An abdominal aorta automatic segmentation method based on split sampling neural network

Abstract

The application discloses an abdominal aorta automatic segmentation method based on a segmentation sampling neural network, and specifically comprises the following steps: obtaining a historical CT sequence scanning image, performing image preprocessing and cutting with a preset size to obtain an original cutting file; inputting the original cutting file into a pre-trained front segmentation network model after down-sampling to perform forward propagation, and obtaining a full receptive field prediction file; performing global segmentation sampling on the original cutting file based on a segmentation step to obtain a plurality of first sub-domain files; performing local segmentation sampling on the original cutting file based on the segmentation step to obtain a plurality of second sub-domain files; constructing and training a sub-domain prediction network model, wherein the sub-domain prediction network model takes the full receptive field prediction file, all the first sub-domain files and all the second sub-domain files as inputs; and using the front segmentation network model and the trained sub-domain prediction network model to perform target segmentation on a real-time CT sequence scanning image to obtain a voxel segmentation model of the abdominal aorta.

Description

Technical Field

[0001] This invention belongs to the field of abdominal aortic CT image processing technology, specifically relating to an automatic segmentation method for the abdominal aorta based on a segmentation sampling neural network. Background Technology

[0002] Abdominal aortic aneurysm (AAA) is a highly prevalent disease in the elderly, characterized by its high incidence and rapid, severe progression. The primary diagnostic method is CT scan combined with medical image reconstruction. Currently, clinically recommended AAA risk assessment uses the maximum diameter of the aneurysm and its annual growth rate. This assessment heavily relies on the accuracy of modeling the abdominal aortic aneurysm portion of the case. Both excessively slow modeling and analysis schemes and poor modeling accuracy can lead to misdiagnosis of the patient's condition.

[0003] Currently, the main methods for modeling CT images include: deep learning-based segmentation methods, region growing-based segmentation methods, and level set evolution-based segmentation methods. Region growing-based segmentation methods rely on selected seed points to continuously expand until reaching the maximum boundary region under given boundary constraints. The advantage of this method is its simplicity and relatively complete boundaries. However, the selection of seed points is manual, resulting in high time complexity during segmentation. Furthermore, its biggest drawback is the potential for image leakage at blurred boundaries. Level set evolution-based methods are susceptible to noise interference and struggle to correctly segment complex edge conditions.

[0004] Therefore, designing an efficient and accurate automatic segmentation and modeling scheme for the abdominal aortic flow channel and abdominal aortic vessels based on CT images is of great significance for assisting medical judgment and guiding analysis. Summary of the Invention

[0005] In view of this, the purpose of this invention is to overcome the technical problems encountered by existing CT image-based modeling methods, such as seed point selection relying on manual methods, high time complexity of the segmentation process, easy occurrence of discontinuous tomography during the segmentation process, and segmentation failure due to noise interference, and to provide an automatic segmentation method for the abdominal aorta based on a segmentation sampling neural network.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] An automatic abdominal aorta segmentation method based on a segmentation sampling neural network includes the following steps:

[0008] Acquire historical CT sequence scan images and perform image preprocessing and cropping to a preset size to obtain the original cropped file;

[0009] The original cropped file is downsampled and then input into a pre-trained pre-segmentation network model for forward propagation.

[0010] To the full receptive field prediction file including the abdominal aortic region;

[0011] Based on the preset segmentation step size, the original cropped file is subjected to global segmentation sampling to obtain multiple first subdomain files;

[0012] Based on a preset segmentation step size, the original cropped file is locally segmented and sampled to obtain multiple second subdomain files;

[0013] Construct and train a subdomain prediction network model, wherein the subdomain prediction network model takes the full receptive field prediction file, all first subdomain files and all second subdomain files as input datasets;

[0014] Using the aforementioned pre-segmentation network model and the trained subdomain prediction network model, target segmentation is performed on real-time CT sequence scan images to obtain a voxel segmentation model of the abdominal aorta.

[0015] The pre-segmentation network model and the subdomain prediction network model are both constructed based on three-dimensional convolutional neural networks, and both take the abdominal aortic channel and blood vessels as segmentation targets.

[0016] Preferably, the three-dimensional convolutional neural network includes mutually symmetrical contraction paths and expansion paths, which are connected by convolution, and both the contraction paths and expansion paths are generated based on a fully convolutional network.

[0017] Preferably, the automatic abdominal aorta segmentation method further includes:

[0018] The voxel segmentation model is then smoothed and visualized.

[0019] Preferably, the image preprocessing includes Hu value normalization based on an upper threshold and a lower threshold.

