Intracranial artery segmentation method and device based on wavelet enhanced space-time connectivity retention network

Abstract

The invention discloses an intracranial artery segmentation method and device based on a wavelet enhanced space-time connectivity retention network, which are used for intracranial artery segmentation in a digital subtraction angiography (DSA) image, and comprises the following steps: constructing a wavelet enhanced space-time connectivity retention network WESCP-Net based on U-Net, the wavelet enhanced space-time connectivity retention network comprising an encoder and a decoder; the encoder comprises a plurality of stages, the input of the first stage is an original DSA image, and the input of the other stages is the output of the previous stage; each stage of the encoder comprises a down-sampling module and a sequence feature extraction module SFEM which are connected with each other; the decoder comprises a plurality of stages corresponding to the encoder, and each stage comprises an up-sampling module; training the wavelet enhanced space-time connectivity retention network to obtain a trained wavelet enhanced space-time connectivity retention network; and segmenting the DSA image to be processed by using the trained wavelet enhanced space-time connectivity retention network. The WESCP-Net provided by the invention can realize accurate blood vessel segmentation in a dynamic DSA sequence.

Description

Technical Field

[0001] This invention relates to the field of graphics processing technology, specifically to a method and apparatus for intracranial artery segmentation based on wavelet-enhanced spatiotemporal connectivity-preserving networks. Background Technology

[0002] Currently, methods for segmenting cerebral blood vessels can be divided into traditional methods and deep learning-based methods.

[0003] Traditional methods mainly rely on image processing and computer vision techniques, and usually include four stages: preprocessing, feature extraction, segmentation and postprocessing. The preprocessing stage mainly removes noise and smooths the image, while the feature extraction stage focuses on extracting features related to vascular structures from the image. The segmentation stage aims to distinguish vascular structures from surrounding tissues, and the postprocessing stage further optimizes the segmentation results. Commonly used techniques include: (1) Threshold segmentation method: The optimal threshold is determined by accurately calculating the normal distribution of foreground pixels and the Gaussian distribution in MRA images, thereby accurately segmenting cerebral blood vessels. This method uses the principle of energy minimization and optimization algorithms to automatically detect and delineate the boundaries of target objects to complete the segmentation task. Segmentation method based on vascular morphology features: Segmentation is achieved by extracting the morphological features of blood vessels. One feasible method proposes a differential structure descriptor that describes each pixel in detail. Using a specific descriptor function, the model can determine whether a pixel belongs to a blood vessel, thereby achieving blood vessel segmentation. Another method innovatively proposes a blood vessel enhancement method based on the scale space theory framework. This method combines the advantages of smoothing blood vessel filters and nonlinear anisotropic diffusion, and achieves significant improvements in displaying blood vessels of different radii and enhancing the appearance of blood vessels.

[0004] Deep learning has made significant breakthroughs in computer vision, particularly in medical image segmentation, where deep convolutional neural networks (CNNs) are gradually replacing traditional image processing methods. The success of AlexNet demonstrated the powerful feature extraction capabilities of CNNs. By leveraging the computational advantages of convolutional layers and GPUs, it efficiently processes and extracts image features, significantly improving segmentation accuracy. Its performance has not only propelled the rapid development of computer vision but also brought new possibilities to medical image segmentation tasks. With the maturation of deep learning technology, researchers have proposed more efficient models for medical image segmentation tasks, especially for cerebral vascular image segmentation. The most representative example is UNet. The UNet network adopts a U-shaped structure, divided into an encoder, a decoder, and skip connections. The encoder gradually extracts high-level features from the image, while the decoder uses these features to recover the spatial resolution of the image. Skip connections can directly pass low-level features to the decoder, thereby improving segmentation accuracy, making it particularly suitable for medical image segmentation tasks with small sample datasets. Building upon UNet, researchers have further optimized the network. Res-UNet, by introducing a weighted attention mechanism and residual connections, effectively improves the model's performance in low-contrast vascular images and addresses the problems of image noise and blurred boundaries. On the other hand, UNet++ further improved the UNet structure by introducing nested and dense skip connections, enhancing the model's ability to capture image details, especially in high-resolution and semantically rich feature maps. Furthermore, with the increasing demand for 3D image segmentation, 3D U-Net emerged. To address the problem of 3D brain tumor image segmentation, 3D U-Net introduced depthwise separable convolutions on top of the traditional UNet, reducing the number of training parameters, and designed extended dense residual modules to expand the receptive field and capture a wider range of feature information, thereby improving the ability to handle complex structures. Through these continuously optimized deep learning models, the accuracy and efficiency of medical image segmentation, especially cerebrovascular image segmentation, have been significantly improved, further driving the development of related fields.

