Deep learning-based CT aorta robustness three-dimensional segmentation method and device
By introducing a memory mechanism and a deep learning method for dynamic tracking quality assessment, the problem of insufficient robustness in CT aortic 3D segmentation is solved, achieving fully automated and efficient segmentation results, supporting clinical applications.
Patent Information
- Application Number
- CN202511506142.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-21
- Publication Date
- 2026-01-09
AI Technical Summary
Existing deep learning models have low robustness and insufficient automation in 3D segmentation of the aorta in CT scans. They struggle to maintain tracking coherence and topological connectivity when dealing with aortic bifurcation or dramatic morphological changes, and thus cannot meet the clinical demand for fully automated processing.
A robust 3D segmentation method for the aorta based on deep learning is adopted. By extracting the aortic structure frame by frame, a medical video segmentation model with a memory mechanism is introduced. Combined with lazy loading and dynamic tracking quality assessment, automatic re-anchoring and memory reset are triggered to ensure the robustness and automation of the segmentation.
It achieves fully automated and robust 3D segmentation of the aorta using CT, providing precise and efficient imaging support and reliable image data for clinical diagnosis, treatment planning, and surgical navigation.
Smart Images
Figure CN121304935A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to a robust three-dimensional segmentation method and apparatus for the aorta based on deep learning in CT imaging. Background Technology
[0002] CT imaging is crucial in the diagnosis and treatment of aortic diseases, and its precise three-dimensional segmentation is of great significance for clinical diagnosis and surgical planning. As the largest artery in the human body, the aorta's accurate three-dimensional segmentation is clinically vital for diagnosing serious diseases such as aortic dissection, aortic aneurysm, and atherosclerosis, as well as for preoperative planning and navigation for complex interventional procedures such as endovascular stenting (EVAR) and transcatheter aortic valve replacement (TAVR). However, traditional manual or image intensity-based methods are time-consuming, subjective, and susceptible to noise interference, making it difficult to accurately delineate the complex morphology of the aorta, especially in the branches and lesion areas.
[0003] In recent years, deep learning technology has made progress in the field of medical image segmentation. Medical video segmentation models with memory mechanisms, as general-purpose segmentation models specifically optimized for medical images, have demonstrated stronger segmentation capabilities in the medical field. However, existing methods or systems that directly apply these models to CT aortic 3D segmentation still face challenges: even these fundamental medical models, while performing well in single-frame segmentation, are susceptible to internal memory drift when handling aortic bifurcation or drastic morphological changes, leading to discontinuous tracking or target loss, and difficulty in maintaining topological connectivity across the entire vascular tree. Existing attempts may use single-frame segmentation results as simple cues for the video model, but lack dynamic monitoring, evaluation, and adaptive correction mechanisms for the video model's memory state. Especially when tracking quality deteriorates, they cannot effectively utilize independent robust segmentation results for memory reset and re-anchoring. Furthermore, the reliance on accurate cues limits the degree of automation, making it difficult to meet the clinical demand for fully automated processing of large-scale CT sequences. Therefore, the existing technology system still lacks an integrated, adaptive solution that can effectively manage video model memory drift and achieve adaptive correction and robust memory anchoring in order to achieve robust, accurate, and fully automated 3D segmentation of the aorta using CT.
[0004] Therefore, a robust 3D segmentation method and device for the aorta based on deep learning was developed to solve the above problems. Summary of the Invention
[0005] This invention proposes a robust 3D segmentation method and apparatus for the aorta based on deep learning, in order to solve the problems of low robustness and low automation of existing segmentation methods.
[0006] The present invention achieves the above objectives through the following technical solutions:
[0007] A deep learning-based robust 3D segmentation method for the aorta in CT scans, comprising:
[0008] This invention discloses a deep learning-based robust 3D segmentation method for the aorta in CT scans, comprising:
[0009] Obtain a two-dimensional slice sequence of the aorta;
[0010] The aortic structure is extracted frame by frame from the slice sequence to obtain the corresponding normalized probability map, which represents the probability that each pixel belongs to the aorta.