[0020] Preferably, the pre-segmentation network model includes a first contraction path and a first expansion path;

[0021] The first shrinkage path, from top to bottom, includes an image input channel that takes the file obtained by downsampling the original cropped file as input, a first merging model connected to the image input channel, a first downsampling network connected to the first merging model through convolution, a second downsampling network connected to the first downsampling network through convolution, a third downsampling network connected to the second downsampling network through convolution, and a fourth downsampling network connected to the third downsampling network through convolution.

[0022] The first extended path, from bottom to top, includes a first upsampling network connected to the fourth downsampling network via convolution, a second upsampling network connected to the first upsampling network via convolution, a third upsampling network connected to the second upsampling network via convolution, a fourth upsampling network connected to the third upsampling network via convolution, a second merging model connected to the fourth upsampling network via convolution, and a first image output channel connected to the second merging model. The first image output channel is used to output the full receptive field prediction file.

[0023] The output of the first merging model is summed with the output of the first downsampling network. The output of this summed network is then dropped out and summed with the output of the second downsampling network. The output of this summed network is then dropped out and summed with the output of the third downsampling network. The output of this summed network is then dropped out and summed with the output of the third downsampling network. The output of this summed network is then dropped out and summed with the output of the second upsampling network. The output of this summed network is then convolved with the output of the first upsampling network. The output of the first upsampling network is then summed with the output of the second upsampling network. The output of this summed network is then summed with the output of the third upsampling network. The output of this summed network is then summed with the output of the fourth upsampling network. The output of this summed network is then summed with the output of the second merging model. The output of this summed network is then passed to the first image output channel.

[0024] Preferably, the step of performing global segmentation sampling on the original cropped file based on a preset segmentation step size to obtain multiple first sub-domain files is as follows:

[0025] Based on the preset segmentation step size, the original cropped file is divided into N equal parts in the X, Y and Z directions to obtain multiple first subdomain files;

[0026] Preferably, the step of performing local segmentation sampling on the original cropped file based on a preset segmentation step size to obtain multiple second sub-domain files specifically involves:

[0027] Based on a preset segmentation step size, the remaining pixel region after discarding the boundary in the X direction of the original cropped file is divided into M equal parts, the remaining pixel region after discarding the boundary in the Y direction of the original cropped file is divided into M equal parts, and the remaining pixel region after discarding the boundary in the Z direction of the original cropped file is divided into M equal parts, resulting in multiple second sub-domain files.

[0028] Preferably, the subdomain prediction network model uses the full receptive field prediction file, all first subdomain files, and all second subdomain files as input datasets, specifically:

[0029] Based on the preset segmentation step size, global segmentation sampling and local segmentation sampling are performed on the full receptive field prediction file to obtain multiple third subdomain files;

[0030] Use all first subdomain files, all second subdomain files, and all third subdomain files as the input dataset for the subdomain prediction network model.

[0031] Preferably, the subdomain prediction network model includes a second shrinking path and a second expanding path;

[0032] The second shrinkage path, from top to bottom, includes a third merging model that takes all first subdomain files and all second subdomain files as input, a fifth downsampling network connected to the third merging model via convolution, a sixth downsampling network connected to the fifth downsampling network via convolution, a seventh downsampling network connected to the sixth downsampling network via convolution, and an eighth downsampling network connected to the seventh downsampling network via convolution, and also takes all third subdomain files as input to the sixth downsampling network;

[0033] The second extension path, from bottom to top, includes a fifth upsampling network connected to the eighth downsampling network via convolution, a sixth upsampling network connected to the fifth upsampling network via convolution, a seventh upsampling network connected to the sixth upsampling network via convolution, an eighth upsampling network connected to the seventh upsampling network via convolution, a fourth merging model connected to the eighth upsampling network via convolution, and a second image output channel connected to the fourth merging model. The second image output channel is used to output the voxel segmentation model.

[0034] The output of the third merging model is summed with the output of the fifth downsampling network. The dropout of this summed output is passed to the eighth upsampling network and then summed with the output of the sixth downsampling network. The dropout of this summed output is passed to the seventh upsampling network and then summed with the output of the seventh downsampling network. The dropout of this summed output is passed to the sixth upsampling network and then summed with the output of the eighth downsampling network. The output of this summed output is convolved with the fifth upsampling network. The output of the fifth upsampling network is summed with the output of the sixth upsampling network. The output of this summed output is summed with the output of the seventh upsampling network. The output of this summed output is summed with the output of the eighth upsampling network. The output of this summed output is then summed with the output of the fourth merging model. The output of this summed output is then passed to the second image output channel.

