A ccta-based combined segmentation method for cto lesions and coronary trees
By employing a combined segmentation method of CTO lesions and coronary artery tree based on CCTA, and utilizing residual network and boundary point detection technology, the centerline of the CTO plaque breakpoint segment is extracted, solving the problem of incomplete segmentation of coronary CTO plaques and achieving complete coronary artery tree segmentation, thus supporting clinical treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2023-12-22
- Publication Date
- 2026-07-21
AI Technical Summary
In existing technologies, segmentation of coronary artery CTO plaques is difficult and incomplete, resulting in a decline in the quality of automatic coronary tree segmentation results and the inability to obtain a complete coronary tree.
A combined segmentation method based on CCTA and coronary artery tree was adopted. Coronary artery segmentation was performed using a residual network model. By combining boundary point detection and CTO plaque mask annotation, the coronary artery centerline of the CTO plaque breakpoint segment was extracted, and CTO plaque segmentation was performed. Finally, the results were stitched together to form a complete coronary artery tree.
It achieves complete and independent segmentation of CTO plaques, solves the problem of inaccurate segmentation of CTO plaques, obtains complete coronary tree structure, and supports the formulation of clinical treatment plans.
Smart Images
Figure CN117726632B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology, and in particular relates to a method for joint segmentation of CTO lesions and coronary artery tree based on CCTA. Background Technology
[0002] Coronary artery disease (CAD) is one of the most common types of cardiovascular disease. Chronic total occlusion (CTO) of the coronary arteries refers to a form of CAD where the diameter of one or more branches of the coronary arteries is completely blocked for more than three months. Clinical studies have shown that CTO accounts for one-third of all patients undergoing coronary angiography and 10%–20% of all patients undergoing interventional cardiac procedures.
[0003] Coronary computed tomography angiography (CCTA) generates three-dimensional images of the coronary arteries by introducing contrast agents and performing X-ray imaging, providing detailed information about the coronary arteries. Since some occluded segments are difficult to observe manually, CTA provides valuable information about CTO characteristics, such as lesion length and degree of calcification. This information helps physicians better understand the condition of the CTO and develop more effective treatment plans.
[0004] With the introduction and development of neural network concepts, deep learning (DL) has made significant progress in tissue segmentation, lesion detection, and disease identification in medical image processing over the past decade. In the field of coronary artery segmentation, numerous related models have been proposed, and current algorithms and commercial workstations can achieve relatively complete automatic segmentation of coronary artery trees. However, CTO plaques, as a type of plaque that completely blocks the coronary arteries, often lead to a decline in the quality of automatic coronary artery segmentation results, or even fragmentation, thus preventing the acquisition of a complete coronary artery tree. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a combined segmentation method for CTO lesions and coronary artery tree based on CCTA, thereby resolving the issues existing in the prior art.
[0006] To achieve the above objectives, this invention provides a method for combined segmentation of CTO lesions and coronary artery tree based on CCTA, comprising the following steps:
[0007] Acquire coronary artery CCTA images, and annotate the complete coronary artery mask and CTO plaque mask of the coronary artery CCTA images to obtain annotated images. Construct a residual network model to segment the annotated images to obtain the segmentation results of healthy coronary arteries.
[0008] Boundary point detection is performed on the coronary artery CCTA images to obtain coronary artery breakpoints. The coronary artery breakpoints are compared with the CTO plaque mask annotations in the labeled images to confirm the CTO plaque breakpoints.
[0009] Based on the tubular features of the coronary arteries, the coronary centerline of the CTO site is extracted from the CTO plaque breakpoint segment. Based on the coronary centerline of the CTO site, CTO plaque segmentation is performed to obtain the CTO plaque segmentation result.
[0010] The CTO plaque segmentation results and the healthy coronary artery segmentation results are stitched together according to the image data hierarchy to obtain a complete coronary tree result.