[0011] Linear interpolation and thresholding of the probability map yield a preliminary segmentation mask;
[0012] Set the initial segmentation mask of the current slice as the initialization prompt of the medical video segmentation model with a memory mechanism;
[0013] Memory initialization: For the initial slice image, the medical video segmentation model calls the initial memory state function to generate the preliminary segmentation result of the initial slice and encode it into the initial memory state;
[0014] Memory propagation: For subsequent slice images, the medical video segmentation model uses its memory state established in the previous slice and the current slice image to call the propagation tracking state function to automatically predict the original aortic mask of the current slice and obtain the predicted mask;
[0015] The predicted mask and the preliminary segmentation mask are weighted and fused to obtain a 2D aortic segmentation mask. Based on the 2D aortic segmentation mask generated for all slices, a three-dimensional reconstruction is performed to obtain a complete three-dimensional model of the CT aorta.
[0016] Furthermore, before extracting the aortic structure frame by frame from the slice sequence, a standardized preprocessing operation is performed on each original slice based on an inert loading mechanism.
[0017] Furthermore, the standardization preprocessing includes window width and window level adjustment, pixel value normalization, and size unification.
[0018] Furthermore, it also includes:
[0019] After processing each slice t, the tracking quality index between the predicted mask and the initial segmentation mask is calculated;
[0020] If the tracking quality index is lower than the preset threshold, or the prediction mask is an empty mask, it is determined that the medical video segmentation model has drifted or lost the target.
[0021] Automatically trigger tracking re-anchoring and memory reset: Clear the internal memory state of the medical video segmentation model and reinitialize its memory state using the initial segmentation mask, prompting the next frame to re-execute the memory initialization phase of the medical video segmentation model.
[0022] Furthermore, the tracking quality indicator is the crossover ratio (CUR) or union ratio (UWR), and the formula for calculating the CUR is as follows:
[0023] For intersection, union, and comparison, To predict the mask, This is a preliminary segmentation mask.
[0024] Furthermore, the aortic structure is extracted frame by frame from the slice sequence by using a pre-trained U2-Net model to perform preliminary frame-by-frame segmentation of each slice in the CT image sequence and extracting the aortic structure.
[0025] Furthermore, in the initial slice of the tracing process Or the internal state of the underlying model when tracking requires recalibration. Initialized, tracking whether recalibration is needed is performed by The expression for initializing memory is determined to be:
[0026]
[0027]
[0028] in, This represents the memory state established by the medical video segmentation model in the previous slice. A unique identifier representing the aorta to be tracked. This indicates the initial segmentation mask. To initialize the memory state function, A value of True indicates that the model's tracking results in the current frame t are unreliable (e.g., low segmentation accuracy, target loss), and need to be re-tracked using accurate methods. Calibrate the internal state of the model , This is the preliminary segmentation result. This is the initial memory state.
[0029] Furthermore, for t> Furthermore, subsequent slices that do not trigger reinitialization undergo memory propagation. The expression for the memory propagation process is as follows:
[0030] )
[0031]
[0032] t represents the time step corresponding to the subsequent slice. The current memory state, The current segmentation result is the final segmentation result, which is a weighted fusion of the memory state and the segmentation result. Setting it to False indicates that the model's tracking results in the current frame t are reliable, requiring no re-initialization, and incremental segmentation can continue based on the memory state of the previous frame. This is the function for propagating the tracking state.
[0033] Furthermore, the medical video segmentation model is MedSAM2.
[0034] The apparatus for the deep learning-based CT aortic robust 3D segmentation method includes:
[0035] The acquisition module is used to acquire a two-dimensional slice sequence of the aorta;
[0036] The extraction module is used to extract the aortic structure frame by frame from the slice sequence to obtain a corresponding normalized probability map, which represents the probability that each pixel belongs to the aorta.
[0037] A linear interpolation and thresholding module is used to perform linear interpolation and thresholding on the probability map to obtain a preliminary segmentation mask;
[0038] The setting module is used to set the initial segmentation mask of the current slice as the initialization prompt of a medical video segmentation model with a memory mechanism;
[0039] The memory initialization module is used to generate the initial segmentation result and initial memory state of the starting slice image by calling the initial memory state function through the medical video segmentation model.
