Deep learning-based abdominal aorta blood vessel segmentation method based on three-dimensional image
Through the three-dimensional image abdominal aortic vascular segmentation method based on deep learning network, the spatial feature map calibration, self-attention and multi-scale feature fusion modules are used to solve the problem of low accuracy of the abdominal aortic vascular reconstruction algorithm, and achieve higher vascular segmentation accuracy and robustness.
Patent Information
- Application Number
- CN202411786269.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-06
- Publication Date
- 2025-05-06
AI Technical Summary
The existing abdominal aortic vascular reconstruction algorithm has limited accuracy, mainly due to the imbalance of vascular and non-vascular voxels and the anisotropic characteristics of vascular distribution.
The three-dimensional image aortic vascular segmentation method based on deep learning network is adopted to improve the accuracy of vascular segmentation through the spatial feature map calibration module, self-attention module and multi-scale feature fusion module.
It effectively solves the problem of sample number imbalance, improves the identification and segmentation accuracy of tiny blood vessels in three-dimensional data, and is suitable for abdominal enhancement CT, plain scanning CT, MR and other image data, and has good robustness.
Smart Images
Figure CN119942102A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical images, and in particular to a method and system for segmenting abdominal aorta vessels in three-dimensional images based on a deep learning network. Background Art
[0002] Accurate vascular segmentation is crucial for the diagnosis and treatment of abdominal vascular diseases; and the abdominal aorta is a key organ. In the traditional mode, clinicians rely on interactive vascular tracking and segmentation. The reconstruction process is very time-consuming, which affects the efficiency of diagnosis and intervention. The vascular reconstruction method based on artificial intelligence can greatly reduce working time through automated segmentation. At the same time, accurate segmentation results play an important role in medical image analysis and accelerate clinical processes. However, the accuracy of existing abdominal aortic reconstruction algorithms is limited. There are two main reasons: first, the abdominal aorta only occupies sparse voxels in three-dimensional CT images, and there is a serious imbalance in the number of voxels between aortic and non-aortic vessels; second, the abdominal aorta is a slender tubular structure, and its distribution is anisotropic. The above factors increase the difficulty of abdominal aortic reconstruction.
[0003] The current segmentation technology of abdominal aorta has the following shortcomings: 1. The accuracy is insufficient, which is far from the clinical application requirements; 2. The reconstruction effect of vascular branches and distal parts is poor. To this end, we propose a method and system for abdominal aorta segmentation in 3D images based on deep learning network. Summary of the invention
[0004] In view of the problems in the prior art, the present invention proposes a method and system for segmenting abdominal aorta vessels in three-dimensional images based on deep learning.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning, comprising the following steps:
[0007] Step S1: Collect three-dimensional image data, perform data preprocessing, and generate a region of interest;
[0008] Step S2: generating a plurality of three-dimensional data blocks including the region of interest according to the three-dimensional data and the region of interest obtained by data preprocessing;
[0009] Step S3: input all three-dimensional data blocks containing the region of interest into the abdominal aorta vessel segmentation deep learning network to obtain the abdominal aorta vessel segmentation result corresponding to each input data; and recombine the segmentation results according to the original relative relationship to generate the abdominal aorta vessel segmentation result.
[0010] Furthermore, the preprocessing of the three-dimensional image data in step S1 includes the following parts: spine extraction, calculation and generation of a region of interest, and pixel value normalization.
[0011] Further, the abdominal spine extraction method is to apply a spine extraction algorithm to the three-dimensional data to extract the abdominal spine region;
[0012] Furthermore, the method for calculating and generating the region of interest is to calculate the spatial range of the abdominal aorta blood vessels as the region of interest based on the spatial range of the abdominal spine segmentation result using prior knowledge.
[0013] Furthermore, the pixel value normalization method is:
[0014] The minimum value of pixel value normalization is min_value, the maximum value is max_value, the pixel value is x, and the normalized output result is y. The calculation process is:
[0015] if x<min_value:
[0016] then x=min_value
[0017] if x>max_value
[0018] then x=max_value
[0019]
[0020] Furthermore, the three-dimensional data in step S2 is the data preprocessed in step S1, and the method for generating multiple three-dimensional data containing the region of interest in step S2 is: selecting a boundary vertex as a starting point in the three-dimensional data of the region of interest, taking a 32×48×48 offset as a step length, and extracting it as a local block three-dimensional image region with a size of 64×128×128; then checking all local block three-dimensional image regions, and discarding local block three-dimensional image regions that do not contain the region of interest; finally, according to the correspondence between the three-dimensional data and the region of interest, extracting the region corresponding to the local block three-dimensional image region in the three-dimensional data, and obtaining multiple three-dimensional data blocks containing the region of interest.