[0011] Optionally, the method for obtaining the segmentation results of healthy coronary arteries includes:
[0012] Doctors annotated the coronary artery CCTA images with complete coronary artery masks and CTO plaque masks, and then divided the annotated coronary artery CCTA images into training set, validation set and test set;
[0013] The annotated coronary CCTA images are subjected to data augmentation and resampling processes, including rotation, flipping, and cropping.
[0014] The initial residual network model is constructed based on a three-dimensional convolutional neural network. The construction method includes: introducing residual blocks into the convolutional blocks of the encoder and decoder. The residual blocks include 3D convolution, BN layers and linear correction units with skip connections.
[0015] The initial residual network model is trained using the training and validation sets, and then tested using the test set to obtain the residual network model. The residual network model is then used to segment the data-augmented coronary artery CCTA image to obtain the segmentation result of the healthy coronary arteries.
[0016] Optionally, methods for obtaining CTO patch breakpoints through comparison include:
[0017] By comparing the breakpoint with the labeled CTO mask through boundary point detection, the image slice at the breakpoint is searched pixel by pixel to see if there is any overlap between the image slice and the CTO mask. If there is no overlap, it is the endpoint of the normal coronary artery structure. If there is an overlap, it is the breakpoint caused by the CTO plaque. The endpoint of the normal coronary artery structure is removed to obtain the CTO plaque breakpoint.
[0018] Optionally, the method for extracting the coronary artery centerline of the CTO plaque breakpoint segment includes:
[0019] Based on the corresponding masks of coronary arteries and CTO plaques, the skeletonization algorithm is used to automatically determine the position of the coronary artery center point at both ends of the rupture site. The local geometric features of the coronary arteries are corrected for the center line. The coronary artery center point between the coronary artery center points at both ends of the rupture site is determined by the central likelihood function. All coronary artery center points are connected together to obtain the coronary artery center line.
[0020] Optionally, the expression for the central likelihood function is:
[0021]
[0022] In the formula, r1 li and r2 li denoted as ray, representing the ray from the edge of the coronary artery to any point within the cross-section of the coronary artery, where n represents the total number of rays and CL represents the likelihood value.
[0023] Optionally, the method for obtaining CTO patch segmentation results includes:
[0024] Based on the coronary artery centerline at the CTO site, CCTA image data slices and CTO plaque mask data corresponding to the CTO plaque were collected to obtain an image dataset;
[0025] The image dataset is preprocessed to obtain a preprocessed dataset;
[0026] An initial segmentation model is constructed, and the patch segmentation model is trained using the preprocessed dataset to obtain the CTO patch segmentation model;
[0027] CTO patch segmentation is performed using the CTO patch segmentation model to obtain CTO patch segmentation results.
[0028] Optionally, the preprocessing method includes: randomly dividing the acquired coronary CCTA images into a training set, a validation set, and a test set; performing data augmentation operations, including random rotation, flipping, and cropping, on the divided coronary CCTA images; and uniformly resampling the images after data augmentation to obtain a preprocessed dataset.
[0029] Optionally, the method for constructing the initial segmentation model includes:
[0030] Construct an encoder network, using DenseNet and dilated convolutions instead of ordinary CNN convolutional blocks;
[0031] A perception enhancement module, including a multi-scale feature extraction module and a channel attention module, is added to the end of the encoder network;
[0032] By adding a skip connection between the encoder and decoder, an initial segmentation model is obtained.
[0033] Optionally, the expression for the multi-scale feature extraction module is:
[0034]
[0035] Where conv1 represents a 1×1 convolution, Cai(·) represents the feature fusion function, and Cai(·) represents a 3×3 dilated convolution with a shrinkage rate of i. This indicates input.
[0036] Optionally, the loss function expression of the CTO patch segmentation model is:
[0037]
[0038] In the formula, y n This is the true label corresponding to the nth pixel. This is the prediction result for the nth pixel, where N is the total number of sample pixels.
