A three-dimensional coronary artery segmentation method based on multi-scale and global features

Abstract

The application discloses a three-dimensional coronary artery segmentation method based on multi-scale and global features, and aims at the problems of easy missed detection of distal fine branches, poor blood vessel tree connectivity and high false positive in a three-dimensional coronary artery CT angiography image, and constructs a convolutional neural network MG-Net of a three-dimensional encoding-decoding structure. A local scale adaptation module is introduced at the encoding end, adaptive enhancement of different vessel feature is realized through multi-branch local convolution, a global context integration module is introduced at the network bottleneck layer, a low-overhead axial serialization global interaction mode is adopted to establish a long-distance voxel dependency relationship, so that the consistency of blood vessel running is enhanced, and a semantic gating fusion module is introduced at the decoding end fusion stage, selective screening is conducted on the skip connection features, so that noise propagation is inhibited. The network is trained to learn a real segmentation mask image, and the trained network is used to output a three-dimensional coronary artery blood vessel segmentation result.

Description

Technical Field

[0001] This invention pertains to the field of image processing technology, specifically to computer-aided medical image segmentation, and more specifically to a three-dimensional coronary artery segmentation method based on multi-scale and global features. Background Technology

[0002] Coronary CT angiography (CCTA) is an important non-invasive imaging technique for screening and assessing coronary artery disease. It is widely used in clinical settings such as coronary artery stenosis, plaque burden, and revascularization decision support. It can accurately segment the coronary arteries in three-dimensional images and support subsequent processes such as extraction of vessel centerline, quantification of lumen diameter, calculation of stenosis rate, preoperative planning, and follow-up comparison.

[0003] Currently, the encoder-decoder architecture based on 3D U-Net is the mainstream method for medical image segmentation, which can fuse shallow details and deep semantics through skip connections. However, because this stitching process lacks a selective filtering mechanism for skip connection information, noise, artifacts, and false vascular textures at the encoder end are easily transmitted to the decoder end along with high-resolution features, leading to increased background missegmentation and false positives. On the other hand, pure convolutional operations mainly rely on local receptive fields and lack the ability to model cross-scale and long-range consistency, easily resulting in the breakage of distal fine branches and discontinuity of the vascular tree.

[0004] Therefore, it is necessary to propose a three-dimensional coronary artery segmentation method that can enhance the adaptation of fine branch scales, improve long-range connectivity consistency, and suppress noise propagation during the fusion stage, so as to meet the dual requirements of accuracy and efficiency for clinical auxiliary diagnosis. Summary of the Invention

[0005] To address key issues in 3D coronary artery CCTA segmentation, such as missed detection of distal fine branches, broken vascular trees, high false positive rates, and large global modeling overhead, this invention provides a 3D coronary artery segmentation method based on multi-scale and global features. This method, built upon a 3D encoder-decoder convolutional neural network, incorporates a local scale adaptation module at the encoder end to enhance the local scale representation of small vessels. At the bottleneck layer, an axially sequential global interaction approach is used to establish long-distance voxel dependencies to enhance course consistency. Finally, a semantically gated fusion module is introduced at the decoding end to selectively introduce skip connection features, outputting the coronary artery segmentation results.

[0006] A three-dimensional coronary artery segmentation method based on multi-scale and global features includes the following steps: Step 1: Use the original 3D coronary CCTA images as samples and the labeled real segmentation mask images as labels to construct a training set.

[0007] Step 2: Construct a coronary artery segmentation network MG-Net with multi-scale global consistency, using the samples obtained in Step 1 as input. The encoding and decoding operations are performed to output the foreground probability map P for each voxel. A binary segmentation mask M is obtained by thresholding. When P≥τ, the corresponding voxel is marked as the coronary artery; otherwise, it is marked as the background. τ is a preset threshold.

[0008] The coronary artery segmentation network is a three-dimensional encoder-decoder structure, which includes an encoder, a bottleneck layer, and a decoder.