[0035] Preferably, the standardization of the Hu value based on the upper and lower thresholds specifically involves:

[0036] The Hu value is standardized according to the first formula in the NumPy standard library, which is: ;

[0037] in, This represents the standardized pixel Hu value. This represents the original Hu value of the pixel. Indicates the lower threshold. This indicates the upper threshold.

[0038] Preferably, the smoothing and visualization of the voxel segmentation model specifically includes:

[0039] Perform opening and closing operations on the voxel segmentation model;

[0040] The VTK library is used to convert the voxel segmentation model after performing opening and closing operations into a triangular facet model.

[0041] Visualize and save the triangular facet model using the VTK library.

[0042] The beneficial effects of this invention are as follows:

[0043] (1) After the pre-segmentation network model performs information sampling and regional connectivity determination, it obtains the coarse segmentation results of the abdominal aortic flow channel and abdominal aortic vessels. Then, based on the segmentation sampling technical route, the original cropped file is subjected to global segmentation and local misalignment segmentation. Then, the full receptive field prediction file, the first subdomain file obtained by global segmentation, and the second subdomain file obtained by misalignment segmentation are used as the input dataset of the subdomain prediction network model. The target is segmented again through the subdomain prediction network to achieve complete, accurate and efficient segmentation of the abdominal aorta without manual intervention. Moreover, the voxel segmentation model obtained by segmentation has a smooth morphological structure and strong segmentation robustness. Compared with the traditional abdominal aorta modeling method, it is more conducive to the auxiliary judgment of medical images and greatly improves the efficiency of clinical morphological detection of abdominal aortic aneurysm, meeting the application requirements.

[0044] (2) The symmetrical shrinking and expanding paths in the three-dimensional convolutional neural network are set. The shrinking path compresses the features of the input image, retains key information, and acts as an encoder. The expanding path restores the feature map, thereby outputting a segmented image of the same size as the original image. This achieves the preservation of detailed information while segmenting the image, and ultimately makes the voxel segmentation model obtained by the subdivision closer to the real state of the abdominal aorta.

[0045] (3) The subdomain prediction network model integrates pruning and downsampling, and has a larger receptive field than the traditional neural network segmentation model;

[0046] (4) By summing the elements in the first contraction path and the first expansion path and using dropout to cross the connection, the training process is optimized and the feature expression is enriched, thereby improving the prediction accuracy of the pre-segmentation network model;

[0047] (5) By summing the elements in the second contraction path and the second expansion path and by dropping out across connections, the training process is optimized and the feature representation is enriched, thereby improving the prediction accuracy of the subdomain prediction network model.

[0048] (6) The voxel segmentation model is smoothed, for example, by using preferred opening and closing operations, which achieves smooth edge processing of the model;

[0049] (7) Visualize the voxel segmentation model based on the VTK library to facilitate the visualization and verification of the segmentation results. Attached Figure Description

[0050] Figure 1 This is a flowchart of an automatic abdominal aorta segmentation method based on a segmentation sampling neural network;

[0051] Figure 2 A schematic diagram of a pre-segmentation network model;

[0052] Figure 3 A schematic diagram of a subdomain prediction network model;

[0053] Figure 4 Another flowchart of an automatic abdominal aorta segmentation method based on a segmentation sampling neural network;

[0054] Figure 5 A schematic diagram of a vascular segmentation plane of the abdominal aorta;

[0055] Figure 6 A visual schematic diagram of a voxel segmentation model of the abdominal aorta;

[0056] Figure 7 This is a schematic diagram of the abdominal aorta voxel segmentation model after smoothing based on the VTK library;

[0057] Figure 8 This is a schematic diagram of a predicted flow path and aneurysm in the abdominal aorta. Detailed Implementation

[0058] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0059] See Figures 1-8 This embodiment provides an automatic abdominal aorta segmentation method based on a segmentation sampling neural network, which specifically includes the following steps:

[0060] S100. Acquire historical CT sequence scan images and perform image preprocessing and cropping to a preset size to obtain the original cropped file.

[0061] In some embodiments, the historical CT sequence scan images are in DICOM format. While acquiring the historical CT sequence scan images, the number of layers in the Z direction of the historical CT sequence scan file is also obtained. S100 also includes case information reading analysis, image pixel spacing reading analysis, and spatial origin position reading analysis. The reading analysis can be performed using the medical image processing software package ITK. The case information includes the case name and the case gender, etc. The preset size is set to 512×512×512, so the size of each original cropped file is 512×512×512.