[0039] Compared with the prior art, the present invention has the following advantages and technical effects:
[0040] The CCTA-based CTO lesion and coronary tree segmentation method described in this invention solves the problems of difficult and incomplete segmentation of CTO plaques in the coronary arteries. An improved model is proposed for CTO plaques, which can achieve complete and independent segmentation of CTO plaques, solving the problem of the lack of accurate segmentation research on CTO plaques, thereby enabling the acquisition of a complete coronary tree. Attached Figure Description
[0041] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0042] Figure 1 This is a flowchart of the CCTA-based CTO lesion and coronary artery tree joint segmentation method according to an embodiment of the present invention;
[0043] Figure 2 This is a schematic diagram of the residual block principle of the coronary artery segmentation network according to an embodiment of the present invention;
[0044] Figure 3 This is a flowchart illustrating the implementation of the coronary artery segmentation method according to an embodiment of the present invention;
[0045] Figure 4 This is a schematic diagram illustrating the principle of extracting the centerline of the coronary artery segmentation and rupture site according to an embodiment of the present invention;
[0046] Figure 5This is a schematic diagram of the overall structure of the coronary artery CTO plaque segmentation neural network according to an embodiment of the present invention;
[0047] Figure 6 This is a schematic diagram of the perception enhancement module in the coronary artery CTO plaque segmentation network according to an embodiment of the present invention;
[0048] Figure 7 This is a schematic diagram of the complete coronary artery prediction results according to an embodiment of the present invention. Detailed Implementation
[0049] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0050] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0051] Example 1
[0052] This embodiment proposes a method for combined segmentation of CTO lesions and coronary artery tree based on CCTA, including the following steps:
[0053] The first stage involves coronary artery segmentation, the second stage involves CTO segment detection, the third stage involves CTO segment centerline growth based on the centerline iteration model, the fourth stage involves extracting CCTA images based on the detected coronary artery rupture areas and performing coronary artery CTO plaque segmentation using the grown centerline, and the fifth stage involves stitching and reconstructing the CTO segmentation results with the coronary artery segmentation results to obtain a complete coronary tree including CTO plaques.
[0054] Specifically:
[0055] The first stage includes: acquiring and segmenting coronary artery image data, with doctors annotating complete coronary artery masks and CTO plaque masks. An optimized residual network model is then constructed to obtain preliminary coronary artery segmentation results.
[0056] The second stage includes: when the tubular structure of the coronary artery is ruptured, the doctor determines whether the rupture is caused by the presence of CTO plaque based on the CTO mask and the boundary point detection unit. If the coronary artery structure segmentation fails due to the presence of CTO plaque, CTO plaque segmentation and reconstruction will be performed.
[0057] The third stage includes: extracting the coronary artery centerline of the CTO plaque segment based on the tubular features of the coronary arteries. First, the endpoints of the coronary artery centerline at both ends of the rupture site are given based on the coronary artery and the CTO plaque mask. Then, the coronary artery centerline is obtained point by point using the tubular local structural features of the coronary artery and the center likelihood function to obtain a complete and continuous coronary artery centerline of the CTO site.
[0058] The fourth stage includes: extracting corresponding CCTA images based on the centerline of the grown coronary artery CTO site, segmenting the CTO plaque using an improved convolutional model, and integrating the results with the coronary artery segmentation. The model incorporates a dilated convolution module, multi-scale feature extraction, and a hybrid attention module, which ensures detection accuracy when detecting small targets. While preserving sufficient spatial information, it fuses deep and shallow information to achieve high-precision segmentation of CTO plaques.
[0059] The fifth stage includes stitching the CTO plaque segmentation results together with the coronary artery tree segmentation results according to the image data hierarchy to obtain a complete coronary artery tree result including the CTO plaque.
[0060] Furthermore, the coronary artery segmentation network phase includes:
[0061] (1) Acquire patient CCTA images. The physician performs two types of annotation: coronary artery mask and CTO plaque mask. The coronary artery mask only includes the corresponding coronary artery portion against a black background, including all left and right coronary arteries and their branches. The physician also needs to annotate the CTO plaques; the CTO mask is independent of the coronary artery mask. The images are randomly divided into training, validation, and test sets. The window width and level of the CCTA images are uniformly adjusted, and data augmentation processing including rotation, flipping, and cropping is performed. Resampling is used to ensure that each dataset is of the same size.