[0009] The encoder is used to progressively reduce the spatial resolution of the input image and extract multi-scale semantic features. It comprises L encoding layers consisting of an initial 3D convolutional layer, a downsampling layer, and residual convolutional blocks. Through progressive downsampling, the receptive field is expanded and multi-scale semantic features are extracted. Simultaneously, features at corresponding scales are output for skip connections. A local scale adaptation module is introduced into the preset high-resolution encoding layer to perform local scale adaptive enhancement and background noise reduction on the high-resolution features, thereby improving the detection capability of distant fine branches and reducing noise-induced false responses. Where L≥2, i represents the encoding level index, F i S represents the original feature output by the i-th layer encoder. i This represents the feature output of the i-th coding layer used for skip connections. This represents the preset high-resolution encoding layer index set, Conv3d() represents a 3D convolutional layer, Down() represents a downsampling layer, ResBlock() represents a residual convolutional block, and LSA() represents a local scale adaptation module.

[0010] The local scale adaptation module obtains branch features of different receptive fields through multi-branch local convolution and outputs weighted fused features. : Among them, Conv3d (j) Let represent the j-th 3D convolution, and k represent the number of 3D convolutions with different kernel sizes or different dilation rates, used to cover different blood vessel diameters and local structural scales. Generate a mapping for the weights, used to extract features from the branch. Generate normalized weights In its specific implementation, the local scale adaptation module consists of a three-dimensional convolution operator, a normalization operator, and an activation function. Its input and output are multi-dimensional tensor data stored in the computer, and it performs convolution and element-wise operations to complete feature fusion.

[0011] The bottleneck layer introduces a global context integration module on the lowest resolution features output by the encoder. It establishes long-distance voxel dependencies through low-overhead serialized global interactions and superimposes the global enhancement features onto the bottleneck input features using residual addition. This achieves global consistency enhancement of the coronary vascular tree and improves connectivity between weak-contrast regions and bifurcation regions, reducing vascular rupture issues. Where B represents the bottleneck layer output feature, Norm() represents the normalization layer, and Linear() represents the linear mapping. AxisMix() represents the serialization and fusion operator for global interaction, which uses axial serialization for global interaction, performs one-dimensional gated mixing or gated recursive updates along the spatial dimension, and fuses the multi-axis outputs to obtain globally consistent enhanced features. The gated recursive update method is as follows: Where, x t Represents axial sequence The t-th element of the expanded sequence, h represents the length of the axial sequence. t Indicates the recursive state, y t Indicates recursive output. , , Represents a linear mapping or a convolutional mapping. This represents the Sigmoid function. This indicates element-wise multiplication.

[0012] The decoder performs upsampling step-by-step through multiple decoding layers to restore spatial resolution, and then uses the decoded features upsampled from each decoding layer. Jump connection features at the corresponding scale Fusion is performed. To suppress high-resolution noise and pseudo-vascular textures introduced by skip connections, a semantically gated fusion module is introduced at the fusion point, which separately processes the decoded features. Features of skip connections Convolutional mapping operations are performed, and gating weights are generated based on the mapping results. Skip connection features are then filtered element-wise before participating in feature fusion. This suppresses the propagation of irrelevant details such as veins, myocardial texture, and artifacts through skip connections, reducing false positives and improving the continuity of bifurcation and weak contrast regions. Among them, U i For the i-th level decoding feature, P represents the skip connection features filtered by the semantically gated fusion module, and P is the decoder output probability map. For upsampling, For fusion operators. G represents the feature concatenation operation. i This indicates the gating weight.

[0013] Step 3: Compare the segmentation mask M obtained in Step 2 with the real segmentation mask image in Step 1, calculate the loss function, and update the parameters of the coronary artery segmentation network MG-Net through backpropagation until the training converges and the final model is obtained.

[0014] Step 4: Input the 3D coronary artery CCTA image to be segmented into the coronary artery segmentation network MG-Net trained in Step 3, and output the coronary artery segmentation results.