[0062] Image preprocessing includes Hu value normalization based on upper and lower thresholds. A specific implementation process is as follows:

[0063] The Hu value is standardized according to the first formula in the NumPy standard library. The first formula is: .in, This represents the standardized pixel Hu value. This represents the original Hu value of the pixel. Indicates the lower threshold. This represents the upper threshold. The lower threshold is preferably -200, and the upper threshold is preferably 700.

[0064] S200. The original cropped file is downsampled and then input into a pre-trained pre-segmentation network model for forward propagation to obtain a full receptive field prediction file containing the abdominal aortic region. The pre-segmentation network model is constructed based on a three-dimensional convolutional neural network and uses the abdominal aortic flow channel and abdominal aortic vessels as segmentation targets.

[0065] In some embodiments, the original 512×512×512 cropped file is downsampled based on intra-block averaging. The resulting file size after downsampling is 128×128×128. The full receptive field prediction file size output after prediction by the pre-segmentation network model is also 128×128×128. A specific implementation process for downsampling the original cropped file is as follows:

[0066] The original cropped file is divided into blocks of 4×4×4 size. Downsampling is achieved by averaging the pixels within each block. The intra-block averaging is based on the second formula shown below:

[0067] ,in, These are the pixel values ​​of the pixels within the divided region, located at the x, y, and z coordinates. This represents the pixel value after averaging within the block region.

[0068] In some embodiments, the forward propagation in step S200 is performed using the forward propagation method under the torch framework; during the training of the front segmentation network model, the sample dataset includes CT images under whole-body scans and corresponding annotation information, including Hu information characterizing the abdominal aortic vascular tissue.

[0069] Figure 2 A preferred architecture for a pre-segmentation network model is shown, specifically including: a first contraction path and a first expansion path that are symmetrical to each other, the first contraction path and the first expansion path are connected by convolution, and both the first contraction path and the first expansion path are generated based on a fully convolutional network.

[0070] The first shrinkage path, from top to bottom, includes an image input channel that takes the downsampled file from the original cropped file as input, a first merging model connected to the image input channel, a first downsampling network connected to the first merging model via convolution, a second downsampling network connected to the first downsampling network via convolution, a third downsampling network connected to the second downsampling network via convolution, and a fourth downsampling network connected to the third downsampling network via convolution. The first merging model performs voxel merging on the input file from the input channel. The convolution connection between the first merging model and the first downsampling network is performed using a filter with a stride of 1 and a size of 5*5*5, performing three convolutions. The convolution connection processes between the second downsampling network and the first downsampling network, the third downsampling network and the second downsampling network, and the fourth downsampling network and the third downsampling network are all the same as the convolution connection process between the first merging model and the first downsampling network. The first downsampling network, the second downsampling network, the third downsampling network, and the fourth downsampling network are all 2*2*2 downsampling networks. The first downsampling network has 16 channels, the second downsampling network has 32 channels, the third downsampling network has 64 channels, and the fourth downsampling network has 128 channels.

[0071] The first extension path, from bottom to top, includes a first upsampling network connected to the fourth downsampling network via convolution, a second upsampling network connected to the first upsampling network via convolution, a third upsampling network connected to the second upsampling network via convolution, a fourth upsampling network connected to the third upsampling network via convolution, a second merging model connected to the fourth upsampling network via convolution, and a first image output channel connected to the second merging model. The first image output channel is used to output the full receptive field prediction file. The fourth downsampling network and the first upsampling network are connected via three 5x5x5 convolutions. The convolution connection process between the first and second upsampling networks is as follows: three convolutions are performed using a filter with a stride of 1 and a size of 5x5x5. The convolution connection processes between the second and third upsampling networks, between the third and fourth upsampling networks, and between the fourth upsampling network and the second merging model are all the same as the convolution connection process between the first and second upsampling networks. The first, second, third, and fourth upsampling networks all use 2x2x2 upsampling. The first upsampling network has 128 channels, the second has 64 channels, the third has 32 channels, and the fourth has 16 channels. The second merging model is connected to the first image output channel after passing through a softmax layer and interpolation.

[0072] The output of the first merging model is summed with the output of the first downsampling network. The output of this summed model is then dropped out and summed with the output of the second downsampling network. The output of this summed model is then dropped out and summed with the output of the third downsampling network. The output of this summed model is then dropped out and summed with the output of the third downsampling network. The output of this summed model is then dropped out and summed with the output of the second downsampling network. The output of this summed model is then convolved with the first upsampling network. The output of the first upsampling network is then summed with the output of the second upsampling network. The output of this summed model is then summed with the output of the third upsampling network. The output of this summed model is then summed with the output of the fourth upsampling network. The output of this summed model is then summed with the output of the second merging model. The output of this summed model is then passed through a softmax layer and interpolated before being sent to the first image output channel.

