A training method of an aortic dissection stress field prediction model and related products

Abstract

The application discloses a training method of an aortic dissection stress field prediction model and related products. First, a plurality of CT images of aortic dissection patients are obtained, and cross sections of the maximum diameter of the dissection are identified to obtain a plurality of target cross sections. Based on the cross sections, initial samples and corresponding geometric mechanics information thereof are constructed, and augmented samples and geometric mechanics information thereof are generated. Then, the initial samples and the augmented samples are determined as training samples, and a three-value label set is constructed for each training sample. Subsequently, the initial stress field of the training sample is physically boundary cleaned according to the three-value label set to obtain a target stress field. Finally, the training sample and a regular coordinate set thereof are taken as input, the target stress field is taken as a label, and a Fourier operator network is trained to generate a prediction model. The application can construct a stress field prediction model with strong generalization ability and high stability under small sample conditions, and realize rapid and accurate mechanical evaluation of individualized anatomical structures of different patients.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, specifically to a training method for a stress field prediction model of aortic dissection and related products. Background Technology

[0002] Aortic dissection (AD) is a high-risk cardiovascular disease characterized by rapid onset, rapid progression, and extremely high mortality. Current research has confirmed that the occurrence, tear propagation, and even rupture risk of aortic dissection are directly related to the distribution of stress field on and within the aortic wall. Therefore, rapid, accurate, and stable assessment of the dissection stress field is of irreplaceable value for preoperative risk stratification, surgical planning, and long-term prognosis.

[0003] Currently, in the field of aortic dissection stress field prediction, deep learning has become the mainstream technical route to replace traditional finite element simulation. Among them, Physical-Informed Neural Network (PINN) and Convolutional Neural Network (CNN) are the two most widely used methods.

[0004] However, in practical applications, both CNN and PINN struggle to meet the real requirements of clinical biomechanical assessment. CNN requires large-scale, high-quality simulated labeled data for training and exhibits poor generalization performance when faced with anatomical structures showing significant individual differences. PINN, on the other hand, is prone to training instability and insufficient fitting accuracy in boundary regions. Both methods share the common shortcomings of limited generalization ability and high dependence on simulated data, making it impossible to achieve stable and reliable stress field prediction for the personalized anatomical structures of different patients under the premise of small sample learning.

[0005] Therefore, how to achieve rapid and accurate prediction of aortic dissection stress field with strong generalization and across individuals under small sample conditions is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a training method and related products for aortic dissection stress field prediction models. This method enables the construction of stress field prediction models with strong generalization ability, high training stability, and no need for extensive simulation data support under small sample training conditions, thereby achieving rapid and accurate mechanical assessment of the personalized anatomical structures of different patients.

[0007] The embodiments of this application disclose the following technical solutions: A training method for a stress field prediction model of aortic dissection, the method comprising: CT images of multiple patients with aortic dissection were acquired, and multiple target cross sections were obtained by identifying the cross section with the maximum diameter of the dissection from the multiple CT images. Based on the multiple target cross sections, multiple initial samples and their corresponding geometric and mechanical information are constructed, as well as multiple amplified samples and their corresponding geometric and mechanical information. Each initial sample and each amplified sample is a two-dimensional solid anatomical model. The geometric and mechanical information includes the regular coordinate set, initial stress field, outer wall contour, true cavity inner wall contour, and false cavity inner wall contour of the two-dimensional solid anatomical model. The regular coordinate set and the initial stress field have been normalized to a unified global spatial coordinate system. Each initial sample and each amplified sample are determined as training samples, and a three-value label set corresponding to each training sample is constructed; the three-value label set is assigned different pixel values ​​according to different regions of the two-dimensional entity anatomical model; Based on the three-valued label set corresponding to each training sample, the initial stress field of each training sample is physically cleaned to obtain the target stress field corresponding to each training sample. The Fourier operator network is trained using the training samples and their corresponding set of rule coordinates as training inputs, and the target stress fields corresponding to each training sample as training labels, to obtain the aortic dissection stress field prediction model.

[0008] In one possible implementation, constructing the set of three-value labels corresponding to each training sample includes: For each training sample, the pixel values ​​of the outer region of the outer wall contour are assigned as X, the pixel values ​​of the inner regions of the true lumen inner wall contour and the false lumen inner wall contour are assigned as Y, and the pixel values ​​of the solid region of the blood vessel wall are assigned as Z, thus obtaining the three-value label set corresponding to the training sample.

[0009] In one possible implementation, the construction of multiple initial samples and their corresponding geometric-mechanical information based on the multiple target cross-sections, as well as multiple amplified samples corresponding to the multiple initial samples and their corresponding geometric-mechanical information, includes: Two-dimensional modeling is performed on the cross-sections of the multiple targets to obtain multiple two-dimensional solid anatomical models; Finite element simulation calculations are performed on the multiple two-dimensional solid anatomical models to obtain the regular coordinate set and initial stress field corresponding to each of the multiple two-dimensional solid anatomical models, and the outer wall contour, the true cavity inner wall contour and the false cavity inner wall contour are extracted from the multiple two-dimensional solid anatomical models respectively. A subset of the multiple two-dimensional solid anatomical models were selected as initial samples. Data amplification is performed on multiple initial samples to obtain multiple amplified samples corresponding to each initial sample; Based on the geometric and mechanical information of the initial sample corresponding to each amplified sample, the geometric and mechanical information corresponding to each amplified sample is constructed.

