CT image blood vessel segmentation method and device based on adaptive feature fusion
The adaptive feature fusion-based liver vessel segmentation method for CT images, utilizing the nnU-Net network model and multi-module design, solves the problems of noise interference and complex structure in liver vessel segmentation in liver CT images, achieving high-precision liver vessel segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2023-01-29
- Publication Date
- 2026-06-16
Smart Images
Figure CN116309308B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to a method for segmenting blood vessels in CT images based on adaptive feature fusion, and a device for segmenting blood vessels in CT images based on adaptive feature fusion. Background Technology
[0002] In clinical practice, information on the structure and morphology of liver vessels is crucial for accurate diagnosis of liver diseases and planning of liver surgery. In routine clinical practice, liver vessels are manually annotated by experts on each slice of CT images, consuming significant time and labor costs. Furthermore, manual annotation is highly dependent on the experience and skill of the experts. Therefore, there is an urgent need for an automated and accurate method for liver vessel segmentation in clinical applications. However, due to the irregular intensity distribution, low contrast, high image noise, and complex structure of liver CT images, automated and accurate liver vessel segmentation remains a challenge.
[0003] Traditional methods for liver vessel segmentation include vessel enhancement, active contouring, tracking, and machine learning. However, most of these methods rely heavily on initial parameter settings and vessel feature selection, making them prone to segmentation errors in complex liver vessel structures. In recent years, a series of deep learning-based methods have been proposed, achieving good performance, but these methods have three common limitations: First, the skip connections in U-shaped networks directly combine low-level and high-level features, introducing too much irrelevant background noise, making it difficult to distinguish liver vessels from noise, especially small vessels; second, most researchers use enhanced vessel images as segmentation priors without extracting the embedded vessel topological information, which could significantly improve segmentation quality; finally, most researchers use intrahepatic images for segmentation, neglecting the connectivity between intrahepatic and extrahepatic vessels, resulting in poor segmentation of vessels around the liver margin. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, the technical problem to be solved by the present invention is to provide a CT image blood vessel segmentation method based on adaptive feature fusion, which can suppress image background noise, accurately extract small blood vessels, improve segmentation integrity, and improve the segmentation accuracy of edge blood vessels.
[0005] The technical solution of this invention is: a CT image vessel segmentation method based on adaptive feature fusion, which includes the following steps:
[0006] (1) Preprocess the collected dataset;
[0007] (2) Construct a CT image liver vessel segmentation network model based on adaptive feature fusion;
[0008] (3) Input the preprocessed training set into the CT image liver vessel segmentation network model and train the network model to obtain the trained network model.
[0009] (4) Input the test image into the trained network model and output the liver blood vessel segmentation result of the test image.
[0010] This invention first preprocesses the acquired dataset; secondly, it constructs a CT image liver vessel segmentation network model based on adaptive feature fusion; then, it inputs the preprocessed training set into the CT image liver vessel segmentation network model to train the network model and obtain a trained network model; finally, it inputs the test image into the trained network model and outputs the liver vessel segmentation result of the test image. Therefore, it can suppress image background noise, accurately extract small blood vessels, improve segmentation integrity, and improve the segmentation accuracy of edge blood vessels.
[0011] A CT image vessel segmentation device based on adaptive feature fusion is also provided, which includes:
[0012] The preprocessing module is configured to preprocess the collected dataset;
[0013] The module is configured to build a CT image liver vessel segmentation network model based on adaptive feature fusion.
[0014] The training module is configured to input the preprocessed training set into the liver vessel segmentation network model of CT images, and train the network model to obtain a trained network model.
[0015] The output module is configured to input the test image into the trained network model and output the liver vessel segmentation result of the test image. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the CT image vessel segmentation method based on adaptive feature fusion according to the present invention.
[0017] Figure 2 This is a CT image liver vessel segmentation network model based on adaptive feature fusion according to the present invention.
[0018] Figure 3 An adaptive feature connection according to the present invention is shown. Detailed Implementation
[0019] The purpose of this invention is to provide a liver vessel segmentation method based on adaptive feature fusion in CT images to address the current challenges in liver vessel segmentation in CT images. This is mainly achieved by constructing an adaptive feature fusion-based CT image liver vessel segmentation network model to automatically and accurately segment liver vessels. The model uses the nnU-Net network as its main architecture and consists of three core modules. First, an adaptive feature connection module is designed to suppress image background noise and accurately extract small vessels. Second, an enhancement auxiliary module is proposed to fully utilize the topological information of the vessels and improve segmentation integrity. Finally, a global information supervision module is introduced to extract liver edge features and improve the segmentation accuracy of edge vessels.
[0020] like Figure 1 As shown, this CT image vessel segmentation method based on adaptive feature fusion includes the following steps:
[0021] (1) Preprocess the collected dataset;
[0022] (2) Construct a CT image liver vessel segmentation network model based on adaptive feature fusion;
[0023] (3) Input the preprocessed training set into the CT image liver vessel segmentation network model and train the network model to obtain the trained network model.
[0024] (4) Input the test image into the trained network model and output the liver blood vessel segmentation result of the test image.