[0073] Figure 2 In this context, m represents the artificial downsampling factor, preferably m=4; n represents the tensor size within the channels of the fourth downsampling network and the first upsampling network.

[0074] S300. Based on the preset segmentation step size, perform global segmentation sampling on the original cropped file to obtain multiple first sub-domain files.

[0075] In some embodiments, a specific implementation of S300 is as follows:

[0076] S301. Based on a preset segmentation step size, the original cropped file of size 512×512×512 is divided into N equal parts in the X, Y, and Z directions to obtain multiple first sub-domain files. Specifically, N is four, and with a segmentation step size of 128, a total of 64 first sub-domain files are obtained. The size of the first sub-domain file is 128×128×128, that is, the original cropped file is divided every 128 data units, with the starting data bits in the X, Y, and Z directions all being 0, and the ending data bits in the X, Y, and Z directions all being 511. Taking the X direction of the original cropped file as an example, the data bits of each segment obtained by the four equal parts are (0:127), (128:255), (256:383), and (384:511), respectively.

[0077] S400. Based on the preset segmentation step size, perform local segmentation sampling on the original cropped file to obtain multiple second subdomain files.

[0078] In some embodiments, a specific implementation of S400 is as follows:

[0079] S401. Based on the preset segmentation step size, the remaining pixel region after discarding the boundary in the X direction of the original cropped file of size 512×512×512 is divided into M equal parts, the remaining pixel region after discarding the boundary in the Y direction of the original cropped file is divided into M equal parts, and the remaining pixel region after discarding the boundary in the Z direction of the original cropped file is divided into M equal parts, to obtain multiple second sub-domain files. Specifically, M is set to three. Following a segmentation step of 128, the region between the 65th and 448th pixels in the X direction of the original cropped file (512×512×512 size) is divided into three equal parts; the region between the 65th and 448th pixels in the Y direction is also divided into three equal parts; and the region between the 65th and 448th pixels in the Z direction is also divided into three equal parts. This results in a total of 27 second sub-domain files, each 128×128×128 in size. That is, the original cropped file is divided every 128 data units, with the starting data position being 64 in each of the X, Y, and Z directions, and the ending data position being 447 in each of the X, Y, and Z directions. Taking the X direction of the original cropped file as an example, the data positions of the segments obtained from the three divisions are (64:191), (192:319), and (320:447). By discarding local segmentation sampling at the boundaries, the continuity of information between image blocks is ensured.

[0080] S500. Construct and train a subdomain prediction network model. The subdomain prediction network model takes the full receptive field prediction file, all first subdomain files, and all second subdomain files as input datasets. The subdomain prediction network model is built based on a three-dimensional convolutional neural network, and the abdominal aortic flow path and abdominal aortic vessels are used as segmentation targets.

[0081] In some embodiments, a specific implementation of the subdomain prediction network model using the full receptive field prediction file, all first subdomain files, and all second subdomain files as input datasets is as follows:

[0082] Based on the preset segmentation step size, global segmentation sampling and local segmentation sampling are performed on the full receptive field prediction file to obtain multiple third subdomain files;

[0083] Use all first subdomain files, all second subdomain files, and all third subdomain files as the input dataset for the subdomain prediction network model.

[0084] Specifically, based on a preset segmentation step size, global segmentation sampling and local segmentation sampling are performed on the full receptive field prediction file to obtain multiple third sub-domain files. The specific process is as follows:

[0085] Based on a preset segmentation step size, the 128×128×128 full receptive field prediction file is divided into N equal parts in the X, Y, and Z directions, resulting in multiple third subdomain files A. Specifically, N is set to four, and with a segmentation step size of 32, a total of 64 third subdomain files A are obtained. The size of each third subdomain file A is 32×32×32, meaning the full receptive field prediction file is segmented every 32 data units. The starting data bits are 0 in the X, Y, and Z directions, and the ending data bits are 127 in the X, Y, and Z directions. Taking the X direction of the full receptive field prediction file as an example, the data bits of each segment obtained from the four equal parts are (0:31), (32:63), (64:95), and (96:127).

