Deep learning based abdominal aortic aneurysm hemodynamics simulation method and system

Abstract

This invention provides a deep learning-based method and system for simulating the hemodynamics of abdominal aortic aneurysms. The segmented image is input into an inverted pyramid structured ATRU-Net network to obtain a 3D label map containing spatial feature information. The parameters of the ATRU-Net network are updated using a loss function. 3DCRF combined with the original image is used to regularize the 3D label map, obtaining the intraluminal geometry of the connected abdominal aortic aneurysm. 3DCCO is used to perform connected component search to obtain a 3D binary mask of the blood vessel. Hemodynamic results are calculated after data processing of the 3D binary mask using a surface reconstruction algorithm combined with boundary conditions. Advantages include: using attention gating to quickly locate the most prominent object in a cluttered visual scene, improving computational efficiency without additional supervision; utilizing long and short jump connections to retain richer spatial information, improving sensitivity to smaller aneurysms and blood vessels, and enhancing the ability to acquire small vessel features; and shortening model reconstruction time to reduce the subjective limitations of manual image segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis, and in particular to a method and system for simulating the hemodynamics of abdominal aortic aneurysm based on deep learning. Background Technology

[0002] Abdominal aortic aneurysm (AAA) is a common vascular disease, typically referring to a localized dilation of the abdominal aorta with a diameter exceeding 1.5 times the normal value. The main danger of this disease is massive bleeding due to aneurysm rupture, a life-threatening acute condition. Furthermore, abdominal aortic aneurysms can lead to complications such as thrombosis, embolism, and renal insufficiency, severely impacting patients' quality of life and prognosis. The exact causes of abdominal aortic aneurysms are not fully understood, but they are associated with factors such as hypertension, smoking, high cholesterol, and arteriosclerosis. Early symptoms are often subtle and easily overlooked or misdiagnosed until the aneurysm enlarges or ruptures, at which point obvious symptoms such as abdominal pain, back pain, and an abdominal mass may appear. Therefore, early diagnosis and treatment are crucial for preventing and controlling the development of abdominal aortic aneurysms.

[0003] Currently, the diagnosis of abdominal aortic aneurysms primarily relies on medical imaging techniques such as CT and MRI. Doctors need to analyze and interpret the images to determine the size, shape, and location of the aneurysm in order to develop an appropriate treatment plan. However, due to the complexity and diversity of medical imaging data, doctors' judgments and interpretations are subject to certain subjectivity and limitations, easily leading to misdiagnosis and missed diagnosis. Therefore, developing automated medical image analysis algorithms to improve diagnostic accuracy and efficiency is of great significance for preventing and controlling the development of abdominal aortic aneurysms.

[0004] In the field of medical image analysis, traditional image segmentation methods mainly rely on prior knowledge, requiring manual annotation and model input. The drawbacks of this method are the significant need for manual intervention and time, and the accuracy and stability of the results are affected by the operator's experience and skill level. Furthermore, due to the subjectivity and limitations of manual annotation, this method has certain limitations when processing complex medical images.

[0005] In recent years, the development of deep learning technology has provided new solutions for medical image analysis. Deep learning algorithms can automatically learn features, thereby reducing reliance on prior knowledge and improving the accuracy and stability of image segmentation. Among deep learning algorithms, the U-Net structure is a commonly used convolutional neural network that can perform image segmentation and feature extraction simultaneously, exhibiting good performance and efficiency. However, existing deep learning algorithms still face some problems and challenges in the field of medical image analysis. First, deep learning algorithms require a large amount of labeled data for training, but the acquisition and labeling of medical image data is costly, thus limiting the size and quality of datasets. Second, deep learning algorithms are complex and computationally intensive, requiring high-performance computing equipment and algorithm optimization; otherwise, training time will be excessively long and model performance will degrade.

[0006] Therefore, how to provide an image processing method that can automatically annotate and is highly sensitive to smaller aneurysms and blood vessels, and has a strong ability to acquire small blood vessel features has become an urgent problem to be solved. Summary of the Invention

[0007] This invention provides a deep learning-based method and system for simulating the hemodynamics of abdominal aortic aneurysms, which addresses the problems of high annotation costs and low annotation efficiency in the annotation of small aneurysms and blood vessels in existing technologies.

[0008] To achieve the above objectives, the present invention provides a deep learning-based method for simulating the hemodynamics of abdominal aortic aneurysm, comprising: inputting the segmented image into an ATRU-Net network with an inverted pyramid structure to obtain a three-dimensional label map containing spatial feature information; specifically, in the ATRU-Net network, the spatial feature information output by each layer of the previous layer and the spatial feature information output by the first computing unit of the current layer are input together into the attention gate of the current layer and then the spatial feature information of the current layer is output. A loss function is set, which calculates the current loss value based on the feature map and corresponding labeled data output by each layer of the ATRU-Net network, and then updates the parameters of the ATRU-Net network through backpropagation; The three-dimensional label map was regularized by combining the fully connected three-dimensional conditional random field 3DCRF with the original image to obtain the processed image, so as to suppress unreasonable adhesions, repair discontinuous vascular segments, and obtain the intraluminal geometry of the abdominal aortic aneurysm. The 3DCCO (Three-Dimensional Connectivity Component Optimization) method is used to perform connected component search on the processed image, retaining the largest connected blood vessel cluster, and obtaining the final 3D binary blood vessel mask. Boundary conditions are set, and data processing is performed on the three-dimensional binary mask of the blood vessel using a surface reconstruction algorithm to obtain the surface of the blood vessel lumen. Then, a CFD computation mesh is generated by dividing the volume mesh using the finite volume method, and the hemodynamic results of the blood vessel corresponding to the three-dimensional binary mask of the blood vessel are calculated.

[0009] As a preferred embodiment of the above technical solution, the inverted pyramid structure of the ATRU-Net network is specifically as follows: the structure of the ATRU-Net network is an inverted pyramid-shaped 4-layer architecture; each layer has a dimensionally symmetrical convolutional layer and a parameter correction linear unit group, and the convolutional layers and parameter correction linear unit groups are arranged in an inverted pyramid arrangement from low resolution to high resolution. Each convolutional layer and parameter correction linear unit group has an attention gate between the two convolutional layers and parameter correction linear units.