[0010] In one possible implementation, performing finite element simulation calculations on the plurality of two-dimensional solid anatomical models to obtain the regular coordinate sets and initial stress fields corresponding to each of the plurality of two-dimensional solid anatomical models includes: Finite element simulation calculations were performed on the multiple two-dimensional solid anatomical models to obtain the set of mesh node coordinates and Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models. The node coordinates of all the mesh nodes corresponding to the multiple two-dimensional entity anatomical models are sequentially translated and mapped to the bounding box regions of each target to obtain the normalized coordinate set corresponding to each of the multiple two-dimensional entity anatomical models. Based on a preset rule grid, the coordinates in each normalized coordinate set are arranged in a regular manner to obtain multiple regular coordinate sets; Based on the regular grid, nearest neighbor interpolation is performed on the Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models to obtain the initial stress field corresponding to each of the multiple two-dimensional solid anatomical models.

[0011] In one possible implementation, constructing the geometric-mechanical information corresponding to each amplified sample based on the geometric-mechanical information of the initial sample corresponding to each amplified sample includes: For each amplified sample, the regular coordinate set of the initial sample corresponding to the amplified sample is rotated and transformed to preserve the initial stress field, outer wall contour, true cavity inner wall contour and false cavity inner wall contour of the initial sample corresponding to the amplified sample, thus obtaining the geometric and mechanical information corresponding to the amplified sample.

[0012] In one possible implementation, the step of physically cleaning the initial stress field of each training sample based on the ternary label set corresponding to each training sample to obtain the target stress field corresponding to each training sample includes: For each training sample, based on the three-value label set of the training sample, the initial stress field of the training sample is spatially matched and the stress value is set to zero. Specifically, the stress values ​​of the regions with pixel values ​​of X and Y in the three-valued marker set are all set to zero at the corresponding positions in the initial stress field.

[0013] In one possible implementation, the method further includes: A subset of the multiple two-dimensional solid anatomical models were selected as test samples. Construct a set of three-valued labels corresponding to each test sample; Based on the three-valued label set corresponding to each test sample, the initial stress field of each test sample is physically cleaned to obtain the target stress field corresponding to each test sample. The performance of the aortic dissection stress field prediction model is evaluated and verified by using each test sample and its corresponding set of regular coordinates as test inputs, and the target stress field corresponding to each test sample as test labels.

[0014] A training device for a stress field prediction model of aortic dissection, the device comprising: The acquisition unit is used to acquire CT images of multiple patients with aortic dissection and to identify multiple target cross-sections from the cross-sections of the maximum diameter of the dissection from the multiple CT images. The sample construction unit is used to construct multiple initial samples and their corresponding geometric and mechanical information based on the multiple target cross-sections, as well as multiple amplified samples and their corresponding geometric and mechanical information. Each initial sample and each amplified sample is a two-dimensional solid anatomical model. The geometric and mechanical information includes the regular coordinate set, initial stress field, outer wall contour, true cavity inner wall contour, and false cavity inner wall contour of the two-dimensional solid anatomical model. The regular coordinate set and the initial stress field have been normalized to a unified global spatial coordinate system. The sample determination unit is used to determine each initial sample and each amplified sample as training samples. The first label field construction unit is used to construct the three-value label set corresponding to each training sample; the three-value label set is assigned different pixel values ​​according to different regions of the two-dimensional entity anatomical model. The first stress cleaning unit is used to perform physical boundary cleaning on the initial stress field of each training sample based on the three-value label set corresponding to each training sample, so as to obtain the target stress field corresponding to each training sample. The model training unit is used to train the Fourier operator network by taking each training sample and the set of rule coordinates corresponding to each training sample as training input, and taking the target stress field corresponding to each training sample as training label, to obtain the aortic dissection stress field prediction model.

[0015] A training device for aortic dissection stress field prediction model includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the training method for the aortic dissection stress field prediction model as described above.

[0016] A computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the training method for the aortic dissection stress field prediction model as described above.

[0017] Compared with the prior art, this application has the following beneficial effects: This application provides a training method and related products for a stress field prediction model of aortic dissection. Specifically, when implementing the training method for the stress field prediction model of aortic dissection provided in this application, firstly, multiple target cross-sections are obtained by acquiring computed tomography (CT) images of multiple aortic dissection patients and identifying the cross-section with the largest diameter of the dissection. This ensures that the samples are taken from key clinical stress sections, improving data validity. Next, initial samples and their corresponding geometric and mechanical information are constructed based on these target cross-sections, while amplified samples and their geometric and mechanical information are generated. These samples are two-dimensional solid anatomical models, and the geometric and mechanical information includes a set of regular coordinates, an initial stress field, an outer wall contour, a true lumen inner wall contour, and a false lumen inner wall contour. All information is normalized to a unified global spatial coordinate system, thereby expanding the scale of training data and unifying the spatial scale of the samples. Subsequently, these initial and amplified samples are determined as training samples, and a three-value label set is constructed for each training sample to assign different pixel values ​​according to different regions. Next, the initial stress field of the training samples is physically cleaned based on the three-value label set to obtain the target stress field. This effectively eliminates interpolated spurious stresses, ensures the physical authenticity of the stress field distribution, and avoids boundary fitting distortion. Finally, the Fourier operator network is trained using the training samples and their regular coordinate set as input, and the target stress field as the label, thereby generating a strong generalization ability and accurate and stable aortic dissection stress field prediction model. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for training a stress field prediction model for aortic dissection, provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for constructing sample and geometric mechanical information provided in this application embodiment; Figure 3A flowchart illustrating a method for constructing regular coordinates and an initial stress field, as provided in an embodiment of this application; Figure 4 A flowchart illustrating a method for testing and verifying aortic dissection stress field prediction model, as provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of a training device for a predictive model of aortic dissection stress field provided in an embodiment of this application. Detailed Implementation

[0020] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0021] Currently, deep learning technology has become the main method for predicting stress fields in Alzheimer's disease (AD), with PINN and CNN being the two most commonly used. However, in practical applications, CNNs typically rely on a large amount of high-quality simulation data for training, and their generalization ability is insufficient when dealing with anatomical structures with significant individual differences. While PINN incorporates physical knowledge, it is prone to instability and inaccurate fitting of boundary regions during training. This makes both methods unable to effectively adapt to the personalized anatomical characteristics of different patients under small sample learning conditions, thus limiting their clinical application.