[0062] (2) Based on the 3D-Unet network, residual blocks are introduced into the convolutional blocks of the encoder and decoder to solve the vanishing and exploding gradient problems. Each residual block consists of a 3D convolution, a BN layer, and a linear correction unit with skip connections. The decoder is terminated by a 1×1 convolutional layer with a sigmoid activation function. The constructed neural network is as follows: Figure 5 As shown;
[0063] (3) Using a coronary artery mask, the pre-defined training and validation sets are input into the network for training. The optimal model is then validated using test set data. The two mean squared error (MSE) loss function can improve the connectivity-aware segmentation results. Complete segmentation results of healthy coronary artery segments are obtained.
[0064] Furthermore, after the above steps, the boundary point detection module can determine whether there are any breakpoints in the coronary artery segmentation results. Since the obtained coronary artery mask is complete, the coronary artery segmentation results should be complete, meaning that the coronary arteries and their related branches and small vessels are completely connected, and the breakpoints are only due to CTO plaques. The boundary point detection module compares the correlation between the breakpoint location and the CTO mask. If the breakpoint location is at the top of the coronary artery or is unrelated to the CTO plaque location, it is discarded. If the coronary artery rupture is indeed caused by the CTO plaque, the next step of obtaining the CTO centerline is performed.
[0065] Furthermore, the centerline acquisition phase includes:
[0066] (1) When the coronary artery is interrupted, the position of the center point of the coronary artery at both ends of the rupture site is automatically determined by the skeletonization algorithm based on the corresponding mask of the coronary artery and CTO plaque.
[0067] (2) Centerline correction of local geometric features of the coronary arteries. The coronary artery centerline is continuously searched and judged point by point along the direction of the two endpoints of the centerline, and the central likelihood function is used to determine the coronary artery center point between the two searchable points. By continuously connecting the judged vessel center points, the coronary artery centerline of the CTO plaque site can be obtained, and the corresponding CCTA image is extracted.
[0068] Furthermore, the CTO plaque segmentation stage includes:
[0069] (1) Data preprocessing. The acquired image dataset was randomly divided into training, validation, and test sets. Data augmentation operations, including random rotation, flipping, and cropping, were performed on the original coronary CCTA images. The CCTA images were then resampled to a uniform size.
[0070] (2) Network Construction. Coronary artery image samples are used as input, and CTO plaque segmentation results are output. The U-shaped basic architecture of the U-Net network is adopted, and the network structure includes an encoder, a decoder, a multi-scale feature extraction module, and a channel attention module. Dilated convolutional layers are added to the encoder, and a perceptual enhancement module (PE) including a multi-scale feature extraction module and a channel attention module for enhancing features is used at the end of the encoder. Upsampling is performed using the decoder;
[0071] (3) The multi-scale feature extraction module includes four branches: one branch with a 1×1 convolutional kernel and three branches with 3×3 dilated convolutions, with dilation rates of 1, 2, and 3, respectively. Connecting the feature maps from different branches yields multi-scale features containing rich semantic information. The calculation formula is as follows:
[0072]
[0073] Where conv1 represents a 1×1 convolution, φ represents the feature fusion function, and Cai(·) represents a 3×3 dilated convolution with a shrinkage rate of i. For input;
[0074] (4) Model Training. The labeled training and validation sets of CCTA image data are input into the patch segmentation model for training. An improved loss function combining BCEloss and Diceloss is used for model training. The formula for the improved loss function is as follows:
[0075]
[0076] Where y n This is the true label corresponding to the nth pixel. The prediction result for the nth pixel is given, where N is the total number of sample pixels. Fusing two different loss functions can comprehensively consider pixel-level classification accuracy (BCE loss) and segmentation result similarity (Dice loss), improving the segmentation accuracy for smaller targets and those with unclear edges.