[0010] As a preferred embodiment of the above technical solution, the ATRU-Net network preferably includes an encoding path and a decoding path: In the encoding path, the output of the first convolutional layer and parameter correction linear unit of the low-resolution layer is used as the input of the first convolutional layer and parameter correction linear unit of the next higher resolution layer, so as to realize the layer-by-layer transfer of features from low resolution to high resolution. In the decoding path, the features processed by the second convolutional layer and parameter correction linear unit of the highest resolution layer are taken as the starting point. After being fused with the output of the second convolutional layer and parameter correction linear unit of the adjacent high resolution layer through the attention gate, they are sequentially input into the second convolutional layer and parameter correction linear unit of the next low resolution layer. This realizes the feature backpropagation from high resolution to low resolution layer by layer, thereby obtaining a 3D label map containing spatial feature information.

[0011] As a preferred embodiment of the above technical solution, the attention gates are configured such that: the attention gate uses the feature map of the previous convolutional layer as a gate signal to optimize the feature map obtained by encoding the gate signal and the feature map of the same layer; specifically, by mapping the concatenated feature map and gate signal to the intermediate space, different activation maps are obtained sequentially through different activation operations, and the nonlinear layer corresponding to the next activation map is subjected to trilinear interpolation resampling to finally obtain the final output of the current attention gate.

[0012] As a preferred embodiment of the above technical solution, preferably, the loss function calculates the current loss value based on the feature map output by each layer of the ATRU-Net network and the corresponding labeled data, and then updates the parameters of the ATRU-Net network through backpropagation, including: In the 3D label map containing spatial feature information output by each layer of the ATRU-Net network, the predicted pixel set of individual voxels belonging to the blood vessel category and the labeled pixel set in this 3D label map are extracted. The loss function calculates the loss value based on the predicted pixel set and the labeled pixel set, combined with the false negative penalty magnitude and the false positive penalty magnitude.

[0013] As a preferred embodiment of the above technical solution, preferably, a fully connected three-dimensional conditional random field (3DCRF) combined with the original image is used to perform regularization processing on the three-dimensional label image to obtain a processed image, thereby suppressing unreasonable adhesions, repairing discontinuous vascular segments, and obtaining the intraluminal geometry of the abdominal aortic aneurysm, including: Based on 3DCRF combined with the original image, the three-dimensional label map is spatially regularized by minimizing the Gibbs energy. Specifically, unreasonable adhesion is suppressed by the spatial distance between adjacent voxels and the gray-level difference between adjacent voxels, and discontinuous blood vessel segments are repaired.

[0014] As a preferred embodiment of the above technical solution, preferably, the 3D Connectivity Component Optimization (3DCCO) method is used to perform connected component search on the processed image, retaining the largest connected vessel cluster to obtain the final 3D vessel binary mask, including: Starting with any vessel label in the binary 3D volume of the image after 3DCRF regularization, 3DCCO recursively searches for all adjacent vessel labels and adds them to the group associated with the first selected vessel label. This process continues until all vessel labels connected to the first selected vessel label are found. If there are vessel labels without a group association, another arbitrary vessel label is initialized, and the region growing process restarts. After processing all vessel labels, at least one connected vessel group is obtained. The vessel group with the most vessel labels is selected as the final segmentation result, resulting in the final 3D binary vessel mask.

[0015] As a preferred embodiment of the above technical solution, preferably, boundary conditions are set, including: An instantaneous volumetric flow rate waveform is applied to the inlet section determined based on the three-dimensional blood vessel binary mask segmentation results, thereby setting the relationship between blood flow velocity or blood flow rate and time. Export boundary conditions: Zero pressure conditions are applied to all outlets, representing the reference value of peripheral pressure; Blood vessel wall boundary conditions: Apply no-slip boundary conditions to the surface of the blood vessel wall formed by the fluid domain, while assuming that the blood vessel wall is rigid and does not deform; Physical properties: The density and dynamic viscosity of blood are uniformly set throughout the entire fluid domain; based on this, hemodynamic results such as local pressure, flow velocity, and wall shear stress are calculated.

[0016] This invention also provides a deep learning-based abdominal aortic aneurysm hemodynamic simulation system, comprising: a spatial feature acquisition unit, an image feature optimization unit, and a hemodynamic result acquisition unit. The spatial feature acquisition unit is used to input the segmented image into the ATRU-Net network with an inverted pyramid structure to obtain a 3D label map containing spatial feature information. Specifically, the spatial feature information output by each layer of the ATRU-Net network from the previous layer and the spatial feature information output by the first computation unit of the current layer are input together into the attention gate of the current layer to output the spatial feature information of the current layer. A loss function is set, which calculates the current loss value based on the feature map output by each layer of the ATRU-Net network and the corresponding labeled data, and then updates the parameters of the ATRU-Net network through backpropagation. The image feature optimization unit is used to perform regularization processing on the three-dimensional label map using a fully connected three-dimensional conditional random field (3DCRF) combined with the original image to obtain a processed image, in order to suppress unreasonable adhesions, repair discontinuous vascular segments, and obtain the intraluminal geometry of the abdominal aortic aneurysm; and to perform connected component optimization (3DCCO) on the processed image to search for connected components, retain the largest connected vascular cluster, and obtain the final three-dimensional vascular binary mask. The hemodynamic results acquisition unit is used to generate data processing of the three-dimensional blood vessel binary mask by setting boundary conditions and using a surface reconstruction algorithm to obtain the surface of the blood vessel lumen, and then generate a CFD calculation mesh by using the finite volume method to calculate the hemodynamic results of the blood vessel corresponding to the three-dimensional blood vessel binary mask.

[0017] This invention provides a deep learning-based method for simulating the hemodynamics of abdominal aortic aneurysms. The segmented image is input into an ATRU-Net network with an inverted pyramid structure to obtain a 3D label map containing spatial feature information. The parameters of the ATRU-Net network are updated using a loss function. 3DCRF is used in conjunction with the original image to regularize the 3D label map, resulting in a processed image that reveals the intraluminal geometry of the abdominal aortic aneurysm. The 3DCCO method is used to perform connected component search on the processed image to obtain a 3D binary vascular mask. Boundary conditions are set, and a surface reconstruction algorithm is used to generate and process the 3D binary vascular mask to obtain the vascular lumen surface. CFD is then used to calculate the hemodynamic results of the vessel corresponding to the 3D binary vascular mask.