[0022] To address this issue, this application provides a training method and related product for aortic dissection stress field prediction model. First, CT images of multiple aortic dissection patients are acquired, and the cross-section with the largest diameter of the dissection is identified to obtain multiple target cross-sections. Next, multiple initial samples and their corresponding geometric and mechanical information are constructed based on these target cross-sections, while multiple amplified samples and their corresponding geometric and mechanical information are generated for each initial sample. These samples are all two-dimensional solid anatomical models, and the geometric and mechanical information includes a set of regular coordinates, an initial stress field, an outer wall contour, a true lumen inner wall contour, and a false lumen inner wall contour. The set of regular coordinates and the initial stress field have been normalized to a unified global spatial coordinate system. Subsequently, these initial and amplified samples are determined as training samples, and a corresponding three-value label set is constructed, which assigns different pixel values ​​according to different regions of the two-dimensional solid anatomical model. Next, the initial stress field of each training sample is physically cleaned to obtain the corresponding target stress field. Finally, the training samples and their corresponding set of regular coordinates are used as training input, and the target stress field is used as training labels to train a Fourier operator network, thereby obtaining the aortic dissection stress field prediction model. This application enables the rapid construction and expansion of samples based on clinical CT scans, achieving stable training with small sample sizes. Simultaneously, through geometric normalization, Fourier operator networks, and physical boundary constraints, it significantly improves the model's generalization ability and prediction accuracy, ensuring the reliability of stress field results.

[0023] 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 embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] See Figure 1 The figure is a flowchart of a method for training a stress field prediction model for aortic dissection according to an embodiment of this application. Figure 1 As shown, the training method for the aortic dissection stress field prediction model may include steps S101-S105: S101: Acquire CT images of multiple patients with aortic dissection, and identify the cross-section with the largest diameter of the dissection from the multiple CT images to obtain multiple target cross-sections.

[0025] In constructing the training samples for the aortic dissection stress field prediction model, multiple CT images of aortic dissection from different patients can be acquired first. These images contain the actual anatomical structure and lesion morphology of the patients' blood vessels. To ensure that the model learns the features with the greatest mechanical risk and clinical representativeness, key sections need to be extracted from each CT image.

[0026] Specifically, for each CT image, layer-by-layer detection and diameter calculation can be performed along the vessel axis to locate the position where the dissection lumen dilation is most significant and the diameter is largest, and the corresponding tomographic image is determined as the target cross-section. By performing the above operation on the CT images of all patients, multiple standardized target cross-sections are finally obtained. This step can effectively focus on the core area of ​​the lesion, eliminate redundant tomographic information, and ensure that the subsequently constructed two-dimensional anatomical models are all derived from the cross-sections with the greatest mechanical analysis value, providing a unified and reliable data foundation for sample construction and model training.

[0027] S102: Based on the multiple target cross sections, construct multiple initial samples and their corresponding geometric and mechanical information, as well as multiple amplified samples corresponding to the multiple initial samples and their corresponding geometric and mechanical information.

[0028] To ensure the authenticity, representativeness, and universality of the model training data, and to provide accurate and unified basic data support for subsequent model training, initial and expanded samples can be constructed based on the obtained multiple target cross-sections, and the corresponding geometric and mechanical information can be extracted.

[0029] Specifically, based on the anatomical structure presented by the target cross-section, multiple initial samples are constructed. Each initial sample is a standardized two-dimensional solid anatomical model that can completely preserve the core anatomical features of aortic dissection. At the same time, in order to enrich the number of training samples and improve the model's ability to adapt to anatomical differences among different individuals, multiple expanded samples are generated for each initial sample. This ensures that the samples cover the anatomical differences among different individuals while further expanding the scale of training data.

[0030] For each initial and amplified sample, corresponding geometric and mechanical information is extracted simultaneously. This information covers the regular coordinate set, initial stress field, outer wall contour, true lumen inner wall contour, and false lumen inner wall contour of the two-dimensional solid anatomical model. The regular coordinate set and initial stress field have undergone unified normalization processing and are strictly adapted to a unified global spatial coordinate system. This ensures that the geometric parameters and mechanical characteristics of different samples are consistent and comparable, effectively avoiding the problem of insufficient model generalization ability caused by differences in spatial scale. At the same time, it clearly conveys the key mechanical characteristics and anatomical information of the vascular structure, providing comprehensive and standardized basic data for subsequent model training.

[0031] S103: Determine each initial sample and each amplified sample as training samples, and construct a set of three-value labels corresponding to each training sample.

[0032] After acquiring and processing the clinical data containing information related to aortic dissection, in order to further ensure the effectiveness and comprehensiveness of model training, all the initial samples and corresponding amplified samples that have been constructed are uniformly determined as the training samples required for model training. This ensures that the training data can cover various scenarios such as differences in anatomical structure and different degrees of lesions among different individuals, providing the model with sufficient and representative learning basis.

[0033] At the same time, for each training sample, combined with its corresponding two-dimensional entity anatomical model, different regions are assigned corresponding pixel values ​​according to their functional and structural differences.

[0034] Specifically, different pixel values ​​are assigned to the blood vessel wall region, lesion region, and background region of the two-dimensional solid anatomical model. This explicit pixel assignment distinguishes the functional regions of different anatomical structures, constructing a three-value label set that corresponds one-to-one with each training sample. This three-value label set clearly defines the attributes of different regions in each training sample, accurately conveying the anatomical features and mechanical information of each region. It provides a clear labeling basis for the subsequent model to learn the differences in anatomical structures among different individuals and improve generalization ability. This ensures that the model can accurately capture and learn the differences in anatomical structures and mechanical features among different individuals, avoiding model training bias caused by sample differences, and laying the foundation for stable model learning and accurate prediction.