[0021] Furthermore, in step S3, the input data of the abdominal aorta vessel segmentation deep learning network includes: a three-dimensional data block of the region of interest, and three-dimensional image data.
[0022] Furthermore, in step S3, the main network structure of the abdominal aorta vessel segmentation deep learning network contains a spatial feature map calibration module, a self-attention module and a multi-scale feature fusion module.
[0023] Furthermore, the spatial feature map calibration module enhances the abdominal aorta blood vessel related areas in the feature map, distinguishes them from the background areas, weakens the sample data imbalance problem, and improves the accuracy of small blood vessel recognition in three-dimensional data.
[0024] Furthermore, the self-attention module upsamples the input features. Then the maximum pooling operation is used to extract the key response features, and each response value is normalized using the sigmoid operation, and then superimposed with the low-level features as the input of the next network.
[0025] Furthermore, the multi-scale feature map fusion module is used to superimpose the semantic information of the abdominal aorta vascular feature maps of different scales at different depths, and then perform upsampling operations to achieve the purpose of fusing feature maps of different scales.
[0026] Furthermore, the output image of the abdominal aorta vessel segmentation deep learning network is a dual-channel image, which is an abdominal aorta vessel probability map and a background probability map.
[0027] The present invention also provides a system for segmenting abdominal aorta vessels in three-dimensional images based on deep learning, comprising:
[0028] Parameter setting interactive module: used to provide users with a visual list or text configuration method to set model parameters;
[0029] Image preprocessing module: used to generate a region of interest and perform preprocessing based on three-dimensional image data, and to generate multiple three-dimensional data blocks containing the region of interest;
[0030] Image segmentation module: used to perform abdominal aortic vascular image segmentation using a deep learning network;
[0031] Image reconstruction module: used to render images based on the abdominal aorta segmentation results, texture data and transformation matrix to obtain reconstructed three-dimensional medical images.
[0032] Compared with the prior art, the present invention has the following beneficial effects: 1. In the method for segmenting the abdominal aorta, the present invention uses a spatial feature map calibration module to effectively solve the problem of sample number imbalance in the deep learning network training process, thereby making the recognition and segmentation accuracy of small blood vessels in three-dimensional data higher;
[0033] 2. In the abdominal aorta segmentation method of the present invention, the self-attention module is used to strengthen the representation of the abdominal aorta in the feature map, so that the network has higher feature recognition and segmentation accuracy for the abdominal aorta in the feature map.
[0034] 3. In the abdominal aorta segmentation method of the present invention, a multi-scale fusion module is used to fuse the semantic information in feature maps of different depths, and effectively identify the thinner blood vessels in the abdominal aorta.
[0035] 4. The abdominal aorta segmentation method of the present invention is applicable to abdominal enhanced CT, plain scan CT, MR, etc., and has good robustness. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] The present invention is further described below in conjunction with the accompanying drawings.
[0037] Figure 1 A flow chart of a method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning provided by an embodiment of the present invention;
[0038] Figure 2 This is a subject network structure diagram of the abdominal aorta blood vessel segmentation deep learning network of the present invention;
[0039] Figure 3 The result is element-wise added to obtain an output feature map for the feature recalibration operation of the present invention;
[0040] Figure 4 The present invention applies sigmoid operation to normalize the low-dimensional feature map to obtain an output feature map;
[0041] Figure 5 A schematic diagram of the structure of a deep learning-based abdominal aorta blood vessel segmentation system in three-dimensional images provided in an embodiment of the present invention.