[0018] The first feature of this invention is the construction of an ATRU-Net network with an inverted pyramid structure, and the introduction of a multi-scale attention gate between the encoding and decoding paths, which enables the synergistic use of shallow spatial detail features and deep semantic features, significantly improving segmentation accuracy and boundary characterization ability for small-diameter branch vessels and irregularly shaped aneurysm neck regions in abdominal aortic aneurysms.

[0019] The second feature of this invention is that it adopts an asymmetric loss function based on Tversky loss during the network training stage, which imposes a higher weight penalty on false negatives. This allows the model to prioritize recall of small blood vessels and tumor edge regions in scenarios with high class imbalance, thus mitigating the problem of small target structures being ignored.

[0020] The third feature of this invention is that it combines fully connected three-dimensional conditional random field (3DCRF) with three-dimensional connectivity component optimization (3DCCO) for post-processing. By combining energy minimization and connected component screening, it not only suppresses unreasonable adhesion between the aneurysm cavity and adjacent tissues, but also removes isolated noise clumps, thereby obtaining a three-dimensional vascular geometric model with a more reasonable topology that can be directly used for subsequent CFD calculations.

[0021] The advantages of this invention are that it enables rapid localization of the most prominent object in cluttered visual scenes by setting attention gates, without the need for additional supervision. Utilizing long-hop and short-hop connections preserves richer spatial information, helping the network integrate multi-level spatial and semantic information, thereby improving sensitivity to smaller aneurysms and blood vessels, and enhancing the ability to acquire features of small blood vessels. It significantly shortens the reconstruction time of the AAA model, and automatic image segmentation greatly reduces the subjectivity and limitations of manual segmentation. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a deep learning-based method for simulating abdominal aortic aneurysm hemodynamics provided by this invention. Figure 1 .

[0024] Figure 2 A flowchart illustrating a deep learning-based method for simulating abdominal aortic aneurysm hemodynamics provided by this invention. Figure 2 .

[0025] Figure 3 for Figure 2The flowchart for step 202.

[0026] Figure 4 This is a rough flowchart illustrating a deep learning-based method for simulating the hemodynamics of an abdominal aortic aneurysm, as provided by the present invention.

[0027] Figure 5 This is a schematic diagram of the ATRU-Net network structure in this invention.

[0028] Figure 6 A diagram of attention gates in the ATRU-Net network Figure 1 .

[0029] Figure 7 A diagram of attention gates in the ATRU-Net network Figure 2 .

[0030] Figure 8 This is a waveform diagram of pulsating flow rate.

[0031] Figure 9 This is a schematic diagram of the structure of a deep learning-based abdominal aortic aneurysm hemodynamic simulation system provided by the present invention.

[0032] Figure 10 The present invention provides a reference flowchart for a deep learning-based method and system for simulating abdominal aortic aneurysm hemodynamics. Detailed Implementation

[0033] 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Figure 1 This is a schematic diagram of a process provided for an embodiment of the present invention, such as... Figure 1 As shown, it includes: Step 101: Input the segmented image into the ATRU-Net network to obtain a 3D label map containing spatial feature information.

[0035] The ATRU-Net network structure is an inverted pyramid, with dimensions increasing from the top to the bottom. Specifically, each layer of the inverted pyramid has dimensionally symmetrical convolutional layers and parameter-correcting linear units (PCUs). The convolutional layers and PCUs are arranged in an inverted pyramid pattern from low to high resolution. The attention gate in each layer is located between two separate units within the convolutional layer and PCU group. During data processing, the spatial feature information output by the first computational unit of each layer is simultaneously input into the higher-precision first computational unit of the next layer and the attention gate of the current layer. The attention gate of the current layer processes the data result of the current layer or processes the data result of the current layer and the spatial feature information output by the higher-precision second computational unit before outputting it to the second computational unit of the lower-precision convolutional layer and PCU group, and so on, outputting a 3D label map containing spatial feature information.

[0036] The above description specifically refers to the encoding and decoding paths of the ATRU-Net network: In the encoding path, the output of the first convolutional layer and parameter correction linear unit of the low-resolution layer is used as the input of the first convolutional layer and parameter correction linear unit of the next higher resolution layer, so as to realize the layer-by-layer transfer of features from low resolution to high resolution.

[0037] In the decoding path, the features processed by the second convolutional layer and parameter correction linear unit of the highest resolution layer are taken as the starting point. After being fused with the output of the second convolutional layer and parameter correction linear unit of the adjacent high resolution layer through the attention gate, they are sequentially input into the second convolutional layer and parameter correction linear unit of the next low resolution layer. This realizes the feature backpropagation from high resolution to low resolution layer by layer, thereby obtaining a 3D label map containing spatial feature information.

[0038] Step 102: Set the loss function and update the parameters of the ATRU-Net network through backpropagation.

[0039] Specifically, the loss function calculates the current loss value based on the feature map and corresponding labeled data output by each layer of the ATRU-Net network, and then updates the parameters of the ATRU-Net network through backpropagation: In the 3D label map containing spatial feature information output by each layer of the ATRU-Net network, the predicted pixel set of voxels belonging to the blood vessel category and the labeled pixel set in this 3D label map are extracted. Then, the loss function calculates the loss value based on the predicted pixel set and the labeled pixel set, combined with the false negative penalty magnitude and the false positive penalty magnitude, and updates the parameters of the ATRU-Net network based on the loss value.

[0040] Step 103: Use 3DCRF combined with the original image to perform regularization processing on the 3D label map.

[0041] Specifically, a fully connected 3D conditional random field (3DCRF) combined with the original image is used to process the 3D label map to suppress unreasonable adhesions, repair discontinuous vascular segments, and obtain the intraluminal geometry of the connected abdominal aortic aneurysm. Based on 3DCRF combined with the original image, the 3D label map is spatially regularized by minimizing the Gibbs energy. Specifically, unreasonable adhesions are suppressed by the spatial distance between adjacent voxels and the gray-level difference between adjacent voxels, so as to repair discontinuous vascular segments.

[0042] Step 104: Use 3DCCO to perform connected component search on the regularized image, retain the largest connected vessel cluster, and obtain the final three-dimensional vessel binary mask.