[0040] The memory propagation module is used to automatically predict the original aortic mask of the current slice by calling the propagation tracking state function for subsequent slice images, using the memory state established by the medical video segmentation model in the previous slice and the current slice image, and to obtain the predicted mask.
[0041] The three-dimensional reconstruction module is used to perform weighted fusion of the predicted mask and the preliminary segmentation mask to obtain a 2D aortic segmentation mask. Based on the 2D aortic segmentation mask generated for all slices, three-dimensional reconstruction is performed to obtain a complete three-dimensional model of the CT aorta.
[0042] The beneficial effects of this invention are as follows:
[0043] The deep learning-based robust 3D segmentation method and device for the aorta proposed in this invention realizes fully automated and robust 3D segmentation of the aorta using CT, providing accurate and efficient image support for clinical diagnosis, treatment planning and surgical navigation. Attached Figure Description
[0044] Figure 1 This is a flowchart of the deep learning-based CT aortic robust 3D segmentation method proposed in this application. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0046] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0047] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0048] The present invention achieves the above objectives through the following technical solutions:
[0049] A deep learning-based robust 3D segmentation method for the aorta in CT scans, comprising:
[0050] like Figure 1 As shown, the present invention provides a robust 3D segmentation method for the aorta based on deep learning, comprising:
[0051] S1: Obtain the two-dimensional slice sequence of the aorta;
[0052] S2: Extract the aortic structure frame by frame from the slice sequence to obtain the corresponding normalized probability map, which represents the probability that each pixel belongs to the aorta;
[0053] S3: Perform linear interpolation and thresholding on the probability map to obtain the initial segmentation mask;
[0054] S4: Set the initial segmentation mask of the current slice as the initialization prompt of the medical video segmentation model with a memory mechanism;
[0055] S5: Memory Initialization: For the initial slice image, the medical video segmentation model calls the initial memory state function to generate the preliminary segmentation result of the initial slice and encode it into the initial memory state;
[0056] S6: Memory propagation: For subsequent slice images, the medical video segmentation model uses the memory state established in the previous few frames and the current slice image to call the propagation tracking state function to automatically predict the original aortic mask of the current slice and obtain the predicted mask.
[0057] S7: The predicted mask and the preliminary segmentation mask are weighted and fused to obtain the 2D aortic segmentation mask. Based on the 2D aortic segmentation mask generated for all slices, three-dimensional reconstruction is performed to obtain the complete three-dimensional model of the CT aorta.
[0058] In one embodiment, before extracting the aortic structure frame by frame from the slice sequence, a standardized preprocessing operation is performed on each raw slice based on an inertial loading mechanism.
[0059] In one embodiment, the standardization preprocessing includes window width and window level adjustment, pixel value normalization, and size unification.
[0060] In one embodiment, it further includes:
[0061] After processing each slice t, the tracking quality index between the predicted mask and the initial segmentation mask is calculated;
[0062] If the tracking quality index is lower than the preset threshold, or the prediction mask is an empty mask, it is determined that the medical video segmentation model has drifted or lost the target.
[0063] Automatically trigger tracking re-anchoring and memory reset: Clear the internal memory state of the medical video segmentation model and reinitialize its memory state using the initial segmentation mask, prompting the next frame to re-execute the memory initialization phase of the medical video segmentation model.
[0064] In one embodiment, the tracking quality metric is the crossover ratio (CURRY), and the CURRY is calculated using the following formula:
[0065]
[0066] For intersection, union, and comparison, To predict the mask, This is a preliminary segmentation mask.
[0067] Furthermore, the aortic structure is extracted frame by frame from the slice sequence by using a pre-trained U2-Net model to perform preliminary frame-by-frame segmentation of each slice in the CT image sequence and extracting the aortic structure.
[0068] In one embodiment, at the beginning slice of the tracing process Or the internal state of the underlying model when tracking requires recalibration. Initialized, tracking whether recalibration is needed is performed by The expression for initializing memory is determined to be:
[0069]
[0070]
[0071] in, This represents the memory state established by the medical video segmentation model in the previous slice. A unique identifier representing the aorta to be tracked. This indicates the initial segmentation mask. To initialize the memory state function, A value of True indicates that the model's tracking results in the current frame t are unreliable (e.g., low segmentation accuracy, target loss), and need to be re-tracked using accurate methods. Calibrate the internal state of the model , This is the preliminary segmentation result. This is the initial memory state.