[0035] In one possible implementation, step S103 involves constructing a set of three-value labels corresponding to each training sample, including: for each training sample, assigning different pixel values ​​to different locations according to the differences in anatomical regions, assigning the pixel value of the background region outside the outer wall contour to X (e.g., -2), assigning the pixel value of the cavity region inside the true lumen inner wall contour and the false lumen inner wall contour to Y (e.g., -1), and assigning the pixel value of the solid region of the blood vessel wall between the outer wall, the true lumen inner wall and the false lumen inner wall to Z (e.g., 1), thereby clearly distinguishing three different regions and obtaining the set of three-value labels corresponding to the training sample.

[0036] S104: Based on the three-valued label set corresponding to each training sample, perform physical boundary cleaning on the initial stress field of each training sample to obtain the target stress field corresponding to each training sample.

[0037] After obtaining the training samples of the aortic dissection stress prediction model, in order to ensure the accuracy and rationality of the subsequent stress prediction results, the initial stress field of each training sample can be physically cleaned based on the three-value label set corresponding to each training sample.

[0038] Specifically, ternary labeling can accurately distinguish between solid and non-solid regions of blood vessels, retaining only the effective stress corresponding to solid structures such as the vessel wall, true lumen, and false lumen, while completely eliminating spurious stress and interpolation noise in non-solid regions such as the background and cavities. Simultaneously, it corrects unreasonable stress abrupt changes and numerical deviations at the boundaries. This boundary constraint processing eliminates spurious stresses in the initial stress field that do not conform to physiological structures, ensuring that the stress distribution strictly matches the physical reality of the anatomical region, thus obtaining an accurate and standardized target stress field.

[0039] In one possible implementation, step S104 involves physical boundary cleaning of the initial stress field of each training sample based on the ternary label set corresponding to each training sample, to obtain the target stress field corresponding to each training sample, including: For each training sample, the ternary label set is matched point-by-point in space with the initial stress field, and the stress values ​​are physically constrained according to the region category. Specifically, for non-solid regions with pixel values ​​X and Y in the ternary label set, the stress values ​​at their corresponding positions in the initial stress field are all set to zero, retaining only the stress values ​​of the blood vessel wall solid region with pixel value Z. For example, when X=-2, Y=-1, and Z=1 in the ternary labels, the stress values ​​corresponding to the background region labeled -2 ​​and the true and false cavity regions labeled -1 are set to 0, retaining only the stress of the blood vessel solid region labeled 1. This eliminates the spurious stress in non-stressed regions, resulting in a target stress field that conforms to the true mechanical distribution.

[0040] S105: Using each training sample and the set of rule coordinates corresponding to each training sample as training input, and using the target stress field corresponding to each training sample as training label, the Fourier operator network is trained to obtain the aortic dissection stress field prediction model.

[0041] After completing sample construction, geometric normalization, ternary labeling, and physical boundary cleaning, the model training phase begins. Standardized training samples and their corresponding regular coordinate sets are used as network input to characterize the complete geometric structure and spatial location information of aortic dissection. Simultaneously, the target stress field, corrected for physical constraints, serves as a supervisory label to provide a reliable mechanical distribution benchmark. The Fourier operator network is trained using these paired data sets, enabling it to learn the mapping relationship from global geometric topology to stress field distribution. This fully captures the intrinsic correlation between complex anatomical structures and mechanical responses, ultimately resulting in a highly accurate, generalizable, and physically consistent aortic dissection stress field prediction model.

[0042] In one possible implementation, the relative error between the predicted stress matrix and the stress matrix solved by finite element analysis (FEA) is minimized during the training of the Fourier operator network to update the neural parameters in the Fourier Neural Operator (FNO) network, thus obtaining the final aortic dissection stress field prediction model.

[0043] Based on the content of S101-S105, CT images of multiple patients with aortic dissection are first acquired, and the cross-section with the largest diameter of the dissection is identified to obtain multiple target cross-sections. Next, multiple initial samples and their corresponding geometric and mechanical information are constructed based on these target cross-sections. Simultaneously, an amplified sample and its corresponding geometric and mechanical information are generated for each initial sample. These samples are two-dimensional solid anatomical models, and their geometric and mechanical information includes a set of regular coordinates, an initial stress field, an outer wall contour, a true lumen inner wall contour, and a false lumen inner wall contour. The regular coordinate set and the initial stress field are normalized to ensure comparison within a unified global spatial coordinate system. Subsequently, all initial and amplified samples are selected as training samples, and a corresponding three-value label set is constructed for each training sample. Then, based on these three-value label sets, the initial stress field of each training sample is physically cleaned to obtain the corresponding target stress field. Finally, the training samples and their corresponding regular coordinate sets are used as training inputs, and the target stress field is used as training labels to train a Fourier operator network, thereby obtaining the final aortic dissection stress field prediction model. This application enables the rapid construction and expansion of samples based on clinical CT scans, achieving stable training even with small sample sizes. Simultaneously, through geometric normalization, Fourier operator networks, and physical boundary constraints, it significantly improves the model's generalization ability and prediction accuracy, ensuring the reliability of stress field results.

[0044] See Figure 2 , Figure 2 This application provides a flowchart of a method for constructing sample and geometric mechanical information. Accordingly, step S102 constructs multiple initial samples and their corresponding geometric mechanical information based on the multiple target cross-sections, as well as multiple amplified samples corresponding to each initial sample and their corresponding geometric mechanical information. This can be specifically implemented through steps S201-S205. S201: Perform two-dimensional modeling on the multiple target cross sections to obtain multiple two-dimensional solid anatomical models.