[0005] Despite some progress in existing brain vessel segmentation methods, key challenges remain, such as maintaining the connectivity of vascular topology and capturing dynamic spatiotemporal features. Summary of the Invention

[0006] The purpose of this application is to propose an intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network to address the aforementioned technical problems. By introducing wavelet-enhanced spatiotemporal feature extraction and topology preservation mechanism, the accuracy and stability of intracranial artery segmentation in digital subtraction angiography (DSA) images are significantly improved.

[0007] On the one hand, an intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving networks is used for intracranial artery segmentation in digital subtraction angiography (DSA) images, including:

[0008] S101, Wavelet-enhanced Spatiotemporal Connectivity Preservation Network Construction Steps: Construct a wavelet-enhanced spatiotemporal connectivity preservation network (WESCP-Net) based on U-Net. The wavelet-enhanced spatiotemporal connectivity preservation network includes an encoder and a decoder. The encoder includes several stages. The input of the first stage is the original DSA image, and the input of other stages is the output of the previous stage. Each stage of the encoder includes a downsampling module and a sequence feature extraction module (SFEM) connected together. The output of the downsampling module is used as the input of the sequence feature extraction module (SFEM). The sequence feature extraction module (SFEM) extracts multi-scale feature maps from the image, which includes a 2D image sequence and a time dimension. The decoder includes several stages corresponding to the encoder. Each stage includes an upsampling module. The sequence feature extraction module (SFEM) and the upsampling module of each stage are residually connected.

[0009] S102, Wavelet-enhanced spatiotemporal connectivity preservation network training steps: Train the wavelet-enhanced spatiotemporal connectivity preservation network to obtain the trained wavelet-enhanced spatiotemporal connectivity preservation network;

[0010] S103, DSA image segmentation step, uses a trained wavelet-enhanced spatiotemporal connectivity preservation network to segment the DSA image to be processed.

[0011] Preferably, the representation of the multi-scale feature map extracted by the encoder at each stage is as follows:

[0012]

[0013] Among them, F j represents the fused multi-scale feature map; j represents the downsampling ratio; N represents the sequence length; W represents a 1×1 convolution kernel; This represents the input sequence of the i-th frame; This represents the input feature map of the i-th frame; * indicates a convolution operation. This indicates a splicing operation.

[0014] Preferably, the downsampling module adopts the Haar wavelet downsampling module (HWD); by performing wavelet transform on the input image, the image is decomposed into four components, the spatial resolution of each component is reduced to half that of the original image, and the number of channels is increased to four times that of the input image.

[0015] Preferably, the upsampling module employs an efficient upsampling convolutional block (EUCB). The efficient upsampling convolutional block (EUCB) includes an upsampling operation Up, a depthwise convolution (DWC), a batch normalization (BN), an activation function ReLU, and a convolution Conv connected in sequence. The upsampling operation Up magnifies the feature map by two times, the depthwise convolution (DWC) refines the upsampled feature map, the batch normalization (BN) and ReLU activation functions normalize the feature map and introduce non-linearity, and the 1×1 convolution Conv adjusts the number of channels to ensure compatibility with the feature map in the next stage.

[0016] The preferred efficient upsampling convolutional block (EUCB) is represented as follows:

[0017] EUCB(x)=C 1×1 (ReLU(BN(DWC(Up(x))))

[0018] Among them, C 1×1 (·) represents a 1×1 convolution Conv; ReLU(·) represents an activation function; BN(·) represents batch normalization; DWC(·) represents depthwise convolution; Up(·) represents upsampling; x represents the input feature map.

[0019] Preferably, the downsampling module, the sequence feature extraction module SFEM, and the convolutions used in the downsampling module all use the WTConv module.

[0020] On the other hand, an intracranial artery segmentation device based on a wavelet-enhanced spatiotemporal connectivity-preserving network is used for intracranial artery segmentation in digital subtraction angiography (DSA) images, comprising:

[0021] A wavelet-enhanced spatiotemporal connectivity preservation network construction module is used to construct a wavelet-enhanced spatiotemporal connectivity preservation network (WESCP-Net) based on U-Net. The wavelet-enhanced spatiotemporal connectivity preservation network includes an encoder and a decoder. The encoder comprises several stages, with the input of the first stage being the original DSA image, and the input of other stages being the output of the previous stage. Each stage of the encoder includes a connected downsampling module and a sequence feature extraction (SFEM) module, and the output of the downsampling module serves as the input of the SFEM module. The SFEM module extracts multi-scale feature maps from images including 2D image sequences and the time dimension. The decoder comprises several stages corresponding to the encoder, with each stage including an upsampling module. The SFEM module and the upsampling module in each stage are residually connected.