[0086] Based on a preset segmentation step size, the remaining pixel region in the X direction of the 128×128×128 full receptive field prediction file after discarding the boundary is divided into M equal parts, the remaining pixel region in the Y direction of the full receptive field prediction file after discarding the boundary is divided into M equal parts, and the remaining pixel region in the Z direction of the full receptive field prediction file after discarding the boundary is divided into M equal parts, resulting in multiple third subdomain files B. Specifically, M is set to three. Using a segmentation step size of 32, the region between the 17th and 112th pixels in the X direction of the 128×128×128 full receptive field prediction file is divided into three equal parts; the region between the 17th and 112th pixels in the Y direction of the full receptive field prediction file is divided into three equal parts; and the region between the 17th and 112th pixels in the Z direction of the full receptive field prediction file is divided into three equal parts. This results in a total of 27 third sub-domain files B, each with a size of 32×32×32. That is, the full receptive field prediction file is divided every 32 data units. The starting data position is 16 in each of the X, Y, and Z directions, and the ending data position is 111 in each of the X, Y, and Z directions. Taking the X direction of the full receptive field prediction file as an example, the data positions of the segments obtained by the three divisions are (16:47), (48:79), and (80:111).

[0087] Third subdomain file A and third subdomain file B together constitute the third subdomain file.

[0088] Figure 3 A preferred architecture for a subdomain prediction network model is shown, specifically including: a second contraction path and a second expansion path that are mutually symmetrical, the second contraction path and the second expansion path are connected by convolution, and both the second contraction path and the second expansion path are generated based on a fully convolutional network.

[0089] The process of constructing a subdomain prediction network model based on the above-mentioned preferred architecture is as follows: the number of channels is doubled to 16 using the dimension augmentation method in the mathematical standard library NumPy; the scale of each layer in the neural network is adjusted by downsampling the data file size to obtain the optimized network under this round of training; the second contraction path is determined; the second expansion path is determined; and the model classification output interface is implemented through the softmax layer in the Torch framework.

[0090] The determined second contraction path, from top to bottom, includes a third merging model that takes all first and second subdomain files as input, a fifth downsampling network connected to the third merging model via convolution, a sixth downsampling network connected to the fifth downsampling network via convolution, a seventh downsampling network connected to the sixth downsampling network via convolution, and an eighth downsampling network connected to the seventh downsampling network via convolution. All third subdomain files are also used as input to the sixth downsampling network. The convolutional connection process between the third merging model and the fifth downsampling network involves three convolutions using a filter with a stride of 1 and a size of 5*5*5. The convolutional connection processes between the sixth and fifth downsampling networks, the seventh and sixth downsampling networks, and the eighth and seventh downsampling networks are all the same as the convolutional connection process between the third merging model and the fifth downsampling network. The fifth downsampling network has 16 channels, the sixth downsampling network has 32 channels, the seventh downsampling network has 64 channels, and the eighth downsampling network has 128 channels. All eight downsampling networks use a 2x2x2 downsampling method.

[0091] The determined second extension path, from bottom to top, includes a fifth upsampling network connected to the eighth downsampling network via convolution, a sixth upsampling network connected to the fifth upsampling network via convolution, a seventh upsampling network connected to the sixth upsampling network via convolution, an eighth upsampling network connected to the seventh upsampling network via convolution, a fourth merging model connected to the eighth upsampling network via convolution, and a second image output channel connected to the fourth merging model. The second image output channel is used to output the voxel segmentation model of the abdominal aorta. Specifically, the eighth downsampling network and the fifth upsampling network are connected via three 5x5x5 convolutions. The convolution connection process between the fifth and sixth upsampling networks involves three convolutions using a 5x5x5 filter with a stride of 1. The convolution connection processes between the sixth and seventh upsampling networks, between the seventh and eighth upsampling networks, and between the eighth upsampling network and the fourth merging model are all the same as those between the fifth and sixth upsampling networks. The fifth through eighth upsampling networks all use 2x2x2 upsampling. The fifth upsampling network has 128 channels, the sixth has 64, the seventh has 32, and the eighth has 16. The fourth merging model is connected to the second image output channel after passing through a softmax layer and interpolation.

[0092] The output of the third merging model is summed with the output of the fifth downsampling network. The output of this summed network is then dropped out and summed with the output of the sixth downsampling network. The output of this summed network is then dropped out and summed with the output of the seventh upsampling network. The output of this summed network is then dropped out and summed with the output of the seventh downsampling network. The output of this summed network is then dropped out and summed with the output of the sixth upsampling network. The output of this summed network is then convolved with the fifth upsampling network. The output of the fifth upsampling network is then summed with the output of the sixth upsampling network. The output of this summed network is then summed with the output of the seventh upsampling network. The output of this summed network is then summed with the output of the eighth upsampling network. The output of this summed network is then summed with the output of the fourth merging model. The output of this summed network is then passed through a softmax layer and interpolated before being passed to the second image output channel.