[0045] To provide accurate and standardized geometric models for subsequent finite element mechanical simulations, multiple target cross-sections can be modeled in two dimensions. This two-dimensional modeling process sequentially includes three steps: image segmentation, geometric reconstruction, and unstructured mesh generation. Image segmentation accurately extracts the contours of the outer wall, true lumen, and false lumen of the aortic dissection. Geometric reconstruction forms a complete closed geometric region, and unstructured mesh generation discretizes the region, ultimately resulting in multiple structurally complete two-dimensional solid anatomical models suitable for mechanical calculations.

[0046] S202: Perform finite element simulation calculations on the multiple two-dimensional solid anatomical models to obtain the regular coordinate set and initial stress field corresponding to each of the multiple two-dimensional solid anatomical models, and extract the outer wall contour, the true cavity inner wall contour and the false cavity inner wall contour from the multiple two-dimensional solid anatomical models respectively.

[0047] In order to obtain geometric and mechanical data that can reflect the true physiological stress state of aortic dissection, finite element simulation calculations can be performed on multiple two-dimensional solid anatomical models.

[0048] The simulation is set as a plane strain problem, using a linear elastic material constitutive model, and a reasonable Young's modulus is set (e.g., ...). ) and Poisson's ratio (e.g., The mesh employs a fine subdivision with a minimum element size of no more than 0.2 mm to ensure computational accuracy; simultaneously, it constrains the circumferential displacement of the blood vessel (i.e., displacement along the circumference of the blood vessel). ), release radial displacement (radial displacement is free and unrestrained, allowing the interstitial space to expand), and apply uniform radial pressure simulating physiological blood pressure to the inner walls of the true and false cavities (e.g. p =10 kPa).

[0049] Through the above simulation, the spatial location information of the nodes of each model is obtained to form a set of regular coordinates, and the initial Von Mises corresponding to each model is calculated. Mises (equivalent force field).

[0050] Simultaneously, the outer wall contour, true cavity inner wall contour, and false cavity inner wall contour are extracted from each two-dimensional solid anatomical model, providing complete and accurate geometric features and mechanical labels for subsequent sample construction, stress field cleaning, and model training.

[0051] S203: Select a portion of the multiple two-dimensional physical anatomical models as initial samples.

[0052] In order to control training costs while ensuring sample diversity and representativeness, models with typical anatomical structures and complete mechanical characteristics were selected from multiple two-dimensional solid anatomical models that have completed finite element simulations as initial samples, providing a stable and reliable original data foundation for subsequent data amplification and model training.

[0053] S204: Perform data amplification on multiple initial samples to obtain multiple amplified samples corresponding to each initial sample.

[0054] To enhance the richness of training data and the generalization ability of the model, and to avoid overfitting of the network due to a single sample, data augmentation operations such as rotation transformation can be performed on multiple initial samples. While keeping the anatomical structure and mechanical properties unchanged, multiple variants of different spatial poses corresponding to each initial sample are generated, thus obtaining a sufficient number of augmented samples with diverse distributions.

[0055] S205: Construct the geometric and mechanical information of each amplified sample based on the geometric and mechanical information of the initial sample corresponding to each amplified sample.

[0056] To ensure the geometric consistency and mechanical authenticity between the amplified samples and the initial samples, the spatial coordinates, contour parameters, and stress distribution are synchronously transformed and mapped based on the geometric and mechanical information of the initial sample corresponding to each amplified sample. This generates a set of regular coordinates, an initial stress field, and various contour information corresponding to the amplified sample, providing reliable standardized input for subsequent network training.

[0057] In one possible implementation, in order to expand the amount of training data and improve the rotation invariance and generalization ability of the model without changing the mechanical response and anatomical boundary features of the samples, step S205 constructs the geometric and mechanical information corresponding to each amplified sample based on the geometric and mechanical information of the initial sample corresponding to each amplified sample, including: For each amplified sample, the initial stress field, outer wall contour, true cavity inner wall contour, and false cavity inner wall contour of the corresponding initial sample are kept unchanged. Only the regular coordinate set of the initial sample is rotated and shifted to obtain the geometric mechanical information corresponding to the amplified sample that is consistent with the mechanical properties and boundary topology of the initial sample.

[0058] Steps S201-S205 can effectively expand the scale of training data while ensuring the anatomical authenticity and mechanical accuracy of the samples, and at the same time maintain the consistency of the geometric features and mechanical distribution of the initial samples and the amplified samples, thereby improving the generalization ability and training stability of the model.

[0059] See Figure 3 , Figure 3 This application provides a flowchart of a method for constructing regular coordinates and an initial stress field. Accordingly, step S202 involves performing finite element simulation calculations on the multiple two-dimensional solid anatomical models to obtain the regular coordinate sets and initial stress fields corresponding to each of the multiple two-dimensional solid anatomical models. Specifically, this can be achieved through steps S301-S304. S301: Perform finite element simulation calculations on the multiple two-dimensional solid anatomical models to obtain the set of mesh node coordinates and Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models.

[0060] To obtain the fundamental geometric and mechanical data required for constructing the regular coordinate field and initial stress field, linear elastic finite element static simulation calculations were performed on the multiple two-dimensional solid anatomical models by applying loads and boundary constraints conforming to human physiological conditions. Then, based on the simulation results, the spatial position information of all mesh nodes corresponding to each two-dimensional solid anatomical model was extracted to form a corresponding set of mesh node coordinates. Simultaneously, the mechanical response data obtained from finite element calculations at each mesh node were acquired to form a corresponding set of Von Mises equivalent stresses.

[0061] S302: The node coordinates of all the mesh nodes corresponding to the multiple two-dimensional entity anatomical models are sequentially translated and mapped to the bounding box regions of each target to obtain the normalized coordinate set corresponding to each of the multiple two-dimensional entity anatomical models.