[0072] In one embodiment, for t> Furthermore, subsequent slices that do not trigger reinitialization undergo memory propagation. The expression for the memory propagation process is as follows:
[0073] )
[0074]
[0075] t represents the time step corresponding to the subsequent slice. The current memory state, The current segmentation result is the final segmentation result, which is a weighted fusion of the memory state and the segmentation result. Setting it to False indicates that the model's tracking results in the current frame t are reliable, requiring no re-initialization, and incremental segmentation can continue based on the memory state of the previous frame. This is the function for propagating the tracking state.
[0076] In one embodiment, the medical video segmentation model is MedSAM2.
[0077] The apparatus for the deep learning-based CT aortic robust 3D segmentation method includes:
[0078] The acquisition module is used to acquire a two-dimensional slice sequence of the aorta;
[0079] The extraction module is used to extract the aortic structure frame by frame from the slice sequence to obtain a corresponding normalized probability map, which represents the probability that each pixel belongs to the aorta.
[0080] A linear interpolation and thresholding module is used to perform linear interpolation and thresholding on the probability map to obtain a preliminary segmentation mask;
[0081] The setting module is used to set the initial segmentation mask of the current slice as the initialization prompt of a medical video segmentation model with a memory mechanism;
[0082] The memory initialization module is used to generate the initial segmentation result and initial memory state of the starting slice image by calling the initial memory state function through the medical video segmentation model.
[0083] The memory propagation module is used to automatically predict the original aortic mask of the current slice by calling the propagation tracking state function for subsequent slice images, using the memory state established by the medical video segmentation model in the previous slice and the current slice image, and to obtain the predicted mask.
[0084] The three-dimensional reconstruction module is used to perform weighted fusion of the predicted mask and the preliminary segmentation mask to obtain a 2D aortic segmentation mask. Based on the 2D aortic segmentation mask generated for all slices, three-dimensional reconstruction is performed to obtain a complete three-dimensional model of the CT aorta.
[0085] The present invention will be further described below through a specific implementation process:
[0086] First, we receive a two-dimensional slice sequence acquired by a CT scan device, represented as { , …, }, where N is the total number of frames in the sequence. For optimization, a lazy loading mechanism is adopted, and for each original slice... The framework performs a standardized preprocessing operation, Standardize(⋅), which includes adjusting window width and level, normalizing pixel values, and unifying dimensions to obtain the processed slices. .
[0087] Subsequently, the processed slices Frame-by-frame processing is performed to independently and robustly extract the aortic structure and generate a high-quality, reliable binarization mask. This step ensures that even if the aorta is branched or discontinuous in the slice, the network can identify and generate all visible portions, providing comprehensive initial information for subsequent processing. This mask is obtained by analyzing the probability graph of the output. Linear interpolation and thresholding are performed to obtain stable initial information and a basis for dynamic correction in subsequent steps.
[0088]
[0089] Where R(⋅) is linear interpolation, and T(⋅,τ) is the thresholding function (threshold). Usually, 0.5 is used.
[0090] Next, a medical video segmentation model with a memory mechanism was integrated for robust tracking and refinement of the aorta in CT image sequences. The specific workflow follows a two-stage refinement paradigm:
[0091] Phase 1: Memory Initialization. Specifically, this occurs in the initial slice of the tracing process. (e.g., the first frame of the sequence) or when tracking requires recalibration (by...) (Decision), the internal state of the basic model Initialized. The current slice is binarized into a mask. This serves as an initialization hint for the underlying model. Based on this complete mask, the underlying model forces the establishment of its internal initial memory state. and position encoding By calling its initialization memory state function And generate the preliminary segmentation result of the slice. This stage ensures that the baseline model is traced from an accurate and reliable memory benchmark that includes the complete morphology of the aorta, rather than merely serving as a cue.