[0043] 3D Connectivity Component Optimization (3DCCO) is used to start with any vessel label in the binary 3D volume of the image after 3DCRF regularization. It recursively searches for all adjacent vessel labels and adds them to the group associated with the first selected vessel label. This process continues until all vessel labels connected to the first selected vessel label are found. If a vessel label does not have a group association, another arbitrary vessel label is initialized, and the region growing process restarts. After processing all vessel labels, at least one connected vessel group is obtained. The vessel group with the most vessel labels is selected as the final segmentation result, yielding the final 3D binary vessel mask.

[0044] Step 105: Set boundary conditions and combine CFD to calculate the hemodynamic results of the blood vessels corresponding to the three-dimensional binary mask.

[0045] Specifically, after setting boundary conditions, a surface reconstruction algorithm is used to generate data processing for the three-dimensional binary mask of the blood vessel, obtaining the surface of the blood vessel lumen. Then, a volume mesh is generated using the finite volume method to generate a CFD computation mesh, and the hemodynamic results of the blood vessel corresponding to the three-dimensional binary mask are calculated. Specifically, for the boundary conditions: an instantaneous volumetric flow rate waveform is applied to the inlet section determined based on the segmentation results of the three-dimensional binary mask, thereby setting the relationship between blood flow velocity or blood flow rate and time. For the outlet boundary conditions: a zero-pressure condition is applied to all outlets, representing the peripheral pressure reference value. For the vessel wall boundary conditions: a no-slip boundary condition is applied to the surface of the vessel wall enveloped by the fluid domain, while assuming the vessel wall is rigid and does not deform. For the physical parameters: the density and dynamic viscosity of blood are uniformly set throughout the entire fluid domain; thereby calculating local pressure, flow velocity, wall shear stress, and other hemodynamic results.

[0046] The embodiments of the present invention will now be described in detail with reference to specific implementation methods and processes, such as... Figure 2 , Figure 4 Combination Figure 10 As shown, specifically: Step 201: Segment the CTA image data to be identified.

[0047] CTA data (Coronary Computed Tomography Angiography) are stored in Digital Imaging and Communications in Medicine (DICOM) format.

[0048] CTA data has an in-plane resolution of 512×512, with the number of slices ranging from 214 to 2433, representing a slice thickness of 0.5-1.25 mm / slice. The in-plane resolution of CTA data also varies from 0.75 × 0.75 mm to 0.95 × 0.95 mm.

[0049] Step 202: Input the segmented image into the ATRU-Net network to obtain a three-dimensional label map of spatial features.

[0050] Similar to the standard U-Net segmentation network, the ATRU-Net network in this application consists of encoding and decoding paths; each path has four pyramid layers with different resolutions, and all layers output independent predictions. In both the encoding and decoding paths, each layer contains two or three 3D convolutions, followed by a parametric rectified linear unit (PReLU). These PReLUs are symmetrically arranged in each layer, forming a group of convolutional layers and PReLUs. Mathematically, PReLU(x) = max(0,x) + a × min(0,x), which improves the model's accuracy. The overall architecture of the ATRU-Net algorithm is shown below. Figure 5 As shown. For ease of description Figure 5 The parametric integer linear unit of the convolutional layer and the parametric correction linear unit group on the left side of the middle layer is called the first computational unit, and the parametric integer linear unit symmetrical to it in the same layer is called the second computational unit.

[0051] like Figure 3 Combination Figure 5 As shown: Step 2021: The segmented image is successively downsampled and convolved by the first computational unit of each layer.

[0052] Specifically, the segmented image (original) is input into the first computing unit of the first layer, and the first computing unit outputs processed image A; processed image A is input into the first computing unit of the second layer, and the first computing unit outputs processed image B; processed image B is input into the first computing unit of the third layer, and the first computing unit outputs processed image C; processed image C is input into the first computing unit of the fourth layer, and the first computing unit outputs processed image D.

[0053] Step 2022: Further spatial feature extraction is performed through attention gates to obtain a spatial feature map.

[0054] Further spatial feature extraction refers to performing spatial feature extraction on the processed images output from this layer (processed image A / processed image B / processed image C / processed image D) and the processed images output from the second computing unit of the next layer (processed image a / processed image b / processed image c / processed image d).

[0055] Specifically, attention gates (AG) and the introduction of long jump connections can provide features related to small blood vessels in advance.

[0056] like Figure 6 As shown, attention gates enable the ATRU-Net model to emphasize relevant spatial information from feature maps at multiple scales and then propagate it to the decoding path.

[0057] Upsampling graph in the upsampling (decoding) path The depth features in the data are used as gate signals. (in each pixel) The above is represented as (to represent the focal region) to optimize the feature map generated during the downsampling (encoding) process. This process suppresses feature responses in irrelevant background regions (i.e., through selective activation, such as...). Figure 6 (As shown). The feature map sampled above the attention gate. For reference, a gate signal is generated at each pixel i through linear mapping and activation operations. (0~1), This is used to indicate the degree to which a pixel belongs to the focal region. Subsequently, the gate signal... With feature map Pixel-by-pixel multiplication, for The regions associated with small blood vessels are enhanced, while the background regions are suppressed.

[0058] Specifically, regarding the four scales of this invention, the processed images (a / b / c / d) in step 2021 correspond to different scales. ,each The focal area of ​​each pixel is used as The feature maps generated during the downsampling (encoding) process correspond to the processed images (A / B / C / D) in step 2021, each corresponding to different scales. .

[0059] By performing convolution, normalization, and activation operations on the feature map output by the encoder, the gate signal of the attention gate is obtained in the network layer with the lowest resolution. During the upsampling convolution process in the decoder (the second computation unit of each layer), a lattice-like structure appears at the edges of the feature map (see...). Figure 3 In These grid structures are also known as "checkerboard artifacts." By setting the activation operation in the attention gate, these background artifacts can be largely eliminated. Figure 7 In ).