[0093] The subdomain prediction network model constructed above is trained based on the annotation files of the abdominal aortic flow channel and blood vessels. A specific training process is as follows: the annotation file is cropped according to the image localization points and the cropped file size; the Dice coefficient is used as the loss function to quantify the loss of the segmentation result; the Adam parameter update method is used for parameter updates; the dataset is divided into training and test sets in a 4:1 ratio; the training set is used as the model input for multiple rounds of training; the Dice loss function on the test set is used as the final training evaluation index for the subdomain prediction network model; the parameters of the subdomain prediction network model are updated based on the optimized parameters obtained after training, thereby generating the final subdomain prediction network model.

[0094] S600. Using the pre-segmentation network model and the trained subdomain prediction network model, target segmentation is performed on real-time CT sequence scan images to obtain a voxel segmentation model of the abdominal aorta. Figure 5 A planar schematic diagram of a voxel segmentation model is shown for comparison, showing the original image, the manually segmented image, and the deep learning segmentation image implemented based on this embodiment. Figure 8 A stereoscopic diagram of a voxel segmentation model is shown, illustrating the predicted results of the abdominal aortic flow channel and aneurysm.

[0095] Example 2

[0096] See Figures 6 to 7 The difference between this embodiment and Embodiment 1 is that, after S600, the following steps are included:

[0097] S700. The abdominal aortic voxel segmentation model obtained in Example 1 is smoothed and visualized.

[0098] One specific implementation process for smoothing and visualization is as follows:

[0099] S701. Perform opening and closing operations on the voxel segmentation model. The preferred step size for the opening and closing operations is 2.

[0100] S702. Use vtkMarchingCubes from the VTK library to convert the voxel segmentation model after performing opening and closing operations into a triangular facet model.

[0101] S703. Visualize and save the triangular facet model using the VTK library.

[0102] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An abdominal aorta automatic segmentation method based on a split sampling neural network, characterized in that, The method comprises the following steps: obtaining a historical CT sequence scanning image, performing image preprocessing and preset size cutting on the image, and obtaining an original cutting file; inputting the original cutting file into a pre-trained pre-segmentation network model after down-sampling to perform forward propagation, and obtaining a full receptive field prediction file containing an abdominal aorta interval; based on a preset segmentation step, performing global segmentation sampling on the original cutting file to obtain a plurality of first sub-domain files; based on a preset segmentation step, performing local segmentation sampling on the original cutting file to obtain a plurality of second sub-domain files; constructing and training a sub-domain prediction network model, wherein the sub-domain prediction network model takes the full receptive field prediction file, all first sub-domain files and all second sub-domain files as input data sets; using the pre-segmentation network model and the trained sub-domain prediction network model to perform target segmentation on a real-time CT sequence scanning image to obtain a voxel segmentation model of the abdominal aorta; wherein the pre-segmentation network model and the sub-domain prediction network model are both constructed based on a three-dimensional convolutional neural network, and both take the abdominal aorta flow passage and blood vessels as segmentation targets. The three-dimensional convolutional neural network comprises mutually symmetrical contraction paths and expansion paths, the contraction paths and the expansion paths are connected through convolution, and the contraction paths and the expansion paths are both generated based on a full convolutional network.

2. The abdominal aorta automatic segmentation method based on the split sampling neural network according to claim 1, characterized in that, The abdominal aorta automatic segmentation method further comprises:

3. The abdominal aorta automatic segmentation method based on split sampling neural network according to claim 1, characterized in that, performing smoothing processing and visualization on the voxel segmentation model. The image preprocessing comprises Hu value standardization based on an upper threshold and a lower threshold.

4. The abdominal aorta automatic segmentation method based on split sampling neural network according to claim 1, characterized in that, The pre-segmentation network model comprises a first contraction path and a first expansion path; 5. The abdominal aorta automatic segmentation method based on split sampling neural network according to claim 2, characterized in that, the first contraction path comprises, from top to bottom, an image input channel taking a file obtained by down-sampling the original cutting file as input, a first merging model connected with the image input channel, a first down-sampling network connected with the first merging model through convolution, a second down-sampling network connected with the first down-sampling network through convolution, a third down-sampling network connected with the second down-sampling network through convolution, and a fourth down-sampling network connected with the third down-sampling network through convolution; the first expansion path comprises, from bottom to top, a first up-sampling network connected with the fourth down-sampling network through convolution, a second up-sampling network connected with the first up-sampling network through convolution, a third up-sampling network connected with the second up-sampling network through convolution, a fourth up-sampling network connected with the third up-sampling network through convolution, a second merging model connected with the fourth up-sampling network through convolution, and a first image output channel connected with the second merging model, wherein the first image output channel is used to output the full receptive field prediction file; ​ The output of the first merging model will be summed with the output of the first down-sampling network, the output after this summation is dropped across the fourth up-sampling network and will also be summed with the output of the second down-sampling network, the output after this summation is dropped across the third up-sampling network and will also be summed with the output of the third down-sampling network, the output after this summation is dropped across the second up-sampling network and will also be summed with the output of the fourth down-sampling network, the output after this summation is connected with the first up-sampling network through convolution, the output of the first up-sampling network will be summed with the output of the second up-sampling network, the output after this summation will be summed with the output of the third up-sampling network, the output after this summation will be summed with the output of the fourth up-sampling network, the output after this summation will be summed with the output of the second merging model, and the output after this summation is transmitted to the first image output channel.

