limited by video memory. During training, single-sample or very small batch training is usually used. Existing normalization or feature perturbation methods that rely on the current batch statistics are prone to statistical bias and are difficult to reflect the overall distribution characteristics of cross-center data.

[0062] (3) Local structural boundaries are easily degraded due to uneven contrast agent distribution: In areas such as thin walls of the heart chambers, myocardial boundaries and small blood vessels, uneven distribution of contrast agent can easily lead to blurred boundaries or structural breaks. It is difficult to simultaneously take into account the continuity of local structures and the overall consistency by simply relying on global style alignment or geometric constraints.

[0063] (4) Existing test-time adaptive methods lack stability: Most test-time adaptive methods rely on backpropagation to update model parameters or are based on a single hypothesis, which can easily introduce computational overhead and prediction oscillations during the inference stage, making it difficult to achieve reliable adaptation for a single sample.

[0064] To address the aforementioned problems, this invention proposes a whole-cardiac CT image segmentation method to cope with contrast agent-induced domain shift. Following the overall technical approach of "generalization during training and adaptive collaboration during testing," it achieves stable segmentation of unseen contrast patterns even without target domain annotation. For example... Figure 1 As shown, the method of the present invention includes the following steps:

[0065] S1. Data Acquisition and Preprocessing: Acquire the three-dimensional cardiac CT angiography images to be processed. The image is preprocessed by resampling, intensity normalization, and heart region cropping to obtain network input image blocks. .

[0066] S2, Intermediate Feature Extraction: Extracting the input image... Input is fed into the encoding path of a segmentation model based on a 3D convolutional neural network to extract intermediate feature representations. The intermediate features are used to characterize the spatial and semantic information of the cardiac anatomy.

[0067] S3. Spatial Consistency Compensation: To alleviate the problem of local structural blurring and breakage caused by uneven distribution of contrast agent, intermediate features are compensated for. Perform spatial consistency compensation operations, which include parallel modeling and residual backinjection of local detail enhancement branches and global guidance branches, satisfying the following relationship:

[0068]

[0069]

[0070]

[0071]

[0072] in, Represents the convolution kernel parameters, " indicates a three-dimensional convolution operation. This indicates the instance normalization operation. Represents a non-linear activation function. , , These are intermediate processing results. Through the above spatial consistency compensation calculation, the structural continuity of regions such as the thin walls of the cardiac chambers, myocardial boundaries, and small blood vessels is enhanced, resulting in compensated features. .

[0073] S4. Online Domain Statistical Coordination: During the training phase, to address the inconsistency in grayscale statistical distribution caused by differences in multi-center imaging modalities, the features obtained after spatial consistency compensation in step S3 are re-coordinated. Perform online domain statistics coordination operations.

[0074] First, calculate the instance mean of the current sample based on the channel dimension for the aforementioned features. Standard deviation of the instance The global statistics are maintained online using an exponential moving average method.

[0075]

[0076] Calculate the global mean and variance based on the maintained first- and second-order statistics:

[0077]

[0078] Similarly, the instance statistic for the characteristic standard deviation Its first-order statistic is maintained using an exponential moving average method. With second-order statistics And calculate its global variance:

[0079]

[0080] Based on this, random perturbation is introduced to generate feature reparameterization coefficients:

[0081]

[0082] in, This represents random noise that follows a standard normal distribution. β represents the feature reparameterization bias term introduced during the online domain statistical coordination process, used to adjust the mean position of the features; The feature reparameterization scaling term introduced during online domain statistical coordination is used to adjust the scale amplitude of features. and They represent the corresponding variances. , The statistical fluctuation range is obtained by taking the square root.

[0083] Finally, regarding the features Performing normalization and affine transformation yields the output features after statistical harmonization in the online domain:

[0084]

[0085] in, This indicates the output feature of step S4. This indicates the output feature of step S3; These represent the instance mean and standard deviation of the current sample, respectively. A smoothing constant is introduced to prevent numerical instability; This is the coefficient of the exponential moving average.

[0086] The output characteristics obtained through the online domain statistical coordination operation are: The These features serve as input to the subsequent segmentation model's decoding path and are used to generate initial segmentation prediction results.