[0062] To eliminate spatial offsets and size differences between different two-dimensional physical anatomical models, and to achieve geometric space unification and standardization, facilitating subsequent regular mesh division and field data alignment, the coordinates of all mesh nodes corresponding to each of the multiple two-dimensional physical anatomical models are sequentially mapped to a preset 0.05m×0.05m unified square target bounding box region through translation transformation. This ensures that the geometric domains of all models are normalized to the same spatial scale and position range, thereby obtaining the normalized coordinate set corresponding to each of the multiple two-dimensional physical anatomical models.

[0063] S303: Based on a preset regular grid, the coordinates in each normalized coordinate set are arranged in a regular manner to obtain multiple regular coordinate sets.

[0064] To convert the discrete coordinates of different models after translation and normalization into field data with a regular structure and consistent dimensions, so as to adapt to the requirements of Fourier operator networks for regular input, the coordinate points in each normalized coordinate set can be spatially aligned, rearranged and normalized sequentially based on a uniform regular grid with a preset resolution of 128×128. This maps and arranges the originally scattered and inconsistent normalized node coordinates onto regular grid points of a uniform specification, ultimately obtaining the regular coordinate sets corresponding to each of the multiple two-dimensional entity anatomical models.

[0065] S304: Based on the regular grid, perform nearest-neighbor interpolation on the Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models to obtain the initial stress field corresponding to each of the multiple two-dimensional solid anatomical models.

[0066] To fully map the discrete, non-uniformly distributed Von Mises equivalent stress values ​​obtained from finite element analysis onto a unified, structured regular mesh, forming continuous stress field data that can be directly used for network training, based on a defined 128×128 regular mesh, and using the stress values ​​of each original mesh node as a basis, nearest neighbor interpolation calculations are performed sequentially on all nodes on the regular mesh. This assigns the discrete stress values ​​to each corresponding position on the regular mesh, thereby generating an initial stress field that is spatially continuous, uniformly formatted, and completely aligned with the regular coordinate field, providing standard label input for subsequent stress field prediction models.

[0067] Steps S301-S304 can effectively obtain the regular coordinate set and initial stress field of the two-dimensional solid anatomical model, laying the foundation for subsequent stress field prediction model training.

[0068] See Figure 4 , Figure 4 A flowchart of a method for testing and verifying aortic dissection stress field prediction model provided in this application embodiment is shown, which can be implemented through steps S401-S404: S401: Select a portion of the multiple two-dimensional physical anatomical models as test samples.

[0069] To objectively, independently, and unbiasedly verify the generalization performance and computational accuracy of the aortic dissection stress field prediction model, some models that did not participate in the model training can be selected from the multiple two-dimensional solid anatomical models that have been constructed as independent test samples, so as to ensure that the subsequent evaluation results can truly reflect the model's predictive ability on unknown anatomical structures.

[0070] S402: Construct a set of three-valued labels for each test sample.

[0071] This step is similar to the process of constructing the three-value label set corresponding to each training sample in step S103, and will not be described again here.

[0072] S403: Based on the three-valued label set corresponding to each test sample, perform physical boundary cleaning on the initial stress field of each test sample to obtain the target stress field corresponding to each test sample.

[0073] In order to eliminate spurious stress values ​​in non-solid regions caused by interpolation discretization in the initial stress field and ensure that the stress field distribution strictly conforms to the mechanical reality of aortic dissection, the initial stress field is cleaned by boundary constraints based on the three-valued label set corresponding to each test sample as a physical mask. Only the stress values ​​in solid regions are retained, and the stress in the cavity and background regions is forced to zero. Finally, a target stress field with accurate boundaries that can be used as a test label is obtained.

[0074] S404: Using each test sample and its corresponding set of regular coordinates as test inputs, and the target stress field corresponding to each test sample as test labels, the performance of the aortic dissection stress field prediction model is evaluated and verified.

[0075] To objectively, rigorously, and unbiasedly verify the generalization ability and computational accuracy of the aortic dissection stress field prediction model, each test sample and its corresponding set of regular coordinates were used as the test input of the model. At the same time, the target stress field after physical boundary cleaning was used as the standard test label and input into the trained aortic dissection stress field prediction model. Predictive inference was performed on multiple new and unseen topological models in the prediction set. The output global test set stress field was extracted, and the coefficient of determination, root mean square error, and other indicators were used to compare and evaluate the model with actual FEA simulation data to verify accuracy. This completed the accuracy test, generalization verification, and comprehensive performance evaluation of the model.

[0076] Steps S401-S404 allow for an objective, accurate, and comprehensive evaluation of the stress field prediction model's generalization ability and computational accuracy on unknown anatomical structures, ensuring that the model's output meets the reliability requirements of clinical mechanical analysis.

[0077] See Figure 5 , Figure 5 This is a schematic diagram of the structure of a training device for a predictive model of aortic dissection stress field provided in an embodiment of this application. Figure 5 As shown, the training device for the aortic dissection stress field prediction model includes: The acquisition unit 501 is used to acquire CT images of multiple patients with aortic dissection, and to identify the cross-section with the maximum diameter of the dissection from the multiple CT images to obtain multiple target cross-sections. The sample construction unit 502 is used to construct multiple initial samples and their corresponding geometric and mechanical information based on the multiple target cross-sections, as well as multiple amplified samples and their corresponding geometric and mechanical information. Each initial sample and each amplified sample is a two-dimensional solid anatomical model. The geometric and mechanical information includes the regular coordinate set, initial stress field, outer wall contour, true cavity inner wall contour, and false cavity inner wall contour of the two-dimensional solid anatomical model. The regular coordinate set and the initial stress field have been normalized to a unified global spatial coordinate system. The sample determination unit 503 is used to determine each initial sample and each amplified sample as training samples; The first label field construction unit 504 is used to construct a three-value label set corresponding to each training sample; the three-value label set is assigned different pixel values ​​according to different regions of the two-dimensional entity anatomical model. The first stress cleaning unit 505 is used to perform physical boundary cleaning on the initial stress field of each training sample based on the three-value label set corresponding to each training sample, so as to obtain the target stress field corresponding to each training sample. The model training unit 506 is used to train the Fourier operator network by taking each training sample and the set of regular coordinates corresponding to each training sample as training input, and taking the target stress field corresponding to each training sample as training label, to obtain the aortic dissection stress field prediction model.