[0092]
[0093]
[0094] in, This represents the memory state of the medical video segmentation baseline model with a memory mechanism up to the current time step (e.g., t−1). A unique identifier representing the aorta to be tracked.
[0095] Phase Two: Memory Transmission. For t> Furthermore, no subsequent slices were reinitialized, and the base model utilizes the memory state established in the previous slice t−1. and the current slice image , through its propagation to trace the state function Automatically predict the original aortic mask logits of the current slice. This phase aims to maintain consistency in the tracking time and to continuously revise and refine the initial predictions from the first phase.
[0096]
[0097]
[0098] Furthermore, to ensure the robustness of tracking, a dynamic feedback correction and memory reset mechanism is introduced. Specifically, after processing each slice t, the predicted mask of the medical video segmentation base model with the memory mechanism is calculated. With the initial segmentation mask Tracking quality metrics (e.g., Intersection over Union (IoU)) , If the tracking quality indicator is lower than the preset threshold. ,or If the mask is empty, it indicates that the underlying model may be drifting or losing its target. In this case, a re-anchoring and memory reset will be automatically triggered: the internal memory state of the underlying model will be cleared, and a reliable mask will be forced to be used. By reinitializing its memory state, the underlying model memory initialization phase is re-executed in the next frame, thereby effectively and adaptively correcting memory drift and ensuring long-term robustness and temporal consistency of tracking.
[0099]
[0100]
[0101] Finally, the final 3D model of the aorta is generated through mask fusion and 3D reconstruction. The mask used in the propagation of the medical video segmentation basic model has a memory mechanism. With binarization mask Weighted fusion is performed to generate the final 2D aortic segmentation mask. .
[0102]
[0103] Here, α and β are the fusion weight coefficients, and they satisfy α+β=1, usually α=β=0.5.
[0104] All slices in the sequence The data is stacked to form three-dimensional voxel data. Finally, a volume rendering algorithm is used for rapid three-dimensional reconstruction to form a complete three-dimensional model of the aorta on CT, providing intuitive and accurate surgical navigation information for clinicians.
[0105] This invention first performs frame-by-frame preliminary segmentation of CT image sequences using initial two-dimensional segmentation to obtain a high-quality binary mask containing the complete morphology of the aorta. Subsequently, this complete mask is used as an initialization cue to force the establishment of an initial memory state, and a second-stage refinement is performed based on this, including video propagation, to ensure temporal consistency in aortic tracking. To further enhance robustness, a dynamic adaptive correction based on tracking quality assessment is introduced: when the propagation result quality deteriorates or the prediction is empty, its internal memory is automatically cleared and reinitialized using the mask from the initial two-dimensional segmentation, effectively correcting tracking drift. Finally, the results from both methods are fused and output through three-dimensional reconstruction. This framework achieves fully automated and robust three-dimensional segmentation of the CT aorta, providing accurate and efficient image support for clinical diagnosis, treatment planning, and surgical navigation.
[0106] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A deep learning based CT aorta robust 3D segmentation method, characterized in that, The method comprises the following steps: obtaining a two-dimensional slice sequence of an aorta; extracting the aorta structure from the slice sequence frame by frame to obtain a corresponding normalized probability map, the probability map representing the probability of each pixel belonging to the aorta; performing linear interpolation and thresholding on the probability map to obtain a preliminary segmentation mask; setting the preliminary segmentation mask of the current slice as an initialization prompt of a medical video segmentation model with a memory mechanism; memory initialization: for the starting slice image, an initialization memory state function is called through the medical video segmentation model to generate the preliminary segmentation result of the starting slice and encode it into an initial memory state; memory propagation: for subsequent slice images, the medical video segmentation model uses the memory state established in the previous few frames and the current slice image to call a propagation tracking state function to automatically predict the aorta original mask of the current slice to obtain a predicted mask; performing weighted fusion on the predicted mask and the preliminary segmentation mask to obtain a 2D aorta segmentation mask, and performing three-dimensional reconstruction on all the 2D aorta segmentation masks generated according to all the slices to obtain a complete three-dimensional model of the CT aorta.
2. The deep learning-based CT aorta robust three-dimensional segmentation method of claim 1, wherein, Before extracting the aorta structure from the slice sequence frame by frame, a standardization preprocessing operation is performed on each original slice based on a lazy loading mechanism.