[0022] The wavelet-enhanced spatiotemporal connectivity preservation network training module is used to train the wavelet-enhanced spatiotemporal connectivity preservation network to obtain the trained wavelet-enhanced spatiotemporal connectivity preservation network.

[0023] The DSA image segmentation module is used to segment the DSA image to be processed using a trained wavelet-enhanced spatiotemporal connectivity preservation network.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) The wavelet-enhanced spatiotemporal connectivity-preserving network WESCP-Net of the present invention uses multi-scale dynamic perception to enhance 3D DSA sequences (combining 2D image sequences and time dimension). WESCP-Net uses wavelet transform for feature enhancement and adopts spatiotemporal feature fusion technology to maintain the topological connectivity of blood vessels while efficiently mapping 3D data to 2D. It significantly improves the segmentation continuity of blood vessel ends in complex lesions such as arteriovenous malformations and achieves accurate blood vessel segmentation in dynamic DSA sequences.

[0026] (2) The sequence feature extraction module SFEM of the present invention can extract dynamic vascular features from complex spatiotemporal data. Unlike traditional methods that mainly focus on static spatial structure, the SFEM module integrates time series information and uses multi-scale feature maps to capture the morphological features of blood vessels at different time points, which significantly improves the accuracy of intracranial artery (IA) segmentation.

[0027] (3) The EUCB of the present invention maintains high segmentation performance while significantly improving computational efficiency. When processing DSA images with complex spatiotemporal characteristics, the EUCB can effectively restore details and provide more accurate feature support for blood vessel segmentation tasks.

[0028] (4) When processing DSA time-series images, HWD can significantly improve the accuracy of intracranial vessel segmentation, especially when processing small vessels or vessels with blurred background and foreground. In addition, HWD can better capture the dynamic changes of vessels in the time dimension, thus providing more comprehensive and accurate feature support for cerebral vessel segmentation tasks.

[0029] (5) The WTConv module of the present invention can significantly expand the receptive field of convolution without significantly increasing the number of parameters, enabling it to better capture low-frequency information, which is beneficial to the recognition of vascular structures in brain blood vessel segmentation tasks. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1This is a flowchart illustrating an embodiment of the intracranial artery segmentation method based on a wavelet-enhanced spatiotemporal connectivity-preserving network according to this application.

[0032] Figure 2 A schematic diagram of the structure of the wavelet-enhanced spatiotemporal connectivity-preserving network WESCP-Net, which is an embodiment of this application;

[0033] Figure 3 This is a schematic diagram of the structure of the Haar wavelet downsampling module (HWD) according to an embodiment of this application;

[0034] Figure 4 This is a decomposed schematic diagram of the Haar wavelet downsampling module (HWD) according to an embodiment of this application;

[0035] Figure 5 This is a schematic diagram of the structure of the efficient upsampling convolutional block EUCB in an embodiment of this application;

[0036] Figure 6 This is a schematic diagram of the processing of the WTConv module in an embodiment of this application;

[0037] Figure 7 A schematic diagram illustrating the segmentation details of the same case using different networks;

[0038] Figure 8 This is a schematic diagram of an intracranial artery segmentation device based on a wavelet-enhanced spatiotemporal connectivity-preserving network, which is an embodiment of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0040] like Figure 1 As shown in the figure, this embodiment provides an intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network for segmenting intracranial arteries in digital subtraction angiography (DSA) images, including the following steps.

[0041] S101, Wavelet-enhanced Spatiotemporal Connectivity Preservation Network Construction Steps: Construct a wavelet-enhanced spatiotemporal connectivity preservation network (WESCP-Net) based on U-Net. The wavelet-enhanced spatiotemporal connectivity preservation network includes an encoder and a decoder. The encoder includes several stages; the input of the first stage is the original DSA image, and the input of other stages is the output of the previous stage. Each stage of the encoder includes a connected downsampling module and a sequence feature extraction module (SFEM), and the output of the downsampling module serves as the input of the SFEM. The SFEM extracts multi-scale feature maps from the image, which includes a 2D image sequence and a time dimension. The decoder includes several stages corresponding to the encoder, and each stage includes an upsampling module. The SFEM and the upsampling module of each stage are residually connected.