[0078] In one possible implementation, the first marker field construction unit 504 is specifically used for: For each training sample, the pixel values ​​of the outer region of the outer wall contour are assigned as X, the pixel values ​​of the inner regions of the true lumen inner wall contour and the false lumen inner wall contour are assigned as Y, and the pixel values ​​of the solid region of the blood vessel wall are assigned as Z, thus obtaining the three-value label set corresponding to the training sample.

[0079] In one possible implementation, the sample construction unit 502 specifically includes: A two-dimensional modeling unit is used to perform two-dimensional modeling on the multiple target cross sections respectively to obtain multiple two-dimensional solid anatomical models; The simulation calculation unit is used to perform finite element simulation calculations on the multiple two-dimensional solid anatomical models to obtain the regular coordinate set and initial stress field corresponding to each of the multiple two-dimensional solid anatomical models. The contour extraction unit is used to extract the outer wall contour, the true cavity inner wall contour, and the false cavity inner wall contour from the multiple two-dimensional solid anatomical models, respectively. An initial sample selection unit is used to select a portion of the models from the plurality of two-dimensional solid anatomical models as initial samples; The data amplification unit is used to amplify multiple initial samples to obtain multiple amplified samples corresponding to each initial sample. The geometric-mechanical information construction unit is used to construct the geometric-mechanical information corresponding to each amplified sample based on the geometric-mechanical information of the initial sample corresponding to each amplified sample.

[0080] In one possible implementation, the simulation calculation unit is specifically used for: Finite element simulation calculations were performed on the multiple two-dimensional solid anatomical models to obtain the set of mesh node coordinates and Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models. The node coordinates of all the mesh nodes corresponding to the multiple two-dimensional entity anatomical models are sequentially translated and mapped to the bounding box regions of each target to obtain the normalized coordinate set corresponding to each of the multiple two-dimensional entity anatomical models. Based on a preset rule grid, the coordinates in each normalized coordinate set are arranged in a regular manner to obtain multiple regular coordinate sets; Based on the regular grid, nearest neighbor interpolation is performed on the Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models to obtain the initial stress field corresponding to each of the multiple two-dimensional solid anatomical models.

[0081] In one possible implementation, the geometric and mechanical information construction unit is specifically used for: For each amplified sample, the regular coordinate set of the initial sample corresponding to the amplified sample is rotated and transformed to preserve the initial stress field, outer wall contour, true cavity inner wall contour and false cavity inner wall contour of the initial sample corresponding to the amplified sample, thus obtaining the geometric and mechanical information corresponding to the amplified sample.

[0082] In one possible implementation, the first stress cleaning unit 505 is specifically used for: For each training sample, based on the three-value label set of the training sample, the initial stress field of the training sample is spatially matched and the stress value is set to zero. Specifically, the stress values ​​of the regions with pixel values ​​of X and Y in the three-valued marker set are all set to zero at the corresponding positions in the initial stress field.

[0083] In one possible implementation, the device further includes: The test sample selection unit is used to select a portion of the multiple two-dimensional solid anatomical models as test samples. The second label field construction unit is used to construct the set of three-valued labels corresponding to each test sample; The second stress cleaning unit is used to perform physical boundary cleaning on the initial stress field of each test sample based on the three-valued label set corresponding to each test sample, so as to obtain the target stress field corresponding to each test sample. The model validation unit is used to evaluate and validate the performance of the aortic dissection stress field prediction model by taking each test sample and its corresponding set of regular coordinates as test inputs and the target stress field corresponding to each test sample as test labels.

[0084] In addition, this application embodiment also provides a training device for aortic dissection stress field prediction model, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the training method for aortic dissection stress field prediction model as described above.

[0085] In addition, this application embodiment also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the training method for the aortic dissection stress field prediction model as described above.

[0086] This application's embodiments analyze CT images of multiple patients with aortic dissection to identify the cross-section with the largest diameter of the dissection, effectively constructing multiple initial and expanded samples. This enhances the model's adaptability to different individual patient anatomical structures, enabling stable training even with small sample sizes. Simultaneously, normalizing the geometric and mechanical information to a unified global spatial coordinate system ensures comparability between different samples. Furthermore, introducing a Fourier operator network effectively captures and expresses complex geometric shapes and physical properties, further improving the model's predictive stability and generalization performance. Additionally, through ternary label sets and physical boundary cleaning, the solid and non-solid regions of the blood vessel are accurately encoded, eliminating interpolated spurious stresses, ensuring the physical realism of the stress field, and avoiding boundary fitting distortion.

[0087] The foregoing provides a detailed description of the training method for aortic dissection stress field prediction model and related products provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0088] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

Claims

1. A training method for a stress field prediction model of aortic dissection, characterized in that, The method includes: CT images of multiple patients with aortic dissection were acquired, and multiple target cross sections were obtained by identifying the cross section with the maximum diameter of the dissection from the multiple CT images. Based on the multiple target cross sections, multiple initial samples and their corresponding geometric and mechanical information are constructed, as well as multiple amplified samples and their corresponding geometric and mechanical information. Each initial sample and each amplified sample is a two-dimensional solid anatomical model. The geometric and mechanical information includes the regular coordinate set, initial stress field, outer wall contour, true cavity inner wall contour, and false cavity inner wall contour of the two-dimensional solid anatomical model. The regular coordinate set and the initial stress field have been normalized to a unified global spatial coordinate system. Each initial sample and each amplified sample are determined as training samples, and a three-value label set corresponding to each training sample is constructed; the three-value label set is assigned different pixel values ​​according to different regions of the two-dimensional entity anatomical model; Based on the three-valued label set corresponding to each training sample, the initial stress field of each training sample is physically cleaned to obtain the target stress field corresponding to each training sample. The Fourier operator network is trained using the training samples and their corresponding set of rule coordinates as training inputs, and the target stress fields corresponding to each training sample as training labels, to obtain the aortic dissection stress field prediction model.