[0077] The network underwent multiple rounds of training until it converged. Model performance was evaluated using a test set and relevant evaluation metrics. The optimal model parameters were selected, and 3D CTO patch results were obtained through segmentation.
[0078] Furthermore, the coronary tree stitching module integrates and stitches the segmented 3D plaque data with the coronary artery data segmented in the first stage to obtain a complete coronary artery structure, including CTO plaques.
[0079] Example 2
[0080] like Figure 1 As shown, this embodiment proposes a method for combined segmentation of CTO lesions and coronary artery tree based on CCTA, including the following steps:
[0081] Coronary artery CCTA images were acquired, and complete coronary artery masks and CTO plaque masks were annotated on the coronary artery CCTA images to obtain annotated images. A residual network model was constructed to segment the annotated images to obtain the segmentation results of healthy coronary arteries.
[0082] Boundary point detection was performed on the coronary artery CCTA images to obtain the coronary artery breakpoints. The coronary artery breakpoints were compared with the CTO plaque mask annotations in the labeled images to obtain the CTO plaque breakpoints.
[0083] Based on the tubular features of the coronary arteries, the coronary centerline of the CTO site is extracted from the segment of the CTO plaque breakpoint. CTO plaque segmentation is then performed based on the coronary centerline of the CTO site to obtain the CTO plaque segmentation results.
[0084] The CTO plaque segmentation results and the healthy coronary artery segmentation results are stitched together according to the image data hierarchy to obtain the complete coronary tree results.
[0085] The specific implementation steps are as follows:
[0086] After obtaining the coronary artery mask and CTO plaque mask, the coronary artery segmentation module is used to perform complete segmentation of the non-CTO portion of the coronary artery. The coronary artery segmentation result does not include the CTO plaque area.
[0087] The boundary point detection module compares the breakpoints in the coronary artery segmentation results with the CTO mask to determine whether the breakpoints are caused by CTO plaques. If the breakpoints coincide with the CTO mask, the next step, centerline acquisition, is performed.
[0088] The centerline acquisition module determines the center point positions of the coronary arteries at both ends of the rupture site based on the corresponding mask of the coronary artery and CTO plaque. Then, it uses the center likelihood function to search for the center points point by point and connect them to obtain the centerline.
[0089] The CTO patch segmentation module obtains the corresponding slices of the CCTA image based on the extracted centerline, and uses the improved 3DU-Net model to segment and obtain three-dimensional CTO patch data.
[0090] The coronary tree stitching module stitches together the segmented CTO plaque data with the previously segmented coronary artery results to obtain a complete coronary tree result including the CTO plaques.
[0091] In this embodiment, the collected patient data are CCTA image data of CTO patients, and the corresponding mask includes coronary arteries and CTO plaques. First, all images are randomly divided into training, validation, and test sets in a 6:2:2 ratio. The image window width and level are uniformly adjusted to 700HU / 200HU to facilitate the extraction of coronary artery features. Data augmentation operations, including rotation, flipping, and cropping, are performed on all CCTA images, followed by resampling to a size of 256×256.
[0092] In this embodiment, a deep learning network incorporating residual blocks is used for coronary artery segmentation to obtain a three-dimensional coronary artery image. Residual convolution is used instead of ordinary convolution in the network encoder and decoder. Specifically, the following steps are included:
[0093] (1) The encoder has four residual blocks. Each residual block consists of a BN layer, two 3×3 3D convolutional layers, a RuLu nonlinear activation function, and a linear correction unit with skip connections. The specific structure is as follows: Figure 2As shown. Each residual block is followed by a 2×2 max pooling operation. The three residual blocks have 32, 64, 128, and 256 convolutional kernels, respectively. A residual block with 256 convolutional kernels is connected at the end of the encoder;
[0094] (2) The decoder has four 2×2 upsampling layers, each followed by a residual block identical to the corresponding layer in the decoder. A 1×1 convolutional layer with a sigmoid activation function is added at the decoder output. The overall network model is as follows: Figure 3 As shown, the output shows the coronary artery segmentation results;
[0095] (3) Training was performed on all patients for 80 epochs using the training and validation sets. The loss function was mean squared error (MSE), which is the mean of the sum of squares of the differences between the predicted value f(x) and the target value y.