[0025] This invention first preprocesses the acquired dataset; secondly, it constructs a CT image liver vessel segmentation network model based on adaptive feature fusion; then, it inputs the preprocessed training set into the CT image liver vessel segmentation network model to train the network model and obtain a trained network model; finally, it inputs the test image into the trained network model and outputs the liver vessel segmentation result of the test image. Therefore, it can suppress image background noise, accurately extract small blood vessels, improve segmentation integrity, and improve the segmentation accuracy of edge blood vessels.
[0026] Preferably, in step (1), a liver region is coarsely cut from the abdominal CT image to generate the original liver image I. origin Secondly, a liver mask image (I) was obtained by removing the background from the abdominal CT image using a liver segmentation mask. liver To avoid the influence of surrounding tissues of the liver; finally, in the liver mask image I liver The hepatic vascular enhancement image was obtained by using the Frangi filter. vessel .
[0027] Preferably, step (2) includes the following sub-steps:
[0028] (2.1) Construct an initial liver vessel segmentation model based on the nnU-Net network model;
[0029] An adaptive feature connection module (AFC) is designed in the encoder-decoder part of the network to achieve effective information fusion between high-level and low-level features;
[0030] (2.2) Constructing an enhancement-assisted module EA to fully utilize hepatic vascular enhancement images I vessel Information on the central vascular structure of the liver mask image I liver and enhanced liver vascular images I vessel The inputs are fed into two branches of the network, which have the same network weight parameters.
[0031] (2.3) Introduce the Global Information Supervision Module (GIS) to improve the segmentation of blood vessels around the liver edge.
[0032] Preferably, in step (2.1), the high-level feature f h First, the input is fed into the Spatial Attention Module (SAM), where average pooling and max pooling are performed along the channel direction of the feature map to obtain different pooling features. Then, these two pooling features are concatenated, followed by a 1×1×1 convolution and a sigmoid activation function to generate a spatial attention weight map A. Finally, attention weight map A is multiplied by the low-level feature f. l , to obtain f l ′; to f l ′ and f h Connecting them together yields the fused feature F.
[0033] Preferably, in step (2.3), I origin As input to another branch of the network, edge vessel information is obtained, resulting in I. origin_pred This includes information on intrahepatic and extrahepatic blood vessels; I origin The network branch does not share network parameters with the other two branches mentioned above, utilizing the liver mask I. liver_mask Cut I origin_pred Obtain vessel segmentation results with more liver edge information I origin_pred_crop Combined with I liver_pred and I origin_pred_crop To obtain accurate hepatic vessel segmentation results I output .
[0034] Preferably, in step (3), the preprocessed liver mask image I liver Enhanced liver vascular images I vessel And the original image of the liver I origin Simultaneously, the data is input into a liver vessel segmentation network model based on CT images for training, using a supervised loss function. To optimize network parameters, the formula is as follows:
[0035]
[0036] in, The calculation is the output result I. liver_pred With label I vessel_mask The segmentation loss includes Dice loss and cross-entropy loss:
[0037]
[0038] Where, p i Let g represent the i-th voxel predicted by the network. i Represents the i-th voxel of the label image. Calculation output result I vessel_pred and label I vessel_mask The segmentation loss between them Measurement output result I origin_pred and label I origin_mask The loss between the segments.
[0039] Preferably, in step (4), the test dataset is first preprocessed, and then the processed CT image is input into the trained liver vessel segmentation network model for segmentation to obtain the liver vessel segmentation result of the test image.
[0040] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a CT image vessel segmentation device based on adaptive feature fusion. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes:
[0041] The preprocessing module is configured to preprocess the collected dataset;
[0042] The module is configured to build a CT image liver vessel segmentation network model based on adaptive feature fusion.
[0043] The training module is configured to input the preprocessed training set into the liver vessel segmentation network model of CT images, and train the network model to obtain a trained network model.
[0044] The output module is configured to input the test image into the trained network model and output the liver vessel segmentation result of the test image.
[0045] Preferably, the building module performs the following steps:
[0046] (2.1) Construct an initial liver vessel segmentation model based on the nnU-Net network model;
[0047] An adaptive feature connection module (AFC) is designed in the encoder-decoder part of the network to achieve effective information fusion between high-level and low-level features;
[0048] (2.2) Constructing an enhancement-assisted module EA to fully utilize hepatic vascular enhancement images I vessel Information on the central vascular structure of the liver mask image I liver and enhanced liver vascular images I vessel The inputs are fed into two branches of the network, which have the same network weight parameters.
[0049] (2.3) Introduce the Global Information Supervision Module (GIS) to improve the segmentation of blood vessels around the liver edge.
[0050] Preferably, the training module performs:
[0051] Preprocessed liver mask image I liver Enhanced liver vascular images I vessel And the original image of the liver I origin Simultaneously, the data is input into a liver vessel segmentation network model based on CT images for training, using a supervised loss function. To optimize network parameters, the formula is as follows:
[0052]
[0053] in, The calculation is the output result I. liver_pred With label I vessel_mask The segmentation loss includes Dice loss and cross-entropy loss:
[0054]
[0055] Where, p i Let g represent the i-th voxel predicted by the network. i Represents the i-th voxel of the label image. Calculation output result I vessel_pred and label I vessel_mask The segmentation loss between them Measurement output result I origin_pred and label I origin_mask The loss between the segments.