[0060] Specifically, in combination Figure 5 As shown, the processed image A generated by the downsampled convolution of the first computation unit in the first layer is output to the first computation unit in the second layer, which has higher precision and is connected to the same-layer attention gate. The first computation unit in the second layer performs downsampled convolution on processed image A to generate processed image B, which is then output to the first computation unit in the third layer, which has higher precision and is connected to the same-layer attention gate. The first computation unit in the third layer performs downsampled convolution on processed image B to generate processed image C, which is then output to the first computation unit in the fourth layer, which has higher precision and is connected to the same-layer attention gate. The first computation unit in the fourth layer performs downsampled convolution on processed image C to generate processed image D, which is then output to the same-layer attention gate. The attention gate in the fourth layer uses the focal region of processed image D obtained in this layer as... The processed image D is used as the feature map of the attention gate in this layer. The processed image d is obtained. After downsampling convolution, upsampling convolution is performed: the attention gate of the third layer processes the image C (as...). The depth features in the data were used as gate signals. To optimize the feature maps generated during the downsampling process. (Processing image d) yields the output result, which is then combined with the processed image d directly transmitted from the second computational unit of the fourth layer to the second computational unit of the third layer to obtain the processed image c of the third layer. The attention gate of the second layer processes image B (as... The depth features in the data were used as gate signals. To optimize the feature maps generated during the downsampling process. (Processing image c) yields the output result, which is then combined with the processed image c directly transmitted from the second computational unit of the third layer to the second computational unit of the second layer to obtain the processed image b of the third layer. The attention gate of the first layer processes image A (as... The depth features in the data were used as gate signals. To optimize the feature maps generated during the downsampling process. (Processing image b) yields the output result, which is then combined with the processed image b directly transmitted from the second computational unit of the second layer to the second computational unit of the first layer to obtain the first-layer processed image a. The final processed image a is the output result of the ATRU-Net model.

[0061] After attention gates were applied to all four pyramid layers, the background artifacts of the lattice structure gradually diminished until they disappeared completely. This effectively highlighted features crucial for the segmentation of abdominal aortic aneurysms.

[0062] Detailed combination Figure 7 As shown: By connecting the features and linear mapping Calculate the first intermediate activation graph in the intermediate space. Second intermediate activation graph ,Right now: in, and These represent the ReLU and Sigmoid activation functions, respectively. , and These are all weight parameters used in the linear transformation process, which are obtained by performing a 1×1×1 convolution on the channels of the input tensor. and All are bias terms.

[0063] exist Figure 7 middle, and There are two non-linear layers that selectively activate attention coefficients in the sequence. Then, the attention coefficients are... Perform trilinear interpolation resampling.

[0064] The final output of the attention gate is , is a feature map Compared with upsampling plot The product is shown below: Four algebraic structures (AGs) were used to process shallow and deep features. The processing results of the four AGs at different (spatial) scales are shown below. Figure 7 As shown. Typically, attention maps have larger and smaller values ​​in the target vessel and background regions, respectively. Therefore, as... Figure 7The attention gating process shown can suppress background noise. Furthermore, AG can improve the accuracy of abdominal aortic aneurysm segmentation by more directionally propagating information across different (spatial) scales.

[0065] In summary, the ATRU-Net network makes predictions independently at each pyramid layer, using the sigmoid function to predict layer by layer, and then sums the predictions from all layers to obtain the final segmentation probability. Since the feature maps generated by the network at different spatial resolutions have significant semantic differences due to their varying depths, predictions using a single high-resolution feature, with its weak semantics, inevitably lead to the loss of information about small blood vessels. In contrast, the multi-scale supervision strategy proposed in this algorithm can integrate low-resolution features with strong semantics and high-resolution features with weak semantics, combining coarse-grained and fine-grained high-density predictions, thereby enhancing the representational ability of small blood vessel segmentation and providing a more accurate description of irregularly shaped AAA (autovascular angiography).

[0066] Based on the above process, long-skip connections are introduced through attention gates (AG) to pre-provide features related to small blood vessels. This allows the network to gradually represent possible features at different spatial scales and orientations more richly as the number of layers in the encoder-decoder structure increases. This solves the problem that in the process of acquiring rich representation information, these high-level (i.e., coarse spatial resolution) feature maps lose spatial details, leading to false detections of small objects. It also solves the problem that it is difficult to reduce false positive detections of small objects exhibiting large shape changes at coarse resolution.

[0067] Step 2023: Obtain the binary pixel set and the true value binary pixel set from the spatial feature map output by ATRU-Net.

[0068] ATRU-Net receives abdominal aortic aneurysm CTA body data as input. After forward inference via an encoder-decoder structure and an attention gating module, it outputs different segmentation probability maps at each pyramid layer. The predicted probability of the i-th voxel belonging to the vessel category is denoted as p. i p i The loss function represents the predicted set of binary pixels.

[0069] With p i Correspondingly, with p i The category label of a voxel at the same location in a manually labeled gold standard segmentation mask is denoted as q. i (Vascular vessels are represented by 1, and non-vascular vessels by 0).

[0070] Step 2024: Update the parameters of ATRU-Net through backpropagation of the loss function.

[0071] Tversky loss function Represented as: Among them, the loss function Directly derived from the output feature map of ATRU-Net, Based on the labeled data for ATRU-Net in the training set, Represents the predicted set of binary pixels. Represents the true binary pixel set, in the predicted binary pixel set. In the middle, the pixel values ​​of the blood vessel area A value of 1 indicates a pixel value in avascular regions. The hyperparameters α and β are set to 0. They are pre-defined based on the imbalance between blood vessel and background samples in the abdominal aortic aneurysm dataset. They are used to increase the penalty for false negatives (FNs) during gradient backpropagation and prioritize improving the recall of small blood vessels and aneurysm edge regions. These two parameters are adjustable weights that are independent of the network structure.

[0072] Hyperparameters and hyperparameters The penalties for false positives (FPs) and false negatives (FNs) are adjusted separately. Since more emphasis is placed on FNs, the larger penalties in the loss function are applied. It will increase the weight of recall rather than prediction rate.

[0073] In this invention, when training on highly imbalanced data, the weight of false negatives is higher than that of false positives, which is beneficial for highlighting critical small blood vessels. During training, the generalization loss function... The larger the value, the higher the generalization rate and the better the performance on imbalanced data. This training strategy can effectively shift the focus of prediction to reducing false negatives and improving recall, thereby avoiding missed detections of small blood vessels to some extent.

