Intracranial aneurysm CTA image segmentation method and device and storage medium

Abstract

The invention discloses an intracranial aneurysm CTA image segmentation method and device and a storage medium. The method comprises the following steps: introducing a multi-path convolution module and a multi-view pooling Transform module in a specific downsampling stage in an encoder path to construct a basic image segmentation model; training the basic image segmentation model by using the preprocessed CTA image data of the intracranial aneurysm patient to obtain a CTA image segmentation model; and based on the model, inputting preprocessed 3D CTA volume data, and outputting a corresponding 3D intracranial aneurysm segmentation binary mask. According to the method, the MPConv module and the MVPFormer module can be efficiently integrated, so that the MPConv module and the MVPFormer module work cooperatively, the segmentation precision of the intracranial aneurysm is remarkably improved on the premise that the calculation burden is not excessively increased, the problems that an existing method is not sensitive to segmentation of the small-size intracranial aneurysm and missing detection is prone to occurring are solved, and the goodness of fit between the segmentation boundary and the real situation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a three-dimensional medical image segmentation method based on deep learning, in particular to a method suitable for accurate and automatic segmentation of intracranial aneurysms in CT angiography (CTA) images. BACKGROUND

[0002] Deep learning techniques, especially models based on convolutional neural networks (CNN) and Transformers, have become the mainstream method for medical image segmentation. For example, models such as 3D U-Net and V-Net can effectively process 3D data, but pure CNN models have bottlenecks in modeling long-distance dependencies. Although Transformers can model global context, direct application to high-resolution 3D medical images will result in excessively high computational complexity.

[0003] To this end, existing technologies have proposed hybrid architectures such as TransUNet and UNETR, trying to combine the local feature extraction capabilities of CNN and the global modeling capabilities of Transformer. In particular, the LW-CTrans model proposes two efficient modules: the multi-path convolution module (MPConv) and the multi-view pooling Transformer module (MVPFormer), which are proven in their respective original papers to effectively improve segmentation performance.

[0004] Defects and deficiencies of existing technologies: 1. Although existing hybrid architectures (such as Cascaded Hybrids) have improved performance, they often come with a significant increase in the number of parameters, low computational efficiency, and difficulty in rapid deployment in clinical practice.

[0005] 2. Existing technologies lack an optimal strategy to effectively integrate high-efficiency modules such as MPConv and MVPFormer into a unified network backbone to fully leverage their synergistic advantages in complex and small target segmentation tasks. Simply stacking or replacing modules may not achieve significant performance improvements and may even cause module conflicts, training instability, and other issues.

[0006] 3. For intracranial aneurysms with varying morphology, small size, and complex relationships with surrounding vascular structures, the accuracy of existing segmentation methods, particularly the recall rate for small aneurysms and the accuracy of segmentation boundaries (IoU), still needs to be further improved. SUMMARY

[0007] Invention purposes: The purpose of the present application is to provide an intracranial aneurysm CTA image segmentation method, which solves the problems of large parameter quantity, low calculation efficiency and low segmentation accuracy of smaller intracranial aneurysms of the existing hybrid architecture. Another purpose of the present application is to provide a computer device and a storage medium, which solves the problem of how to perform the intracranial aneurysm CTA image segmentation method.

[0008] Technical scheme: The intracranial aneurysm CTA image segmentation method provided by the present application comprises the following steps: Based on the encoder-decoder segmentation network comprising an encoder path, a decoder path and a jump connection, a multi-path convolution module and a multi-view pooling Transformer module are introduced in the encoder path to construct a basic image segmentation model; the encoder path in the model comprises the following 5 down-sampling stages: In the first stage, standard 3D convolution is used to extract the bottom features of the image; In the second stage, the multi-view pooling Transformer module is introduced to begin to establish the mid-range context dependence; In the third stage, the multi-path convolution module is configured to capture complex morphological features using its multi-scale receptive field; In the fourth stage, the multi-view pooling Transformer module is introduced again to deeply model the high-level semantic features and integrate the global context; In the fifth stage, the final feature compression and global information aggregation are performed in the deepest layer; The CTA image segmentation model is obtained by training the basic image segmentation model using the preprocessed CTA image data of the intracranial aneurysm patient; Based on the CTA image segmentation model, the preprocessed 3D CTA volume data is input, and the corresponding 3D intracranial aneurysm segmentation binary mask is output.