[0096]
[0097] The Adam optimizer is used with a batch size of 4. Validation is performed using a validation set, with mIoU as the evaluation metric. mIoU represents the average intersection-union ratio (IU) of each class in the dataset, calculated as follows:
[0098]
[0099] Where k represents the total number of categories, p ij p represents the number of cases where the true value is i and the prediction is j, where k+1 is the number of classes (including the empty class). ii This is the number of true positives. ij p ji These represent false positives and false negatives, respectively.
[0100] In this embodiment, a boundary point detection module is used to compare the correlation between the breakpoint locations of the coronary artery segmentation results and the CTO plaque locations. The breakpoint locations of the coronary artery segmentation results are compared with the CTO mask, and a pixel-by-pixel search is performed to check whether the slice at the breakpoint overlaps with the CTO mask, determining whether the breakpoint is caused by a CTO plaque. If the breakpoint is not related to a CTO plaque, it is excluded, for example, if it is at the tip of a coronary artery branch; if the coronary artery rupture is indeed caused by a CTO plaque, the next step is to obtain the centerline of the CTO location.
[0101] In this embodiment, centerline extraction is performed on the coronary artery portion where CTO plaques are detected. Specifically, this includes the following steps:
[0102] (1) When the coronary artery is interrupted, the skeletonization algorithm is used to automatically determine the position of the center point of the coronary artery at both ends of the rupture site based on the corresponding mask of the coronary artery and CTO plaque. First, the single-layer CCTA image slice corresponding to the junction of the coronary artery mask and the CTO plaque mask is extracted. The skeletonization algorithm traverses all non-zero pixels in the CTO mask corresponding to the slice, removes edge pixels, and retains only the center point as the upper and lower endpoints of the center line;
[0103] (2) Centerline correction of local geometric features of the coronary arteries. A point-by-point search and assessment is performed along the direction of the two endpoints of the coronary artery centerline, based on the CTO plaque location in the CCTA imaging data, which generally has a slightly higher CT value. The central likelihood function CL is used to determine the coronary artery center point between the two searchable points; the calculation formula is as follows:
[0104]
[0105] This means that if a ray is emitted from any point within the cross-section of the coronary artery and reaches the edge of the coronary artery, for each set of rays, there are two rays r1 with opposite directions. li and r2 li , where n is the total number of rays. The closer CL is to 1, the closer this point is to the center of the coronary artery.
[0106] (3) By continuously connecting the determined coronary artery center points, the coronary artery centerline at the CTO plaque site can be obtained, such as... Figure 4 As shown.
[0107] In this embodiment, after extracting the centerline, CTO patch segmentation is performed, specifically including the following steps:
[0108] (1) Based on the extracted CTO centerline, obtain the corresponding CCTA image data slices and their CTO patch masks. The images are randomly divided into training, validation, and test sets in a 6:2:2 ratio.
[0109] (2) Constructing the encoder network. The encoder network mainly consists of CNN convolutional layers, batch normalization layers, and DenseNet convolutional layers. Each convolutional layer uses the ReLU activation function to learn non-linear features. DenseNet uses dense connections to improve gradient backpropagation. Two DenseBlocks (basic modules) are added to the DenseNet convolutional layers to prevent overfitting while ensuring a large number of parameters. Dilated convolutions are used instead of ordinary convolutional blocks to increase the receptive field and obtain multi-scale contextual information while extracting features.