[0074] Therefore, the probability map {p} output by the ATRU-Net network in each iteration i} and annotation mask {q i The values ​​are then substituted into the Tversky loss function to calculate the loss value for the current batch, and the ATRU-Net network parameters are updated through backpropagation. Thus, the Tversky loss function tightly couples the ATRU-Net output with the ground truth annotations during training, achieving targeted optimization for imbalanced small vessel segmentation tasks.

[0075] Therefore, this invention addresses the data imbalance problem in image segmentation using the Tversky loss function. In the presence of small blood vessel structures, the loss function achieves a relatively better trade-off between precision and recall.

[0076] Step 203: Optimize the 3D label map predicted by ATRU-Net by using 3DCRF combined with the original image.

[0077] Specifically, after the initial ATRU-Net model prediction, a 3D label map is obtained. Although the generated result has a relatively complete vascular structure, the soft segmentation map obtained by the network is often relatively smooth because adjacent pixels share a large number of similar spatial receptive fields, and boundary pixels can be identified as foreground or background. Therefore, modeling adjacent target regions can also lead to unnecessary adhesions between adjacent blood vessels or between viscera and their adjacent blood vessels, because the distance between blood vessels and their surrounding environment is very small, making them easy to predict as vascular regions. This step aims to solve this problem.

[0078] Specifically, a 3DCRF-based framework is used to regularize the initial binary mask for segmentation, making the resulting 3D label map more spatially coherent. The aforementioned binary mask refers to the 3D label map output by the ATRU-Net segmentation network for AAACTA volume data.

[0079] 3DCRF takes the original image I and the initial binary mask X as input, and obtains the optimized label field x by minimizing the Gibbs energy E(X=x|I). ∗ The tags that are optimized away or changed in this step are voxel tags that have a high cost in the energy function.

[0080] In the aforementioned steps, ATRU-Net outputs the probability P(p) that each voxel belongs to a blood vessel / aneurysm cavity. i |I), and then thresholding is used to obtain the initial binary mask X={p i (Vascular = 1, Background = 0). Specifically, one type is a "bridge-like" region with a width of only 1-2 voxels connecting adjacent blood vessels or viscera to the mother artery; the other type is small, isolated positive voxels with large spatial intensity differences or large distances from the main vascular cluster. These two types of voxels are often relabeled from 1 (blood vessel) to 0 (background) after 3DCRF optimization. Here, 1 for blood vessel and 0 for background corresponds to the random field. Label variable x i The value of x is defined i ∈{0,1}.

[0081] In the following formula, The segmentation label (original image) for the 3D input image. For the pre-determined segmentation labels, the 3DCRF model ( , The Gibbs energy in () is calculated using the following formula: in, and Representing random fields respectively and test images The number of pixels in the image. Unidimensional potential energy. It is the probability output of ATRU-Net. The results are as follows, among which This is the 3D feature map output by ATRU-Net. In the above equation, the paired potential energy... The definition is as follows: in, This is represented by the Bode model. The resulting label compatibility function provides compatibility between different adjacent pixel pairs with different labels. Indicates like and The spatial distance between them, and This represents the intensity difference between them in the input image. Hyperparameters , and Used to adjust the nearest neighbor and similarity in the loss function.

[0082] in, and These are non-negative weighting coefficients used to adjust the relative contributions of the two types of Gaussian kernels in the paired potential energy, specifically: coefficient The first weighted term depends on the spatial distance between pixels. This is used to encourage consistent labeling among neighboring pixels, achieving spatial smoothness in the overall segmentation result; it is determined by coefficients. The second weighted term also considers spatial distance. and grayscale difference This is used to preserve boundaries in areas of high grayscale contrast and suppress noise in areas of low contrast. By adjusting... and The value of can achieve a balance between overall smoothness and boundary preservation, as described in the embodiments of the present invention. and The constant is taken as an empirically determined value.

[0083] Best / Final Labels for Segmentation Map Minimize using a mean-field approximation algorithm By optimizing pixel labels in adjacent spaces through spatial consistency, 3DCRF can, to some extent, eliminate adhesions between adjacent blood vessels or between viscera and their maternal artery.

[0084] In this step, whether blood vessels are adhered is reflected in the local connectivity and morphology of the binary mask: If two anatomically separate vascular structures in the initial mask are connected by a small number of positively labeled (vessel=1) voxels to form a narrow "bridge", then topologically they are represented as a connected component, which is the "adhesion" in this invention.

[0085] This step utilizes the spatial distance between voxels through paired potential terms in 3DCRF. and grayscale difference Imposing a large energy penalty on "unnatural connections" makes the optimal label x ∗ Disrupt these bridging voxels and relabel them as background; perform spatial regularization of the mask by minimizing the Gibbs energy, utilizing the spatial distance between adjacent voxels. and grayscale difference It inhibits unreasonable adhesions and repairs discontinuous vascular segments; on the other hand, it removes isolated small noise areas to obtain a smoother and more connected intraluminal geometry of the AAA lumen.

[0086] Step 204: Use 3DCCO to retain the largest connected vessel cluster to obtain the final three-dimensional vessel binary data.

[0087] Specifically, 3DCCO uses the 3D label map optimized by 3DCRF as the target binary 3D volume. From any vessel label within this binary 3D volume, it selects one as the first vessel label. Starting from the first vessel label, it recursively searches for all adjacent vessel labels and adds them to the group associated with the selected first vessel label.

[0088] The above process is region growing, which iterates through the target binary 3D volume until all vessel labels connected to the first selected vessel label are found. If a vessel label is found without a group association, another arbitrary vessel label is initialized, and the region growing process restarts. After processing all vessel labels, several connected vessel groups are generated. The vessel group with the most vessel labels is selected as the final segmentation result, and morphological operations (such as dilation and erosion) are used to eliminate gaps in the vessel regions.

[0089] Therefore, 3DCCO retains only the largest connected component containing AAA and its parent artery. After further removing small connected clusters that are not related to the main geometry, the largest connected vessel cluster is retained, forming the final three-dimensional vessel binary data.

[0090] Step 205: Calculate the hemodynamic results of the blood vessels.

[0091] A high-precision, patient-specific three-dimensional geometric model of the abdominal aortic aneurysm is constructed based on the CTA images, and a CFD computational mesh is generated by combining the final three-dimensional vascular binary volume obtained from the aforementioned steps.