[0015] The present invention has the following beneficial effects: By introducing three key units into the 3D encoding-decoding backbone, the coronary artery cross-scale and connectivity problems are synergistically optimized. Specifically, a local scale adaptation module is set at the encoding end to achieve local scale adaptive modeling, enhance the feature expression of coronary arteries of different diameters, and improve the recall rate of fine branches. A global context integration module is set at the bottleneck layer, and a low-overhead global interaction method is used to establish long-distance dependencies, thereby enhancing global consistency, improving the continuity of the coronary artery tree, and reducing breaks. A semantic gating fusion module is set at the fusion position at the decoding end to gating and filtering skip connection features guided by decoding semantics, suppressing noise propagation caused by skip connections, reducing false positives, and improving the reconstruction quality of boundary and bifurcation regions. At the same time, the network as a whole still maintains the engineering feasibility of the structure, avoiding the high memory and computational overhead caused by 3D global self-attention. Attached Figure Description

[0016] Figure 1 This is a flowchart of a three-dimensional coronary artery segmentation method based on multi-scale and global features.

[0017] Figure 2This is a schematic diagram of the coronary artery segmentation network MG-Net structure in the embodiment.

[0018] Figure 3 This is a schematic diagram of the local scale adaptation module structure in the embodiment.

[0019] Figure 4 This is a schematic diagram of the global context integration module structure in the embodiment.

[0020] Figure 5 This is a schematic diagram of the semantic gating fusion module in the embodiment. Detailed Implementation

[0021] The present invention will be further explained below with reference to the accompanying drawings; This embodiment is used to explain the principle and implementation of the present invention and does not constitute a limitation on the scope of protection of the present invention; without departing from the concept of the present invention, those skilled in the art can make equivalent substitutions for the number of network layers, number of channels, convolution kernel parameters, training strategies, etc.

[0022] A three-dimensional coronary artery segmentation method based on multi-scale and global features, such as Figure 1 As shown, the specific steps include: Step 1: In this embodiment, the ImageCAS dataset for 3D coronary CCTA is used as the original data source. Each original 3D coronary CCTA volume data is preprocessed and used as a sample, with the corresponding real segmentation mask image as the label. The dataset is randomly divided into training set, validation set, and test set. s1.1 Read the single-instance 3D coronary CCTA volume data and its corresponding ground truth segmentation mask, and simultaneously read the spatial information, including voxel spacing, orientation information, or affine transformation information. The ground truth segmentation mask is a binary mask or a multi-class mask, with the coronary artery region as the foreground and the rest as the background.

[0023] s1.2. Perform intensity window width clipping and normalization on the volume data. In this embodiment, if the volume data is HU value volume data, then it is processed according to the preset window width range. Clipping is performed, and linear normalization is used to map the intensity to... This reduces intensity differences caused by different scanning parameters. For other types of volume data, z-score normalization or other monotonic normalization methods can be used.

[0024] s1.3 Resample the normalized volume data to the target voxel spacing. The data is processed in mm format, and linear interpolation is used for the volume data, while nearest neighbor interpolation is used for the actual segmentation mask to ensure that the labels do not alias. The specific target voxel spacing can be adjusted according to the device protocol and memory constraints.

[0025] s1.4. Locate non-air regions using threshold rules and calculate their 3D bounding boxes. After expanding the boundaries of the bounding boxes, clip them to obtain the clipping volume data. With clipping mask Alternatively, coarse localization of the cardiac ROI, aortic localization, or fixed ROI trimming strategies can be used.

[0026] s1.5. To adapt to GPU memory and improve training efficiency, channel dimensions were added to the pruned volume data, and a sliding window approach was used to... medium cutting size is The three-dimensional patch is used as the training sample.

[0027] s1.6 For each patch generated by the sliding window, record its coordinates in the original volume data. It also includes cropping origin and resampling information to support the patch prediction results being back-fused during the inference stage to obtain the whole instance segmentation result.