[0042] Figure 6 This is a schematic diagram of the abdominal aorta blood vessel reconstructed by the present invention. DETAILED DESCRIPTION
[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0044] Figure 1 A flow chart of a method for segmenting abdominal aorta in three-dimensional images based on deep learning provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention proposes a method for segmenting abdominal aorta vessels in three-dimensional images based on a deep learning network, comprising the following steps:
[0045] Step S1: Collect three-dimensional image data, perform data preprocessing, and generate a region of interest;
[0046] In step S1 of the embodiment of the present invention, generating a region of interest from three-dimensional image data includes the following steps: threshold binarization and abdominal spine region extraction;
[0047] Among them, the threshold binarization method is:
[0048] For each pixel value ct_value in the three-dimensional image data, for the threshold thresh:
[0049] if ct_value<thresh then ct_value=0else ct_value=1
[0050] The method for extracting the abdominal spine region is as follows: according to connectivity, the binary three-dimensional data is divided into different connected region sets, and then the connected region set of the abdominal spine region is determined according to the relative relationship of the spatial position.
[0051] In step S1 of the embodiment of the present invention, data preprocessing includes the following steps: spine extraction, calculation and generation of regions of interest, and pixel value normalization.
[0052] The method for extracting the abdominal spine region is as follows: according to connectivity, the binary three-dimensional data is divided into different connected region sets, and then the connected region set of the abdominal spine region is determined according to the relative relationship of the spatial position.
[0053] The method for calculating the region of interest is as follows: according to the spatial range, spatial position relationship and prior knowledge of the abdominal spine region, the region of interest including the abdominal aorta is calculated.
[0054] Among them, the pixel value normalization method is:
[0055] The minimum value of pixel value normalization is min_value, the maximum value is max_value, the pixel value is x, and the normalized output result is y. The calculation process is:
[0056] if x<min_value
[0057] then x=min_value
[0058] if x>max_value
[0059] then x=max_value
[0060]
[0061] Step S2: generating a plurality of three-dimensional data including the region of interest according to the three-dimensional data and the region of interest obtained by data preprocessing;
[0062] In step S2 of the embodiment of the present invention, the method for generating multiple three-dimensional data including the region of interest is: selecting a boundary vertex as a starting point in the three-dimensional data of the region of interest, taking a 32x32x32 offset as a step size, and extracting it as a local block three-dimensional image with a size of 64x128x128; then checking all local block three-dimensional regions, and discarding local block three-dimensional images that do not include the region of interest; finally, according to the positional correspondence between the three-dimensional data and the region of interest, extracting the region in the three-dimensional data corresponding to the local block three-dimensional image, to obtain multiple three-dimensional data blocks including the region of interest.
[0063] Step S3: input all three-dimensional data blocks containing the region of interest into the abdominal aorta vessel segmentation deep learning network to obtain the abdominal aorta vessel segmentation result corresponding to each input data position; and recombine the segmentation results according to the original relative relationship to generate the abdominal aorta vessel segmentation result.
[0064] Among them, the output image of the abdominal aorta vessel segmentation deep learning network is a dual-channel image, namely the abdominal aorta vessel probability map and the background probability map.
[0065] like Figure 2 As shown in the figure, the subject network structure of the deep learning network for abdominal aorta segmentation has a spatial feature map calibration module, a self-attention module, and a multi-scale feature fusion module, which is a framework that integrates the three. Convolutional pooling operations are used to extract high-dimensional features, and then feature recalibration operations are performed, and then upsampling operations are performed to restore the initial resolution.
[0066] like Figure 3 As shown, the feature recalibration operation sets parameters in three dimensions, namely, depth, height, and width, and convolves the input feature map. The results are added element by element to obtain the output feature map. This achieves the purpose of recalibrating the abdominal aorta blood vessel related features in the feature map and improves its recognition probability.
[0067] like Figure 4 As shown in the figure, the self-attention module performs an upsampling operation on the input high-dimensional feature map, then performs a maximum pooling operation, and then applies a sigmoid operation to normalize it, and multiplies it with a low-dimensional feature map of the same resolution to obtain an output feature map.
[0068] The multi-scale fusion module can fuse feature maps containing semantic information of different dimensions, effectively utilize low-latitude and high-latitude semantic information to segment and identify the abdominal aorta, and segment the complete abdominal aorta results.