[0092] Based on the geometric and preset physical property parameters and inflow waveform, the CFD solver calculates hemodynamic results such as local pressure, flow velocity, and wall shear stress, enabling the acquisition of hemodynamic parameters through medical images.

[0093] The simulation of blood flow within and around the lumen was calculated using a finite volume-based CFD solver by applying the three-dimensional Navier-Stokes equations. in, Indicates blood density, Indicates pressure, Represents a three-dimensional velocity vector (blood flow velocity). Indicates viscosity. Indicates time period, This represents the gradient operator with respect to three-dimensional spatial coordinates. The divergence of the velocity vector is represented. This represents the Laplace operator.

[0094] Assume the blood vessel wall is rigid, the boundary condition is no slippage, and the blood is an incompressible Newtonian model. The dynamic viscosity and mass density of the blood are set to 0.004 kg / m³. s and 1050 kg / m 3 .

[0095] like Figure 8 As shown, the pulsating flow waveform measured by magnetic resonance flow imaging is used as the inlet boundary condition, and all outlets are under zero-pressure conditions.

[0096] Four cardiac cycles were simulated, each with 2000 steps (0.0005 seconds / step). By selecting 2000 time steps, the coulomb number of the above waveform was less than 1. Time-resolved hemodynamic data (20 time points) were saved in the last cardiac cycle. The results show that using ATRU-Net to segment AAA significantly shortens the model reconstruction time, and the segmentation results show a good correlation with those of experienced human operators.

[0097] The ATRU-Net structure provided by this invention has the following advantages over existing technologies: 1. It has different preprocessing and geometric post-processing functions for automatic CTA image segmentation. The attention mechanism can quickly locate the most prominent objects in cluttered visual scenes, thus automatically learning in chaotic environments, focusing on object structure, ignoring irrelevant parts, without additional supervision; 2. It adopts a residual learning structure, utilizing long-hop and short-hop connections to retain richer spatial information, which helps the network integrate multi-level spatial and semantic information, thereby improving sensitivity to smaller aneurysms and blood vessels, and improving the ability to acquire small vessel features; 3. Compared with the existing technology of manually (computing hemodynamics) reconstructing the model, the ATRU-Net algorithm can significantly shorten the AAA model reconstruction time, and automatic image segmentation can greatly reduce the subjectivity and limitations of manual segmentation.

[0098] This invention also provides a deep learning-based abdominal aortic aneurysm hemodynamic simulation system, such as... Figure 9 As shown, it includes: a spatial feature acquisition unit 31, an image feature optimization unit 32, and a hemodynamic result acquisition unit 33.

[0099] The spatial feature acquisition unit 31 is used to input the segmented image into the ATRU-Net network with an inverted pyramid structure to obtain a three-dimensional label map containing spatial feature information.

[0100] Specifically, it includes a spatial feature extraction subunit 311, which is used to input the segmented image into the ATRU-Net network. The spatial feature information output by each layer of the ATRU-Net network from the previous layer and the spatial feature information output by the first computing unit of the current layer are input together into the attention gate of the current layer and then output the spatial feature information of the current layer. This process is repeated iteratively to obtain a three-dimensional label map.

[0101] The loss function update subunit 312 is used to calculate the current loss value based on the feature map and corresponding labeled data output by each layer of the ATRU-Net network in the spatial feature extraction subunit according to the set loss function, and then update the parameters of the ATRU-Net network in the spatial feature extraction subunit through backpropagation based on the loss value.

[0102] The image feature optimization unit 32 is used to perform image processing on the three-dimensional label map output by the spatial feature acquisition unit to obtain an image result (three-dimensional blood vessel binary mask) that can be used for data processing by the hemodynamic result acquisition unit.

[0103] The system includes a 3DCRF processing subunit 321, which uses 3DCRF combined with the original image to perform regularization processing on the three-dimensional label map obtained by the spatial feature acquisition unit to obtain a processed image, so as to suppress unreasonable adhesions, repair discontinuous vascular segments, and obtain the intraluminal geometry of the abdominal aortic aneurysm.

[0104] The 3DCCO processing subunit 322 is used to perform connected component search on the image processed by the 3DCRF processing subunit using the 3D connected component optimization 3DCCO method, retain the largest connected blood vessel cluster, and obtain the final 3D blood vessel binary mask.

[0105] The hemodynamic results acquisition unit 33 is used to set boundary conditions, use a surface reconstruction algorithm to generate a three-dimensional blood vessel binary mask output by the image feature optimization unit, process the data to obtain the surface of the blood vessel lumen, and then use the finite volume method to divide the volume mesh to generate a CFD calculation mesh to calculate the hemodynamic results of the blood vessel corresponding to the three-dimensional blood vessel binary mask.

[0106] In summary, the technical solution provided by this invention solves the problems of time-consuming manual annotation, subjectivity and limitations in existing research techniques. It can better preserve detailed information and local features, thereby better handling complex lesion structures and edge information, and improving the accuracy and stability of segmentation.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for simulating the hemodynamics of abdominal aortic aneurysms, characterized in that, The method includes: The segmented image is input into the ATRU-Net network with an inverted pyramid structure to obtain a 3D label map containing spatial feature information. Specifically, the spatial feature information output by each layer of the ATRU-Net network to the previous layer and the spatial feature information output by the first computing unit of the current layer are input together into the attention gate of the current layer to output the spatial feature information of the current layer. A loss function is set, which calculates the current loss value based on the feature map and corresponding labeled data output by each layer of the ATRU-Net network, and then updates the parameters of the ATRU-Net network through backpropagation; The three-dimensional label map was regularized by combining the fully connected three-dimensional conditional random field 3DCRF with the original image to obtain the processed image, so as to suppress unreasonable adhesions, repair discontinuous vascular segments, and obtain the intraluminal geometry of the abdominal aortic aneurysm. The 3DCCO (Three-Dimensional Connectivity Component Optimization) method is used to perform connected component search on the processed image, retaining the largest connected blood vessel cluster, and obtaining the final 3D binary blood vessel mask. Boundary conditions are set, and the data of the three-dimensional blood vessel binary mask is processed by the surface reconstruction algorithm to obtain the surface of the blood vessel lumen. Then, the volume mesh is generated by the finite volume method to generate the CFD computing mesh, and the hemodynamic results of the blood vessel corresponding to the three-dimensional blood vessel binary mask are calculated.