[0009] Preferably, the encoder-decoder segmentation network comprises any one of a 3D UX-Net network, a V-Net variant and an nnU-Net framework.

[0010] The present application proposes a new hybrid 3D segmentation network. The core is to take 3D UX-Net as the backbone network, and innovatively integrate MPConv module and MVPFormer module at different levels of the encoder path, to construct an efficient module integrated architecture.

[0011] Preferably, in the encoder-decoder segmentation network, the encoder is responsible for hierarchical feature extraction, and the outputs of each level are connected through jump connection and the corresponding decoder to fuse shallow details and deep semantics; the decoder restores the spatial resolution by transposed convolution step by step, and finally outputs the segmentation map through 1x1x1 convolution and Sigmoid activation function.

[0012] Preferably, the configuration multi-path convolution module includes four deep wise convolution branches and one point wise convolution branch, the input features are simultaneously sent to the four parallel deep wise convolution branches, and features are extracted from different dimensions respectively; the outputs of the four deep wise convolution branches are combined through splicing operation, and finally the feature fusion and channel number adjustment are performed by the point wise convolution branch to output the enhanced feature map.

[0013] Preferably, the multi-view pooling Transformer module includes multi-view pooling, layer normalization, multi-layer perceptron, splicing and three-linear upsampling, and the input 3D feature is projected into three 2D feature maps along the sagittal, axial and coronal three orthogonal directions through global average pooling; each 2D feature map is sequentially subjected to layer normalization and multi-layer perceptron with shared weights for nonlinear transformation; the three 2D features after transformation are spliced and reconstructed into 3D features with the same size as the input through three-linear upsampling.

[0014] Preferably, the training of the basic image segmentation model uses the AdamW optimizer, the initial learning rate is 0.0001, the cosine annealing scheduling is adopted, and the DiceCELoss is used as the loss function.

[0015] Further, gradient accumulation and gradient clipping are adopted in the training process to maintain training stability.

[0016] Preferably, the preprocessing includes uniformizing the format of the CTA image data of the intracranial aneurysm patient to NIfTI, intensity normalization and spatial resampling to isotropic resolution.

[0017] The second aspect of the present application discloses a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor realizes the above method when executing the computer program.

[0018] The third aspect of the present application discloses a computer storage medium, which stores instructions, and the instructions make the computer execute the above method when executed on the computer.

[0019] Advantages: Compared with the prior art, the present application has the following remarkable advantages: The present application can efficiently integrate MPConv and MVPFormer modules, so that they work cooperatively, significantly improve the segmentation accuracy of intracranial aneurysms without excessive increase of computing burden, solve the problem that the existing method is not sensitive to small size intracranial aneurysms and is prone to miss detection, and improve the coincidence degree of the segmentation boundary and the true situation. At the same time, the present application can achieve an excellent balance between the parameter quantity, inference speed and segmentation accuracy of the model to meet the actual application requirements of clinical automatic diagnosis. Attached Figure Description

[0020] Figure 1 This is a diagram of the overall architecture of the present invention; Figure 2 Here is a diagram of the MPConv module structure; Figure 3 Here is a diagram of the MVPFormer module structure; Figure 4 The results of CTA image segmentation for intracranial aneurysms using different models are shown. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0022] like Figure 1 As shown, a method for segmenting intracranial aneurysm CTA images includes the following steps: (1) Using 3D UX-Net as the backbone network, the MPConv module and MVPFormer module are innovatively and hierarchically integrated at different levels of its encoder path to construct a new hybrid 3D segmentation network, namely the basic image segmentation model. The specific process is as follows: The 3D UX-Net network is a U-Net-like encoder-decoder structure, including an encoder path, a decoder path, and skip connections. The encoder is responsible for hierarchical feature extraction, and its outputs at each level are concatenated with the corresponding decoder through skip connections to fuse shallow details and deep semantics. The decoder gradually upsamples through transposed convolutions to restore spatial resolution, and finally outputs a segmentation map through 1x1x1 convolutions and a sigmoid activation function.