3. The deep learning-based CT aorta robust three-dimensional segmentation method of claim 2, wherein, The standardization preprocessing includes window width and window level adjustment, pixel value normalization, and size unification.
4. The deep learning based CT aorta robust 3D segmentation method of claim 1, wherein, Further comprising: After processing each slice t, a tracking quality indicator between the predicted mask and the preliminary segmentation mask is calculated; if the tracking quality indicator is lower than a preset threshold, or the predicted mask is a null mask, it is judged that the medical video segmentation model has drifted or lost the target; automatically triggering re-anchoring and memory resetting of tracking: emptying the internal memory state of the medical video segmentation model and reinitializing its memory state using the preliminary segmentation mask to prompt the next frame to re-execute the memory initialization stage of the medical video segmentation model.
5. The deep learning based CT aorta robust 3D segmentation method of claim 4, wherein, The tracking quality indicator is the intersection over union, and the calculation formula of the intersection over union is as follows: ， is the intersection over union, is the prediction mask, is the preliminary segmentation mask.
6. The deep learning based CT aorta robust three-dimensional segmentation method of claim 1, wherein, extracting the aorta structure from the slice sequence frame by frame includes using a pre-trained U2-Net model to perform preliminary segmentation on each slice in the CT image sequence frame by frame to extract the aorta structure.
7. The deep learning based CT aorta robust three-dimensional segmentation method of claim 1, wherein, at the start slice of the tracking process or when the tracking needs recalibration, the internal state of this base model is initialized, whether the tracking needs recalibration is decided by the expression memorized for the initialization is: ， ， wherein, represents the memory state of the medical video segmentation model established on the previous slice, represents a unique identifier of the aorta to be tracked, represents a preliminary segmentation mask, is an initialization memory state function, is True, indicating that the tracking result of the model at the current frame t is unreliable (such as too low segmentation accuracy, target loss), and needs to be recalibrated with accurate calibrate the internal state of the model , is a preliminary segmentation result, is an initial memory state.
8. The deep learning-based CT aorta robust three-dimensional segmentation method of claim 7, wherein, for t and no subsequent slice of the reinitialization is triggered, a memory propagation is performed, the expression of the memory propagation process being as follows: )， ， t is the time step corresponding to the subsequent slice, is the current memory state, is the current segmentation result, and the final segmentation result is a weighted fusion of the memory state and the segmentation result, is False, indicating that the tracking result of the model at the current frame t is reliable, and there is no need to reinitialize, and the incremental segmentation can be continued based on the memory state of the previous frame, is the propagation tracking state function.
9. The deep learning based CT aorta robust three-dimensional segmentation method of claim 1, wherein, The medical video segmentation model is MedSAM2.
10. The device for the robust three-dimensional segmentation of the CT aorta based on deep learning according to any one of claims 1-9, characterized in that, The method comprises the following steps: an acquisition module for acquiring a two-dimensional slice sequence of an aorta; an extraction module for extracting the aorta structure from the slice sequence frame by frame to obtain a corresponding normalized probability map, the probability map representing the probability of each pixel belonging to the aorta; a linear interpolation and thresholding module for performing linear interpolation and thresholding on the probability map to obtain a preliminary segmentation mask; a setting module for setting the preliminary segmentation mask of the current slice as an initialization prompt of a medical video segmentation model with a memory mechanism; a memory initialization module for calling an initialization memory state function through the medical video segmentation model for the starting slice image to generate the preliminary segmentation result of the starting slice and an initial memory state; a memory propagation module, configured to, for a subsequent slice image, call a propagation tracking state function to automatically predict an aorta original mask of the current slice by a medical video segmentation model using a memory state established by the medical video segmentation model for a previous slice and the current slice image, to obtain a predicted mask; a three-dimensional reconstruction module, configured to perform weighted fusion on the predicted mask and the preliminary segmentation mask to obtain a 2D aorta segmentation mask, and perform three-dimensional reconstruction on all slice corresponding generated 2D aorta segmentation masks to obtain a complete three-dimensional model of the CT aorta.