2. The method according to claim 1, characterized in that, The inverted pyramid structure of the ATRU-Net network is specifically as follows: The ATRU-Net network has an inverted pyramid-shaped four-layer architecture; each layer has a dimensionally symmetrical convolutional layer and a parameter correction linear unit group, and the convolutional layers and parameter correction linear unit groups are arranged in an inverted pyramid shape from low resolution to high resolution. Each convolutional layer and parameter correction linear unit group has an attention gate between the two convolutional layers and parameter correction linear units.

3. The method according to claim 1, characterized in that, The ATRU-Net network includes an encoding path and a decoding path: In the encoding path, the output of the first convolutional layer and parameter correction linear unit of the low-resolution layer is used as the input of the first convolutional layer and parameter correction linear unit of the next higher resolution layer, so as to realize the layer-by-layer transfer of features from low resolution to high resolution. In the decoding path, the features processed by the second convolutional layer and parameter correction linear unit of the highest resolution layer are taken as the starting point. After being fused with the output of the second convolutional layer and parameter correction linear unit of the adjacent high resolution layer through the attention gate, they are sequentially input into the second convolutional layer and parameter correction linear unit of the next low resolution layer. This realizes the feature backpropagation from high resolution to low resolution layer by layer, thereby obtaining a 3D label map containing spatial feature information.

4. The method according to claim 3, characterized in that, For each of the attention gates mentioned: The attention gate takes the feature map output from the previous convolutional layer as a gate signal, encodes the gate signal with the feature map output from the same layer to obtain an encoded feature map, and optimizes the encoded feature map accordingly. Specifically, the feature map and the gate signal are concatenated and mapped to the intermediate space. Different activation maps are obtained sequentially through different activation operations. The nonlinear layer corresponding to the next activation map is resampled by trilinear interpolation to finally obtain the output of the current attention gate.

5. The method according to claim 1, characterized in that, The loss function calculates the current loss value based on the feature maps and corresponding labeled data output by each layer of the ATRU-Net network, and then updates the parameters of the ATRU-Net network through backpropagation, including: In the 3D label map containing spatial feature information output by each layer of the ATRU-Net network, the predicted pixel set of individual voxels belonging to the blood vessel category and the labeled pixel set in the 3D label map are extracted. The loss function calculates the loss value based on the predicted pixel set and the labeled pixel set, combined with the false negative penalty magnitude and the false positive penalty magnitude.

6. The method according to claim 1, characterized in that, The process involves using a fully connected 3D conditional random field (3DCRF) combined with the original image to regularize the 3D label map, resulting in a processed image. This process suppresses unreasonable adhesions, repairs discontinuous vascular segments, and yields the intraluminal geometry of the connected abdominal aortic aneurysm, including: Based on 3DCRF combined with the original image, the three-dimensional label map is spatially regularized by minimizing the Gibbs energy. In this process, the spatial distance between adjacent voxels and the gray-level difference between adjacent voxels are used to suppress unreasonable adhesions and repair discontinuous vascular segments.

7. The method according to claim 1, characterized in that, The process employs the 3D Connectivity Component Optimization (3DCCO) method to perform connected component search on the processed image, retaining the largest connected vessel clusters to obtain the final 3D binary vessel mask, including: Using 3DCCO, starting with any blood vessel label in the binary 3D volume of the image after 3DCRF regularization, all adjacent blood vessel labels are recursively searched and added to the group associated with the selected first blood vessel label, until all blood vessel labels connected to the selected first blood vessel label are found; if there are ungrouped blood vessel labels, another blood vessel label is initialized and the region growing process is restarted; after processing all blood vessel labels, at least one connected blood vessel group is obtained, and the blood vessel group with the most blood vessel labels is taken as the final segmentation result to obtain the final 3D blood vessel binary mask.

8. The method according to claim 7, characterized in that, The boundary conditions include: An instantaneous volumetric flow rate waveform is applied to the inlet section determined based on the three-dimensional blood vessel binary mask segmentation results, thereby setting the relationship between blood flow velocity or blood flow rate and time. Export boundary conditions: Zero pressure conditions are applied to all outlets, representing the reference value of peripheral pressure; Blood vessel wall boundary conditions: Apply no-slip boundary conditions to the surface of the blood vessel wall formed by the fluid domain, and assume that the blood vessel wall is rigid and does not deform; Physical properties: The density and dynamic viscosity of blood are uniformly set throughout the entire fluid domain, and hemodynamic results such as local pressure, flow velocity, and wall shear stress are calculated accordingly.

9. A deep learning-based abdominal aortic aneurysm hemodynamic simulation system, characterized in that, The system includes: a spatial feature acquisition unit, an image feature optimization unit, and a hemodynamic result acquisition unit; The spatial feature acquisition unit is used to input the segmented image into the pyramid-structured ATRU-Net network to obtain a 3D label map containing spatial feature information. Specifically, the spatial feature information output by each layer of the ATRU-Net network from the previous layer and the spatial feature information output by the first computing unit of the current layer are input together into the attention gate of the current layer to output the spatial feature information of the current layer. The unit calculates the current loss value based on the feature map and corresponding labeled data output by each layer of the ATRU-Net network according to the set loss function, and updates the parameters of the ATRU-Net network through backpropagation. The image feature optimization unit is used to perform regularization processing on the three-dimensional label map using a fully connected three-dimensional conditional random field (3DCRF) combined with the original image to obtain a processed image, in order to suppress unreasonable adhesions, repair discontinuous vascular segments, obtain the intraluminal geometry of the abdominal aortic aneurysm, and use the three-dimensional connectivity component optimization (3DCCO) method to perform connected component search on the processed image, retain the largest connected vascular cluster, and obtain the final three-dimensional vascular binary mask. The hemodynamic results acquisition unit is used to obtain the surface of the blood vessel lumen by setting boundary conditions and using a surface reconstruction algorithm to process the data of the three-dimensional blood vessel binary mask, and then generate a CFD calculation mesh by dividing the volume mesh using the finite volume method, and calculate the hemodynamic results of the blood vessel corresponding to the three-dimensional blood vessel binary mask.

Images