[0087] S5. Initial Segmentation Prediction Generation: This involves generating features that have undergone spatial consistency compensation and online domain statistical coordination. Input to the segmentation model's decoding path to obtain the initial segmentation prediction. and the corresponding deep feature representation .

[0088] S6. Uncertainty Modeling and Reliable Region Selection: The initial segmentation prediction results obtained in step S5 are then processed... An uncertainty modeling mechanism based on evidence theory (a well-known method) is introduced to quantify the reliability of the prediction results for each voxel. For the initial segmentation prediction results... ,set up Represents any voxel In the The predicted logit value for each segmentation category is then defined as the evidence quantity:

[0089]

[0090] Based on the amount of evidence, construct the Dirichlet distribution parameters:

[0091]

[0092] Further define the voxel-level uncertainty measure Among them, a single voxel Uncertainty Represented as:

[0093]

[0094] in, The number of segmentation categories; Voxel representation In the Predict the strength of evidence for each segmentation category; This is an evidence mapping adjustment parameter used to control the degree of numerical scaling when mapping the predicted logit to the evidence quantity.

[0095] Further combining predicted category information, a reliability mask is constructed:

[0096]

[0097] in, This represents the category index in the segmentation task. The total number of categories to be divided; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This represents the voxel-level uncertainty measure calculated in step S6; Indicates the first The reliability mask corresponding to the class is used to characterize the sample that simultaneously satisfies "predicted as the th". A set of voxels that is "classified" and has "low prediction uncertainty".

[0098] S7. Sample-level Online Calibration and Result Fusion: Under the reliability mask constraint obtained in step S6, sample-level online calibration is performed on the current test sample using the deep features output in step S5. Specifically, the deep features are weighted and converged under the reliability mask constraint to construct the first... The dynamic classifier weight vector corresponding to each class is calculated as follows:

[0099]

[0100] in, Indicates the first The dynamic classifier weight vector of the class; Indicates voxel index; The deep feature indicates that the depth feature is in the first... The feature vector at the location of an individual element; This indicates the first step constructed in step S6. The value of the reliability mask at the corresponding voxel position;

[0101] Based on the dynamic classifier weight vector, the deep features Classification is performed to obtain sample-level calibration prediction results:

[0102]

[0103] Finally, the sample-level calibration prediction results are weighted and fused with the initial segmentation prediction results to obtain the final segmentation result:

[0104]

[0105] in, The fusion weight parameter is used to balance the proportion of the initial prediction results and the sample-level calibration prediction results in the final segmentation result.

[0106] The sample-level online calibration process does not involve backpropagation updates of the segmentation model parameters.

[0107] This invention also proposes a whole-cardiac CT image segmentation system to address contrast agent-induced domain shift, comprising a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the aforementioned whole-cardiac CT image segmentation method. In use, this invention only requires inputting the three-dimensional cardiac CTA image data to be processed; the system can automatically complete image preprocessing and feature analysis, and output segmentation results of the entire cardiac anatomy. Users can use these segmentation results for three-dimensional reconstruction and subsequent quantitative analysis or clinical auxiliary evaluation.

[0108] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for segmenting whole-cardiac CT images to address contrast agent-induced domain shift, characterized in that, Specifically, the steps include the following: S1. Acquire the three-dimensional cardiac CT angiography image to be processed and perform preprocessing operations; S2. Input the preprocessed image into the encoding path of the segmentation model based on the three-dimensional convolutional neural network, extract the intermediate features that represent the spatial and semantic information of the heart's anatomical structure, perform spatial consistency compensation operation on the intermediate features, and input the spatial consistency compensated features into the decoding path of the segmentation model to obtain the initial segmentation prediction results and the corresponding deep features. During the training phase, the segmentation model performs online domain statistical coordination operations on the features after spatial consistency compensation, and uses the features after online domain statistical coordination as the input to the segmentation model's decoding path. S3. For the initial segmentation prediction results, an uncertainty modeling mechanism based on evidence theory is introduced to construct a reliability mask; S4. Under the constraint of the reliability mask, the deep features output in step S2 are used to perform sample-level online calibration to obtain sample-level calibration prediction results. The sample-level calibration prediction results are then weighted and fused with the initial segmentation prediction results to obtain the final segmentation results.