[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A CT image vessel segmentation method based on adaptive feature fusion, characterized in that: It includes the following steps: (1) Preprocess the collected dataset; (2) Construct a CT image liver vessel segmentation network model based on adaptive feature fusion; (3) Input the preprocessed training set into the CT image liver vessel segmentation network model to train the network model and obtain the trained network model. (4) Input the test image into the trained network model and output the liver vessel segmentation result of the test image; In step (1), a liver region is coarsely cut from the abdominal CT image to generate the original liver image. Secondly, a liver mask image was obtained by removing the background from the abdominal CT image using a liver segmentation mask. To avoid the influence of surrounding tissues; finally, in the liver mask image. The hepatic vascular enhancement image was obtained by using the Frangi filter. ; Step (2) includes the following sub-steps: (2.1) Construct an initial liver vessel segmentation model based on the nnU-Net network model; design an adaptive feature connection module (AFC) in the encoder-decoder part of the network to achieve effective information fusion of high-level features and low-level features; (2.2) Constructing an enhancement-assisted module (EA) to fully utilize hepatic vascular enhancement images Information on the central vascular structure of the liver mask image. and enhanced images of liver vessels The inputs are fed into two branches of the network, which have the same network weight parameters. (2.3) Introduce the Global Information Supervision Module (GIS) to improve the segmentation of blood vessels around the liver edge.
2. The CT image vessel segmentation method based on adaptive feature fusion according to claim 1, characterized in that: In step (2.1), high-level features First, the input is fed into the Spatial Attention Module (SAM), where average pooling and max pooling are performed along the channel direction of the feature map to obtain different pooled features. The two are then concatenated, followed by a 1×1×1 convolution and a sigmoid activation function to generate a spatial attention weight map. A； Pay attention to the weight graph A Multiply by low-level features ,get ;Will and Connect them to obtain the fusion features F .
3. The CT image vessel segmentation method based on adaptive feature fusion according to claim 2, characterized in that: In step (2.3), the following steps will be taken: As input to another branch of the network, edge vessel information is obtained, resulting in... This includes information on intrahepatic and extrahepatic blood vessels; The network branch does not share network parameters with the other two branches mentioned above, utilizing the liver mask. Cutting Obtain vessel segmentation results with more information about the liver margins. , combined and To obtain accurate results of hepatic vessel segmentation .
4. The CT image vessel segmentation method based on adaptive feature fusion according to claim 3, characterized in that: In step (3), the preprocessed liver mask image is... Enhanced images of liver blood vessels and raw liver images Simultaneously, the data is input into a liver vessel segmentation network model based on CT images for training, using a supervised loss function. To optimize network parameters, the formula is as follows: ， in, The calculation is the output result. With tags The segmentation loss includes Dice loss and cross-entropy loss: ， in, The network prediction of the first i Individual factors, The first label image represents the first label image. i Individual factors, Calculation output results and tags The segmentation loss between them Measurement output results and tags The loss between the segments.
5. The CT image vessel segmentation method based on adaptive feature fusion according to claim 4, characterized in that: In step (4), the test dataset is first preprocessed, and then the processed CT image is input into the trained liver blood vessel segmentation network model for segmentation to obtain the liver blood vessel segmentation result of the test image.
6. The apparatus for CT image vessel segmentation based on adaptive feature fusion according to claim 1, characterized in that: It includes: The preprocessing module is configured to preprocess the collected dataset; The module is configured to build a CT image liver vessel segmentation network model based on adaptive feature fusion. The training module is configured to input the preprocessed training set into the liver vessel segmentation network model of CT images, and train the network model to obtain a trained network model. The output module is configured to input the test image into the trained network model and output the liver vessel segmentation result of the test image. The building module performs the following steps: (2.1) Construct an initial liver vessel segmentation model based on the nnU-Net network model; design an adaptive feature connection module (AFC) in the encoder-decoder part of the network to achieve effective information fusion of high-level features and low-level features; (2.2) Constructing an enhancement-assisted module (EA) to fully utilize hepatic vascular enhancement images Information on the central vascular structure of the liver mask image. and enhanced images of liver vessels The inputs are fed into two branches of the network, which have the same network weight parameters. (2.3) Introduce a global information supervision module (GIS) to improve the segmentation of blood vessels around the liver margin; The training module executes: Preprocessed liver mask image Enhanced images of liver blood vessels and raw liver images Simultaneously, the data is input into a liver vessel segmentation network model based on CT images for training, using a supervised loss function. To optimize network parameters, the formula is as follows: ， in, The calculation is the output result. With tags The segmentation loss includes Dice loss and cross-entropy loss: ， in, The network prediction of the first i Individual factors, The first label image represents the first label image. i Individual factors, Calculation output results and tags The segmentation loss between them Measurement output results and tags The loss between the segments.