[0110] (3) Add a perception enhancement module (PE) to the end of the encoder network, such as Figure 6As shown, it consists of a multi-scale feature extraction module and a channel attention module for enhancing features;
[0111] (3.1) The multi-scale feature extraction module can improve segmentation accuracy by extracting semantic information from different receptive fields, and can increase the receptive field size without significantly increasing computational cost and reducing resolution loss. It includes four branches: one branch with a 1×1 convolutional kernel and three branches with 3×3 dilated convolutions, with dilation rates of 1, 2, and 3, respectively. Figure 6 As shown. Connecting the feature maps of different branches yields multi-scale features containing rich semantic information, calculated using the following formula:
[0112]
[0113] Where conv1 represents a 1×1 convolution, φ represents the feature fusion function, and Cai(·) represents a 3×3 dilated convolution with a shrinkage rate of i. For input;
[0114] (3.2) The channel attention module (SE) can model channel attention and calculate the relationship between different channels. It stretches the original image through global pooling and fully connected layers, and then multiplies it with the original image to assign different weights to each channel, thus solving the loss problem caused by different channels having different importance.
[0115] (4) Add skip connections between the corresponding network layers of the encoder and decoder to add original features;
[0116] (5) Input the labeled training and validation sets into the model for multiple rounds of training. Use the BCEdiceloss loss function for training, and combine Diceloss and BCEloss to evaluate the consistency between the network's prediction results and the true labels. BCEloss is binary cross-entropy, which can be used to measure the difference between the predicted probability distribution and the target probability distribution, and penalize the absolute difference in predicted probabilities. The Dice coefficient measures the similarity or overlap between the predicted results and the target. Combining the two loss functions can balance classification accuracy and target region shape matching during training. The specific formula is as follows:
[0117]
[0118] Where y n This is the true label corresponding to the nth pixel. Let N be the predicted result for the nth pixel, where N is the total number of sample pixels. Fusing two different loss functions allows for a comprehensive consideration of pixel-level classification accuracy (BCE loss) and segmentation result similarity (Dice loss).
[0119] (6) After generating the three-dimensional segmentation results of CTO plaques, quantitative analysis of plaques can be performed, plaque volume and length can be calculated, etc., to provide guidance for clinical treatment.
[0120] In this embodiment, the segmented CTO plaque results are stitched together layer by layer and block by block into the coronary artery segmentation results according to imaging features, such as the CCTA image slice order and centerline position, to obtain a complete three-dimensional segmentation result of the coronary artery including both the coronary artery and the CTO plaque. Figure 7 As shown.
[0121] In summary, the significant advantages of this invention compared to existing technologies are summarized as follows:
[0122] 1) In this embodiment of the invention, the invention combines several different modules, including coronary artery segmentation, centerline extraction, and CTO plaque segmentation, to achieve complete coronary artery segmentation, including CTO plaques, and to enable quantitative analysis of CTO plaques. The technical solution of this invention is relatively simple in engineering and highly feasible. During code training, only the corresponding masks need to be obtained for the coronary arteries and the regions corresponding to CTO plaques. In practical applications, the optimal results of the trained model can be directly used to achieve complete coronary tree segmentation, including CTO plaques.
[0123] 2) An improved model is proposed for CTO patches, which can achieve complete and independent segmentation of CTO patches, solving the current problem of no accurate segmentation research for CTO patches. In particular, the addition of dilated convolution to the model can increase the receptive field; multi-scale feature extraction can improve segmentation accuracy by extracting semantic information from different receptive fields; and the channel attention module can calculate the relationship between different channels of the feature map and assign different weights to solve the loss problem caused by different channels having different importance.