[0023] The MPConv and MVPFormer modules are introduced in each stage of the encoder (Stages 1-4 and Bottleneck) to build a basic image segmentation model. The encoder path in the basic image segmentation model includes the following 5 downsampling stages: Stage 1: Use standard 3D convolution to extract low-level features of the image; Stage 2: Introduce the multi-view pooling Transformer module and begin establishing mid-range context dependencies; Stage 3: Configure a multi-path convolution module to capture complex morphological features using its multi-scale receptive field; Stage 4: The multi-view pooling Transformer module is introduced again to perform deep modeling of high-level semantic features and global context integration; Bottleneck: Performs final feature compression and global information aggregation at the deepest layer; As shown in Figure 2 , the MPConv module includes four deep wise convolution (DWConv) branches and a point wise convolution (PWConv, 1x1x1) branch, the input features are simultaneously sent to the four parallel deep wise convolution branches, and the features are extracted from different dimensions respectively; the outputs of the four deep wise convolution branches are merged through the concatenation (Concat) operation, and finally the feature fusion and channel number adjustment are performed by the point wise convolution branch, and the enhanced feature map is output.

[0024] The MPConv module adopts a multi-branch parallel architecture, uses anisotropic deep wise convolution kernels (3x3x3, 1x3x3, 3x1x3, 3x3x1), fully captures 3D local features from multiple spatial dimensions, strengthens the network's ability to capture complex local patterns in 3D space, and is particularly suitable for complex morphological changes of aneurysms.

[0025] As shown in Figure 3 , the MVPFormer module includes multi-view pooling (GAP), layer normalization (Norm), multi-layer perceptron (MLP), concatenation (Concat) and three-linear upsampling (Upsample), the input 3D feature is projected into three 2D feature maps along the sagittal, axial and coronal orthogonal directions through global average pooling; each 2D feature map is sequentially subjected to layer normalization and multi-layer perceptron with shared weights for nonlinear transformation; the three 2D features after transformation are concatenated and reconstructed into a 3D feature with the same size as the input through three-linear upsampling.

[0026] The MVPFormer module projects the 3D feature map along the sagittal, axial and coronal orthogonal views through global average pooling, decomposes the complex 3D global modeling problem into multiple 2D problem processing, efficiently models the global context dependency relationship with low computational cost, greatly reduces the computational complexity while effectively capturing the global context information.

[0027] (2) The CTA image segmentation model is obtained by training the basic image segmentation model using the preprocessed CTA image data of intracranial aneurysm patients, specifically: The 3D CTA image data of intracranial aneurysm patients is collected. The data is preprocessed, including format unification to NIfTI, intensity normalization (window width 400, window level 40), spatial resampling to isotropic resolution (0.5x0.5x0.5mm 3 ); Using the method of step (1), the LWUX-Net network is built using the PyTorch deep learning framework, and the total parameter quantity is about 53.7M; The number of encoder channels is configured as: 1→48→96→192→384→768, and the MPConv and MVPFormer modules are embedded at the specific stage according to the foregoing scheme; The data set is retrospectively collected from the imaging data of patients with intracranial aneurysms who underwent head and neck CTA examination in the hospital. After a series of screening, 500 cases are used for experiments in the embodiment. The initial file format of the data image is DICOM (Digital Imaging Communications in Medicine), and then the image is converted into nii.gz format for further experiments. The focus of the study is the CTA image with a thickness of 1.5 mm, the size of each image is 512*512 pixels, and the slice range is 117 to 175. According to the ratio of 8:1:1, the 500 data cases are randomly divided into a training set (400), a validation set (50) and a test set (50). The CTA image data of the intracranial aneurysm patient after preprocessing is used to train the basic image segmentation model, the AdamW optimizer is used for training, the initial learning rate is 0.0001, the cosine annealing scheduling is used, and the DiceCELoss combined with the Dice loss and the cross entropy loss is used as the loss function to handle the class imbalance. Gradient accumulation (Batch Size is set to 1 due to the memory limitation of 3D data), gradient clipping (max_norm=1.0) and other techniques are used in the training process to ensure the stability of the training.

[0028] (3) Based on the trained CTA image segmentation model, the preprocessed 3D CTA volume data is input, and the corresponding 3D intracranial aneurysm segmentation binary mask is output.