6. The abdominal aorta automatic segmentation method based on the segmentation sampling neural network according to claim 1, characterized in that, the original cutting file is globally segmented and sampled based on the preset segmentation step to obtain a plurality of first sub-domain files, specifically: the original cutting file is divided into N equal parts in the X direction, the Y direction and the Z direction based on the preset segmentation step to obtain a plurality of first sub-domain files; the original cutting file is locally segmented and sampled based on the preset segmentation step to obtain a plurality of second sub-domain files, specifically: the remaining pixel region after discarding the boundary in the X direction of the original cutting file is divided into M equal parts, the remaining pixel region after discarding the boundary in the Y direction of the original cutting file is divided into M equal parts, and the remaining pixel region after discarding the boundary in the Z direction of the original cutting file is divided into M equal parts based on the preset segmentation step to obtain a plurality of second sub-domain files.

7. The abdominal aorta automatic segmentation method based on split sampling neural network according to claim 2, characterized in that, the sub-domain prediction network model takes the full receptive field prediction file, all first sub-domain files and all second sub-domain files as input data sets, specifically: the full receptive field prediction file is globally segmented and sampled and locally segmented and sampled based on the preset segmentation step to obtain a plurality of third sub-domain files; all first sub-domain files, all second sub-domain files and all third sub-domain files are taken as input data sets of the sub-domain prediction network model.

8. The abdominal aorta automatic segmentation method based on the split sampling neural network according to claim 7, characterized in that, the sub-domain prediction network model includes a second contraction path and a second expansion path; the second contraction path includes, from top to bottom, a third merging model taking all first sub-domain files and all second sub-domain files as input, a fifth down-sampling network connected with the third merging model through convolution, a sixth down-sampling network connected with the fifth down-sampling network through convolution, a seventh down-sampling network connected with the sixth down-sampling network through convolution, an eighth down-sampling network connected with the seventh down-sampling network through convolution, and all third sub-domain files are taken as input of the sixth down-sampling network; The second expansion path comprises, from bottom to top, a fifth upsampling network connected to an eighth downsampling network through convolution, a sixth upsampling network connected to the fifth upsampling network through convolution, a seventh upsampling network connected to the sixth upsampling network through convolution, an eighth upsampling network connected to the seventh upsampling network through convolution, a fourth merging model connected to the eighth upsampling network through convolution, and a second image output channel connected to the fourth merging model, the second image output channel being used for outputting the voxel segmentation model; The output of the third merging model is summed with the output of the fifth downsampling network, the summed output is dropped to the eighth upsampling network, and is further summed with the output of the sixth downsampling network, the summed output is dropped to the seventh upsampling network, and is further summed with the output of the seventh downsampling network, the summed output is dropped to the sixth upsampling network, and is further summed with the output of the eighth downsampling network, the summed output is connected to the fifth upsampling network through convolution, the output of the fifth upsampling network is summed with the output of the sixth upsampling network, the summed output is summed with the output of the seventh upsampling network, the summed output is summed with the output of the eighth upsampling network, the summed output is summed with the output of the fourth merging model, and the summed output is transmitted to the second image output channel.

9. The abdominal aorta automatic segmentation method based on split sampling neural network according to claim 4, characterized in that, The Hu value normalization based on the upper threshold and the lower threshold is specifically: Hu value normalization according to the first formula in the mathematical standard library numpy, which is: ; wherein, represents the normalized pixel Hu value, represents the Hu original value of the pixel, represents the lower threshold value, represents the upper threshold value.

10. The abdominal aorta automatic segmentation method based on split sampling neural network according to claim 3, characterized in that, The voxel segmentation model is smoothed and visualized, specifically: The voxel segmentation model is subjected to opening and closing operations; The voxel segmentation model subjected to the opening and closing operations is converted into a triangular surface format model by using the VTK library; The triangular surface format model is visualized and saved by using the VTK library.

Images