2. The method according to claim 1, characterized in that, The construction of the three-value label set corresponding to each training sample includes: For each training sample, the pixel values ​​of the outer region of the outer wall contour are assigned as X, the pixel values ​​of the inner regions of the true lumen inner wall contour and the false lumen inner wall contour are assigned as Y, and the pixel values ​​of the solid region of the blood vessel wall are assigned as Z, thus obtaining the three-value label set corresponding to the training sample.

3. The method according to claim 1, characterized in that, The construction of multiple initial samples and their corresponding geometric and mechanical information based on the multiple target cross-sections, as well as multiple amplified samples corresponding to each of the initial samples and their corresponding geometric and mechanical information, includes: Two-dimensional modeling is performed on the cross-sections of the multiple targets to obtain multiple two-dimensional solid anatomical models; Finite element simulation calculations are performed on the multiple two-dimensional solid anatomical models to obtain the regular coordinate set and initial stress field corresponding to each of the multiple two-dimensional solid anatomical models, and the outer wall contour, the true cavity inner wall contour and the false cavity inner wall contour are extracted from the multiple two-dimensional solid anatomical models respectively. A subset of the multiple two-dimensional solid anatomical models were selected as initial samples. Data amplification is performed on multiple initial samples to obtain multiple amplified samples corresponding to each initial sample; Based on the geometric and mechanical information of the initial sample corresponding to each amplified sample, the geometric and mechanical information corresponding to each amplified sample is constructed.

4. The method according to claim 3, characterized in that, The step of performing finite element simulation calculations on the multiple two-dimensional solid anatomical models to obtain the regular coordinate set and initial stress field corresponding to each of the multiple two-dimensional solid anatomical models includes: Finite element simulation calculations were performed on the multiple two-dimensional solid anatomical models to obtain the set of mesh node coordinates and Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models. The node coordinates of all the mesh nodes corresponding to the multiple two-dimensional entity anatomical models are sequentially translated and mapped to the bounding box regions of each target to obtain the normalized coordinate set corresponding to each of the multiple two-dimensional entity anatomical models. Based on a preset rule grid, the coordinates in each normalized coordinate set are arranged in a regular manner to obtain multiple regular coordinate sets; Based on the regular grid, nearest neighbor interpolation is performed on the Von Mises equivalent stress set corresponding to each of the multiple two-dimensional solid anatomical models to obtain the initial stress field corresponding to each of the multiple two-dimensional solid anatomical models.

5. The method according to claim 3, characterized in that, The construction of geometric-mechanical information for each amplified sample based on the geometric-mechanical information of the initial sample corresponding to each amplified sample includes: For each amplified sample, the regular coordinate set of the initial sample corresponding to the amplified sample is rotated and transformed to preserve the initial stress field, outer wall contour, true cavity inner wall contour and false cavity inner wall contour of the initial sample corresponding to the amplified sample, thus obtaining the geometric and mechanical information corresponding to the amplified sample.

6. The method according to claim 1, characterized in that, The method involves physically cleaning the initial stress field of each training sample based on the ternary label set corresponding to each training sample, to obtain the target stress field corresponding to each training sample, including: For each training sample, based on the three-value label set of the training sample, the initial stress field of the training sample is spatially matched and the stress value is set to zero. Specifically, the stress values ​​of the regions with pixel values ​​of X and Y in the three-valued marker set are all set to zero at the corresponding positions in the initial stress field.

7. The method according to claim 1, characterized in that, The method further includes: A subset of the multiple two-dimensional solid anatomical models were selected as test samples. Construct a set of three-valued labels corresponding to each test sample; Based on the three-valued label set corresponding to each test sample, the initial stress field of each test sample is physically cleaned to obtain the target stress field corresponding to each test sample. The performance of the aortic dissection stress field prediction model is evaluated and verified by using each test sample and its corresponding set of regular coordinates as test inputs, and the target stress field corresponding to each test sample as test labels.

8. A training device for a stress field prediction model of aortic dissection, characterized in that, The device includes: The acquisition unit is used to acquire CT images of multiple patients with aortic dissection and to identify multiple target cross-sections from the cross-sections of the maximum diameter of the dissection from the multiple CT images. The sample construction unit is used to construct multiple initial samples and their corresponding geometric and mechanical information based on the multiple target cross-sections, as well as multiple amplified samples and their corresponding geometric and mechanical information. Each initial sample and each amplified sample is a two-dimensional solid anatomical model. The geometric and mechanical information includes the regular coordinate set, initial stress field, outer wall contour, true cavity inner wall contour, and false cavity inner wall contour of the two-dimensional solid anatomical model. The regular coordinate set and the initial stress field have been normalized to a unified global spatial coordinate system. The sample determination unit is used to determine each initial sample and each amplified sample as training samples. The first label field construction unit is used to construct the three-value label set corresponding to each training sample; the three-value label set is assigned different pixel values ​​according to different regions of the two-dimensional entity anatomical model. The first stress cleaning unit is used to perform physical boundary cleaning on the initial stress field of each training sample based on the three-value label set corresponding to each training sample, so as to obtain the target stress field corresponding to each training sample. The model training unit is used to train the Fourier operator network by taking each training sample and the set of rule coordinates corresponding to each training sample as training input, and taking the target stress field corresponding to each training sample as training label, to obtain the aortic dissection stress field prediction model.

9. A training device for a stress field prediction model of aortic dissection, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements a training method for aortic dissection stress field prediction model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the training method for the aortic dissection stress field prediction model as described in any one of claims 1-7.

Images