[0124] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for combined segmentation of CTO lesions and coronary artery tree based on CCTA, characterized in that, Includes the following steps: Acquire coronary artery CCTA images, and annotate the complete coronary artery mask and CTO plaque mask of the coronary artery CCTA images to obtain annotated images. Construct a residual network model to segment the annotated images to obtain the segmentation results of healthy coronary arteries. Boundary point detection is performed on the coronary artery CCTA images to obtain coronary artery breakpoints. The coronary artery breakpoints are compared with the CTO plaque mask annotations in the labeled images to confirm the CTO plaque breakpoints. Methods for obtaining CTO patch breakpoints through comparison include: By comparing the breakpoint with the labeled CTO mask through boundary point detection, the image slice at the breakpoint is searched pixel by pixel to see if there is any overlap between the image slice and the CTO mask. If there is no overlap, it is the endpoint of the normal coronary artery structure. If there is an overlap, it is the breakpoint caused by CTO plaque. The endpoint of the normal coronary artery structure is removed to obtain the CTO plaque breakpoint. Based on the tubular features of the coronary arteries, the coronary centerline of the CTO site is extracted from the CTO plaque breakpoint segment. Based on the coronary centerline of the CTO site, CTO plaque segmentation is performed to obtain the CTO plaque segmentation result. The method for extracting the coronary artery centerline of the CTO plaque breakpoint segment includes: Based on the corresponding masks of coronary arteries and CTO plaques, the skeletonization algorithm is used to automatically determine the position of the coronary artery center point at both ends of the rupture site. The local geometric features of the coronary arteries are corrected for the center line. The coronary artery center point between the coronary artery center points at both ends of the rupture site is determined by the center likelihood function. All coronary artery center points are connected together to obtain the coronary artery center line. The CTO plaque segmentation results and the healthy coronary artery segmentation results are stitched together according to the image data hierarchy to obtain a complete coronary tree result.
2. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 1, characterized in that, The method for obtaining the segmentation results of healthy coronary arteries includes: Doctors annotated the coronary artery CCTA images with complete coronary artery masks and CTO plaque masks, and then divided the annotated coronary artery CCTA images into training set, validation set and test set; The annotated coronary CCTA images are subjected to data augmentation and resampling processes, including rotation, flipping, and cropping. The initial residual network model is constructed based on a three-dimensional convolutional neural network. The construction method includes: introducing residual blocks into the convolutional blocks of the encoder and decoder. The residual blocks include 3D convolution, BN layers and linear correction units with skip connections. The initial residual network model is trained using the training set and validation set, with mean squared error as the loss function. The trained residual network model is then tested using the test set to obtain the residual network model. The residual network model is then used to segment the data-augmented coronary artery CCTA image to obtain the segmentation result of the healthy coronary arteries.
3. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 1, characterized in that, The expression for the central likelihood function is: In the formula, r1 li and r2 li denoted as ray, representing the ray from the edge of the coronary artery to any point within the cross-section of the coronary artery, where n represents the total number of rays and CL represents the likelihood value.
4. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 1, characterized in that, The method for obtaining CTO patch segmentation results includes: Based on the coronary artery centerline at the CTO site, CCTA image data slices and CTO plaque mask data corresponding to the CTO plaque were collected to obtain an image dataset; The image dataset is preprocessed to obtain a preprocessed dataset; An initial segmentation model is constructed, and the patch segmentation model is trained using the preprocessed dataset to obtain the CTO patch segmentation model; CTO patch segmentation is performed using the CTO patch segmentation model to obtain CTO patch segmentation results.
5. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 4, characterized in that, The preprocessing method includes: randomly dividing the acquired coronary CCTA images into training set, validation set and test set; performing data augmentation operations, including random rotation, flipping and cropping, on the divided coronary CCTA images; and uniformizing the size of the data augmentation images by resampling to obtain a preprocessed dataset.
6. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 4, characterized in that, The method for constructing the initial segmentation model includes: Construct an encoder network, using DenseNet and dilated convolutions instead of ordinary CNN convolutional blocks; A perception enhancement module, including a multi-scale feature extraction module and a channel attention module, is added to the end of the encoder network; By adding a skip connection between the encoder and decoder, an initial segmentation model is obtained.
7. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 6, characterized in that, The expression for the multi-scale feature extraction module is: in This represents a 1×1 convolution. Represents the feature fusion function. Indicates the shrinkage rate 3×3 dilated convolution, This indicates input.
8. The method for combined segmentation of CTO lesions and coronary artery tree based on CCTA according to claim 4, characterized in that, The loss function expression for the CTO patch segmentation model is as follows: In the formula, For the first The actual label corresponding to each pixel. For the first The prediction result for each pixel. This represents the total number of sample pixels.