[0028] s1.7 To alleviate class imbalance caused by sparse foreground elements, patches are classified as positive or negative during the training phase. If a foreground voxel exists in the mask block corresponding to a patch, it is classified as a positive patch; otherwise, it is classified as a negative patch. The proportion of positive patches is set between 0.3 and 0.7. When there are insufficient positive patches, they are resampled to make up for the deficiency, or hard sample mining or priority sampling based on centerline / fine branch regions is used.

[0029] s1.8 Divide the processed data into training set, validation set and test set according to a preset ratio; during the training phase, data augmentation operations such as random flipping, random rotation / scaling, noise perturbation, and Gamma perturbation can be performed on the input patch; random augmentation is not performed during the validation and testing phases.

[0030] Step 2, construct a... Figure 2 The MG-Net network shown takes the training set samples X obtained in step 1 as input into the network. After encoding, bottleneck global consistency enhancement, and decoding reconstruction operations, it outputs the predicted coronary artery segmentation probability map P, and then thresholds it to obtain the segmentation mask M. The specific steps are as follows: s2.1, The encoder performs an initial three-dimensional convolution on the input X to obtain Subsequently, features are extracted through stepwise downsampling and residual convolutional blocks using four coding layers: Where L=4, representing the number of encoding levels in the encoder. F irepresents the original encoded features output by the i-th encoding layer, Conv3d() represents the three-dimensional convolutional mapping, Down() represents the downsampling operation, and ResBlock() represents the residual convolutional block.

[0031] Introduced on the first and second coding layers Figure 3 The Local Scale Adaptation (LSA) module shown enhances the adaptation capability of small blood vessels to local scales and outputs features for skip connections. : in, ={1,2} represents the preset high-resolution coding level index set. The Local Scale Adaptation (LSA) module obtains the original coding features F. i Branching features of different receptive fields were weighted and fused: Among them, Conv3d (j) For the j-th 3D convolution, each 3D convolution Conv3d (j) Different kernel sizes or different dilation rates can be used. Generate mappings for the weights. This represents the normalized weight.

[0032] s2.2, The bottleneck layer's lowest resolution feature at encoder output. The above introduction is as follows Figure 4 The global context integration module shown establishes dependencies between long-distance voxels through global interactions of axial serialization, and outputs bottleneck features by residual summation. : Wherein, Norm() represents a normalization layer, Linear() represents a linear mapping, and AxisMix() represents an operator that performs serialized global interaction and fusion along at least one spatial dimension. The AxisMix uses axial serialization for global interaction, preferably performing one-dimensional gated recursive updates along the depth, height, or width axis, and fusing the multi-axis outputs to obtain globally consistent enhanced features. As a specific implementation, the above one-dimensional gated feature update operation can be implemented through a feature processing module based on intermediate feature state updates, for example, using a sequence feature hybrid structure that includes state updates and gate control mechanisms. In a preferred embodiment, the sequence feature hybrid structure can be implemented using a Mamba structure.

[0033] The decoder shown in s2.3 uses the bottleneck output as the initial decoding feature, then recovers the spatial resolution by upsampling through four decoding layers, and fuses it with the skip connection features of the corresponding scale. Simultaneously, at each fusion stage, features such as... Figure 5 The semantically gated fusion module SGF shown here applies skip connection features. Gating and screening are performed before fusion: Among them, U i For the i-th level decoding feature, P represents the skip connection features filtered by the semantically gated fusion module, and P is the decoder output probability map. For upsampling, For fusion operators, G represents the feature concatenation operation. i This indicates the gating weight.

[0034] s2.4 For the foreground probability map of each voxel output by the decoder Set threshold ,when When the corresponding voxel is in time, it is labeled as a coronary artery; otherwise, it is labeled as background. A binary segmentation mask is obtained through thresholding. .