[0029] The image segmentation performance of the present application and a plurality of other encoder-decoder segmentation networks is compared, and the results are as follows: Table 1 Performance test results of different encoder-decoder segmentation networks

[0030] As can be seen from the results in Table 1, the CTA image segmentation model (LWUX-Net (our)) in the present application performs well in the intracranial aneurysm segmentation task. In the core evaluation index Dice, LWUX-Net (our) reaches 85.1%, which is significantly higher than 80.0% of SCUX-Net which performs best in the table, indicating that the CTA image segmentation model in the present application has higher segmentation accuracy between the intracranial aneurysm region and the background, and can more accurately capture the target region.

[0031] The accuracy of 81.4% of the LWUX-Net (our) is similar to that of the SCUX-Net of 81.2%, but combined with a high recall rate of 90.2%, it indicates that the LWUX-Net (our) effectively recalls the real intracranial aneurysm region while minimizing false positives and effectively avoiding missed detection.

[0032] Compared with 68.3% of the SCUX-Net, the under-the-line index of 74.4% of the LWUX-Net (our) is significantly improved, which indicates that the segmented region of the LWUX-Net has a higher overlap with the ground truth region, thereby obtaining better segmentation performance.

[0033] In terms of optimizers and loss functions, the LWUX-Net (our) adopts the AdamW optimizer and the Dice + CE loss function. This configuration helps the model converge more effectively in the process.

[0034] Changing the embedding stage and embedding number of the MPConv module and the MVPFormer module in the application or changing the down-sampling stage in the encoder, a plurality of similar experimental comparison models are obtained, the performance of these models and the application is compared, and the results are as follows: Table 2 Influence of embedding mode and down-sampling stage of MPConv module and MVPFormer module on model performance

[0035] As can be seen from the results in Table 2, the LW-UXNet with the complete [Conv, MVPFormer, MPConv, MVPFormer] configuration achieves a Dice coefficient of 85.1%, which is the best among all experimental configurations. This indicates that the synergistic work of each module maximizes the segmentation performance of the model.

[0036] After removing the MPConv, the Dice coefficient is significantly reduced to 63.4%. Although the number of parameters remains unchanged, the segmentation effect is obviously decreased, highlighting the important role of MP-Conv in feature extraction and segmentation tasks.

[0037] When Conv and MVPFormer are removed at the same time and only MPConv is retained, the Dice coefficient is only 48.6%, which is much lower than the complete model, indicating that Conv and MVPFormer are essential for improving segmentation accuracy and are important components of an effective segmentation model.

[0038] After exchanging the positions of MVPFormer and MPConv, the Dice coefficient is 59.4%, which is significantly lower than the complete model, reflecting that the order of the modules affects the performance of the model, and the original module order is more conducive to intracranial aneurysm segmentation.

[0039] After removing Conv, the Dice coefficient dropped to 54.7%, indicating that Conv plays an irreplaceable role in early feature extraction of the model and is important for the smooth progress of subsequent segmentation tasks.

[0040] The results of actual segmentation processing of intracranial aneurysm CTA images using different models are shown in Table 1. Figure 4 Figure 4 Among them, Axial represents the axial plane, GT represents the true segmentation result, Swin UNETR, 3D UX-Net, SCUX-Net represent the image segmentation results of the corresponding network architecture, and LWUX-Net represents the image segmentation result of the present application. Figure 4 As can be seen from Table 1, compared with other models, the LWUX-Net is closer to the true standard.

[0041] In summary, the present application has the following advantages: Synergistic performance improvement: Through the above hierarchical integration strategy, the MPConv and MVPFormer modules realize the synergistic perception from local details to global semantics. On a CTA dataset containing 500 patients, the Dice similarity coefficient of the present application reaches 85.10%, and the intersection over union (IoU) reaches 74.48%, which is significantly better than existing mainstream methods (such as SCUX-Net with 80.0% Dice).

[0042] High recall rate advantage: The recall rate of the present application is as high as 90.25%, which indicates that the model has a significant advantage in avoiding aneurysm detection and has a very high value for clinical preliminary screening.