[0069] Based on any of the above embodiments, Figure 5 A schematic diagram of the structure of a deep learning-based abdominal aorta segmentation system in three-dimensional images provided by an embodiment of the present invention, the system comprising:
[0070] Data preprocessing and region of interest generation module: used to perform data preprocessing and generate regions of interest based on three-dimensional image data;
[0071] Image re-segmentation module: used to re-segment and generate multiple three-dimensional data including the region of interest based on the three-dimensional data and the region of interest obtained by data preprocessing;
[0072] Abdominal aorta segmentation module: used to input all three-dimensional data containing the region of interest into the abdominal aorta segmentation deep learning network, obtain the abdominal aorta segmentation results corresponding to each input data, and recombine the segmentation results according to the original relative relationship to generate the abdominal aorta segmentation results.
[0073] The above contents are merely examples of the structure of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they shall all fall within the protection scope of the present invention.
Claims
1. A method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning, characterized in that: The following steps are involved: Step S1: Collect three-dimensional image data, perform data preprocessing, and generate a region of interest; Step S2: generating a plurality of three-dimensional data blocks including the region of interest according to the three-dimensional data and the region of interest obtained by data preprocessing; Step S3: input all three-dimensional data blocks containing the region of interest into the abdominal aorta vessel segmentation deep learning network to obtain the abdominal aorta vessel segmentation result corresponding to each input data; and recombine the segmentation results according to the original relative relationship to generate the abdominal aorta vessel segmentation result.
2. The method for segmenting abdominal aorta in three-dimensional images based on deep learning according to claim 1, characterized in that: The region of interest in step S1 is the abdomen region in the three-dimensional image data.
3. The method for segmenting abdominal aorta in three-dimensional images based on deep learning according to claim 1, characterized in that: The preprocessing of the three-dimensional image data in step S1 includes the following parts: spine extraction, calculation and generation of regions of interest, and pixel value normalization.
4. The method for segmenting abdominal aorta in three-dimensional images based on deep learning according to claim 1, characterized in that: The method for generating multiple three-dimensional data containing the region of interest in step S2 is to select a boundary vertex in the three-dimensional data of the region of interest as a starting point, use a 32x32x32 offset as a step size, and extract it as a local block three-dimensional image with a size of 64x128x128; then check all local block three-dimensional regions, and discard the local block three-dimensional images that do not contain the region of interest; finally, according to the correspondence between the three-dimensional data and the region of interest, extract the region in the three-dimensional data corresponding to the local block three-dimensional image to obtain multiple three-dimensional data blocks containing the region of interest.
5. The method for segmenting abdominal aorta in three-dimensional images based on deep learning according to claim 1, characterized in that: The main network structure of the deep learning network for abdominal aorta segmentation in step S3 includes a spatial feature map calibration module, a self-attention module and a multi-scale fusion module.
6. The method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning according to claim 5, characterized in that: The spatial feature map calibration module sets parameters in three dimensions, namely, depth, height and width, and convolves with the input feature map respectively, and then adds the results element by element to obtain the output feature map; thereby achieving the purpose of recalibrating the abdominal aorta blood vessel related features in the feature map and improving its recognition probability.
7. The method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning according to claim 5, characterized in that: The self-attention module performs an upsampling operation on the input high-dimensional feature map, then performs a maximum pooling operation, and then applies a sigmoid operation to normalize it, and multiplies it with a low-dimensional feature map of the same resolution to obtain an output feature map.
8. The method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning according to claim 5, characterized in that: The multi-scale fusion module fuses the feature maps containing semantic information of different depths, effectively utilizes low-dimensional and high-dimensional semantic information to segment and identify the abdominal aorta, and segments out a complete abdominal aorta result.
9. The method for segmenting abdominal aorta vessels in three-dimensional images based on deep learning according to claim 5, characterized in that: The abdominal aorta vessel segmentation deep learning network outputs two channels: an abdominal aorta vessel probability map channel and a background probability map channel.
10. A system for segmenting abdominal aorta vessels in three-dimensional images based on deep learning, characterized in that: include: Region of interest generation and data preprocessing module: used to generate regions of interest and preprocess data based on three-dimensional image data; Image re-segmentation module: used to re-segment and generate multiple three-dimensional data blocks containing the region of interest based on the three-dimensional data and the region of interest obtained by data preprocessing; Abdominal aorta segmentation module: used to input all three-dimensional data blocks containing the area of interest into the abdominal aorta segmentation deep learning network, obtain the abdominal aorta segmentation results corresponding to each input data, and recombine the segmentation results according to the original relative relationship to generate the abdominal aorta segmentation results.