[0035] Step 3: Apply the binary segmentation mask obtained in Step 2. The model is compared with the real mask, the loss is calculated and the parameters are updated until the training converges to obtain the final model: s3.1 Input the training set samples into the coronary artery segmentation network MG-Net constructed in step 2 to obtain a binary segmentation mask. , and the corresponding real mask Calculate the loss function And update the network parameters through backpropagation: in, Represents the Dice class loss. Represents the cross-entropy loss class. , This is a weighting coefficient, which can be set based on the performance of the validation set.

[0036] An adaptive optimizer is used to update the parameters, and the optimal model parameters and threshold are selected based on the validation set metrics. .

[0037] After multiple rounds of training, the performance of the coronary artery segmentation network MG-Net trained by our method and the existing 3D segmentation network architecture U-Net3D were tested using test set data. The average Dice score of our method is approximately 79.07, which is a significant improvement compared to the average Dice score of approximately 73.13 obtained by U-Net3D.

[0038] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A three-dimensional coronary artery segmentation method based on multi-scale and global features, which utilizes a trained coronary artery segmentation network to segment the coronary artery region from a three-dimensional coronary CT angiography image, characterized in that: The coronary artery segmentation network includes an encoder, a bottleneck layer, and a decoder. It outputs a predicted coronary artery segmentation probability map by encoding and decoding the input raw three-dimensional coronary CT angiography image. The encoder downsamples the input samples step by step through multiple coding layers to extract multi-scale semantic features, and introduces a local scale adaptation module in the preset high-resolution coding layer to enhance the expression of local scale features of coronary arteries with different diameters in the skip connection features. The bottleneck layer introduces a global context integration module on the lowest resolution features output by the encoder. It establishes dependencies between distant voxels through a global interaction method of axial serialization to enhance the consistency of coronary artery course. The decoder upsamples the bottleneck layer output features step by step through multiple decoding layers. At the position where the decoded features are fused with the skip connection features of the corresponding scale, a semantic gating fusion module is introduced to selectively introduce the skip connection features to obtain the predicted coronary artery segmentation probability map.

2. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: The original 3D coronary CT angiography images were intensity-cropped and normalized, and the images and corresponding segmentation masks were spatially resampled to unify the voxel spacing, which served as training data for the coronary artery segmentation network.

3. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: The local scale adaptation module is set in at least one high-resolution coding layer in the encoder, and its output is passed to the decoder as a skip connection feature.

4. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: The local scale adaptation module first uses three-dimensional convolution with different kernel sizes or different dilation rates to obtain local features of different receptive fields of the original coding features output by the coding layer. Then, it generates normalized weights for each local feature and finally performs weighted fusion of the features of each branch.

5. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: The global context integration module performs axial serialization expansion of features along at least one of the spatial dimensions of depth, height, or width, and performs one-dimensional feature mixing or gated recursive update on the expanded feature sequence to establish dependencies between distant voxels.

6. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 5, characterized in that: The one-dimensional feature mixing or gated recursive update consists of a linear mapping layer, a one-dimensional convolutional layer, and a gate function, and outputs globally consistent enhanced features by adding residuals.

7. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: The semantic gating fusion module performs convolutional mapping on the decoded features output by the decoding layer and the skip connection features of the corresponding layer, generates gating weights based on the mapping results, filters the skip connection features element by element, and then fuses them with the decoded features as the input of the next decoding layer.

8. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: By thresholding the foreground probability map output by the coronary artery segmentation network using a preset threshold τ, a binary segmentation mask is obtained.

9. The three-dimensional coronary artery segmentation method based on multi-scale and global features as described in claim 1, characterized in that: The binary segmentation mask obtained from the foreground probability map output by the coronary artery segmentation network is compared with the segmentation mask of the original 3D coronary CT angiography image, and the loss function is calculated. The training process is completed by updating the parameters of the coronary artery segmentation network through backpropagation. in, Represents the Dice class loss. Represents the cross-entropy loss. , These are the weighting coefficients.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1 to 9.

Images