[0043] Efficient calculation: Compared with pure Transformer models (such as UNETR), the present application reduces the parameter amount by about 42% while maintaining better performance, and the inference speed is improved by about 44%, which is more conducive to clinical deployment.

[0044] Friendly to small targets: The architecture also shows strong segmentation ability for small aneurysms with a diameter less than 3mm, effectively alleviating the class imbalance problem in medical image segmentation.

[0045] The core innovation of the present application lies in the hierarchical integration strategy of MPConv and MVPFormer in the 3D UX-Net skeleton. This strategy can also be applied to other similar encoder-decoder segmentation networks (such as V-Net variants, nnU-Net framework, etc.) for other complex small target 3D medical image segmentation tasks, such as lung nodule segmentation, liver tumor segmentation, etc.​

Claims

1. An intracranial aneurysm CTA image segmentation method, characterized in that, Comprising the following steps: Based on an encoder-decoder segmentation network comprising an encoder path, a decoder path and a skip connection, a multi-path convolution module and a multi-view pooling Transformer module are introduced in the encoder path to construct a basic image segmentation model; the encoder path in the model comprises the following 5 down-sampling stages: In the first stage, standard 3D convolution is used to extract the low-level features of the image; In the second stage, a multi-view pooling Transformer module is introduced to begin to establish the mid-range context dependence; In the third stage, a multi-path convolution module is configured to capture complex morphological features using its multi-scale receptive field; In the fourth stage, the multi-view pooling Transformer module is introduced again to deeply model the high-level semantic features and integrate the global context; In the fifth stage, the final feature compression and global information aggregation are performed at the deepest layer; The CTA image segmentation model is trained by using the preprocessed CTA image data of the intracranial aneurysm patient based on the basic image segmentation model. Based on the CTA image segmentation model, the preprocessed 3D CTA volume data is input to output the corresponding 3D intracranial aneurysm segmentation binary mask.

2. The method of claim 1, wherein the method further comprises: The encoder-decoder segmentation network includes any one of a 3D UX-Net network, a V-Net variant and an nnU-Net framework.

3. The method of claim 1, wherein the method further comprises: In the encoder-decoder segmentation network, the encoder is responsible for hierarchical feature extraction, and the outputs of each level are connected through a skip connection with the corresponding level of the decoder to fuse shallow details and deep semantics; the decoder gradually up-samples through transposed convolution to restore the spatial resolution, and finally outputs a segmentation map through 1x1x1 convolution and Sigmoid activation function.

4. The method of claim 1, wherein the method further comprises: The multi-path convolution module includes four deep wise convolution branches and one point wise convolution branch, the input features are simultaneously sent to the four parallel deep wise convolution branches to extract features from different dimensions respectively; the outputs of the four deep wise convolution branches are merged through a concatenation operation, and finally the point wise convolution branch is used for feature fusion and channel number adjustment to output enhanced feature maps.

5. The method of claim 1, wherein the method further comprises: The multi-view pooling Transformer module includes multi-view pooling, layer normalization, multi-layer perceptron, concatenation and tri-linear up-sampling, the input 3D features are projected into three 2D feature maps along the sagittal, axial and coronal orthogonal directions through global average pooling; each 2D feature map is sequentially subjected to layer normalization and multi-layer perceptron with shared weights for nonlinear transformation; the three 2D features after transformation are concatenated and reconstructed into 3D features with the same size as the input through tri-linear up-sampling.

6. The method of claim 1, wherein the method further comprises: The training of the basic image segmentation model uses AdamW optimizer with an initial learning rate of 0.0001, cosine annealing scheduling and DiceCELoss as the loss function.

7. The intracranial aneurysm CTA image segmentation method of claim 6, wherein, Gradient accumulation and gradient clipping are used during training to maintain training stability.

8. The method of claim 1, wherein the method further comprises: The preprocessing includes uniformizing the format of the CTA image data of the intracranial aneurysm patient to NIfTI, intensity normalization and spatial resampling to isotropic resolution. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor implements the method in any one of claims 1-8 when executing the computer program.

10. A computer storage medium, characterized in that, The computer storage medium has stored therein instructions which, when executed on a computer, cause the computer to perform the method in any one of claims 1-8.

Images