2. The whole-cardiac CT image segmentation method for addressing contrast agent-induced domain shift according to claim 1, characterized in that, The preprocessing operations include image resampling, intensity normalization, and heart region cropping.

3. The whole-cardiac CT image segmentation method for addressing contrast agent-induced domain shift according to claim 1, characterized in that, The spatial consistency compensation operation performed on intermediate features includes parallel modeling of local detail enhancement branches and global guidance branches, as well as residual backinjection, as detailed below: in, Indicates intermediate features, Represents the convolution kernel parameters. Represents a three-dimensional convolution operation. This indicates the instance normalization operation. Represents a non-linear activation function. , , These are the intermediate processing results. These are features after spatial consistency compensation.

4. The whole-cardiac CT image segmentation method for addressing contrast agent-induced domain shift according to claim 1, characterized in that, The online domain statistical coordination operation performed on the features after spatial consistency compensation is as follows: Features of the current sample after spatial consistency compensation Calculate the instance mean corresponding to the current sample based on the channel dimension. Standard deviation of the instance The global statistics are maintained online using an exponential moving average method to obtain the global mean and variance. and global variance ; Introducing random perturbations to generate feature reparameterization coefficients: in, For random noise that follows a standard normal distribution; β represents the feature reparameterization bias term introduced during the online domain statistical coordination process, used to adjust the mean position of the features; The feature reparameterization scaling term introduced during online domain statistical coordination is used to adjust the scale amplitude of features. and They represent the corresponding variances. , The statistical fluctuation range obtained by taking the square root; Features Performing normalization and affine transformation yields the output features after statistical harmonization in the online domain: in, This represents the characteristics after online domain statistical reconciliation. A smoothing constant is introduced to prevent numerical instability.

5. A whole-cardiac CT image segmentation method for addressing contrast agent-induced domain shift according to claim 4, characterized in that, The global statistics are maintained online using an exponential moving average method to obtain the global mean and variance. and global variance The details are as follows: The instance statistics of the characteristic means are maintained using an exponential moving average method to maintain their first-order statistics. With second-order statistics : in The coefficient of the exponential moving average; Calculate the global mean and variance based on the maintained first- and second-order statistics: The instance statistic of the characteristic standard deviation is maintained using an exponential moving average method to maintain its first-order statistic. With second-order statistics And calculate the global variance: 。 6. The whole-cardiac CT image segmentation method for addressing contrast agent-induced domain shift according to claim 1, characterized in that, For the initial segmentation prediction results, a reliability mask is constructed by introducing an uncertainty modeling mechanism based on evidence theory, as detailed below: For any voxel position, let Represents any voxel In the The initial predicted logit value for each segmentation category is then defined as the evidence quantity: in, Voxel representation In the Predict the strength of evidence for each segmentation category. This is an evidence mapping adjustment parameter used to control the degree of numerical scaling when mapping the predicted logit to the evidence quantity; The number of segmentation categories; Constructing Dirichlet distribution parameters based on evidence quantity : Define voxel-level uncertainty measure Among them, a single voxel Uncertainty Represented as: By combining the predicted category information, a reliability mask is constructed: in, This represents the category index in the segmentation task; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Indicates the first Reliability masks corresponding to each segmentation category.

7. A whole-cardiac CT image segmentation method for addressing contrast agent-induced domain shift according to claim 6, characterized in that, Under the constraint of the reliability mask, sample-level online calibration is performed using the deep features output in step S2 to obtain sample-level calibration prediction results, as follows: Deep features under reliability mask constraints Perform weighted aggregation to construct the first The dynamic classifier weight vector corresponding to each segmentation category : in, Indicates the first The dynamic classifier weight vector for each segmentation category; Indicates voxel index; Indicates deep features in the th The feature vector at the location of an individual element; Indicates the first The reliability mask corresponding to the segmentation category is in the first... The value at the position of the individual element; Based on the dynamic classifier weight vector, the deep features Classification is performed to obtain sample-level calibration prediction results. : 。 8. A whole-cardiac CT image segmentation system for addressing contrast agent-induced domain shift, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the whole-heart CT image segmentation method as described in any one of claims 1-7.

Images