[0042] Traditional DSA segmentation methods primarily focus on processing single-frame images, which only partially displays vascular contrast and ignores the dynamic changes of vessels over time. This limitation leads to incomplete representation of vascular structures, especially in complex lesions (such as arteriovenous malformations or small vessel diseases), making it difficult to accurately capture the spatiotemporal features of vessels. To overcome this problem, this invention proposes a network architecture based on an improved VSS-Net, namely the Wavelet Enhanced Spatiotemporal Connectivity Preservation Network (WESCP-Net). This network treats intracranial artery (IA) segmentation as a dimensionality reduction task, reducing 3D data (2D image sequences + time dimension) to 2D, thereby more effectively processing and analyzing the spatiotemporal features of vascular lesions. See also Figure 2 The diagram shows the architecture of WESCP-Net.

[0043] In the architecture of WESCP-Net, the Sequence Feature Extraction (SFEM) module is a core component, extracting dynamic vascular features from complex spatiotemporal data. Unlike traditional methods that primarily focus on static spatial structure, SFEM integrates time-series information, enabling it to capture subtle changes in blood vessels during dynamic processes, thus achieving more accurate identification and segmentation of vascular lesions. Specifically, SFEM utilizes multi-scale feature maps to capture the morphological features of blood vessels at different time points. This multi-dimensional feature extraction method significantly improves the accuracy of intracranial artery (IA) segmentation. In WESCP-Net, the overall structure follows a U-Net-style "symmetric encoder-decoder" framework, with the encoder representing a vertical contraction path on the left and the decoder representing a vertical expansion path on the right. SFEM is cleverly embedded in each encoder stage; in the encoder stage, an SFEM is inserted immediately after downsampling. SFEM takes a 2D temporal feature block at the current scale as input, while simultaneously integrating the feature stream passed down from the previous stage. Its main task is to extract multi-scale feature maps from images containing 2D image sequences and a temporal dimension. By integrating feature maps passed down from previous stages, this module generates a 2D feature representation that combines spatial and temporal information. As the encoder progresses, the resolution of the input data gradually decreases, while the receptive field expands accordingly. This dynamic process enables the network to generate higher-dimensional spatiotemporal feature representations that capture not only local details but also global contextual information. This provides richer and more comprehensive semantic information for subsequent segmentation tasks, ultimately significantly improving the accuracy and robustness of segmentation.

[0044] At each stage (each resolution layer in the encoder path), the input feature map is integrated with the original sequence to supplement fine-grained semantic information. The sequence is resized using bilinear interpolation to match the dimension of the corresponding feature map. Then, the image in the sequence is fused with the corresponding feature map using a 1×1 convolution with shared weights (WSC). The calculation process for the fused multi-scale feature map at scale j is as follows:

[0045]

[0046] Among them, F j represents the fused multi-scale feature map; j represents the downsampling ratio; N represents the sequence length; W represents a 1×1 convolution kernel; This represents the input sequence of the i-th frame; This represents the input feature map of the i-th frame; * indicates a convolution operation. This indicates a splicing operation.

[0047] The feature maps output by the encoder are spatially downsampled using stride convolution to provide input for the next stage of the encoder. These extracted feature maps are highly generalizable and compatible, seamlessly integrating with any encoder-decoder-based 2D segmentation model, thus providing robust feature support for subsequent vessel segmentation tasks. To further improve segmentation accuracy, an innovative feature fusion mechanism is introduced at each stage of the encoder: the extracted feature maps are deeply fused with the original sequence images to supplement finer-grained semantic information (except in the initial stage). Figure 2 The blue horizontal arrows and WSC indicate the operation of the fusion mechanism. The core of this design lies in utilizing the detailed texture and structural information in the original image to further optimize the expressive power of the feature map, enabling it to more accurately capture subtle vascular features. Simultaneously, global max pooling is performed along the sequence dimension of the feature map to reduce dimensionality. The generated multi-scale feature map is compatible with any U-shaped segmentation network. When combined with FR-UNet, the multi-scale output of SFEM can be input into each stage of FR-UNet, thereby improving the accuracy of vessel segmentation. The symmetrical, efficient upsampled voxel stacks on the right constitute the decoder; they are directly connected to the encoder stages in the same layer via dashed blue arrows, representing residual connections—the process of fusing the multi-scale output with the decoder.

[0048] In this embodiment, the downsampling module employs the Haar wavelet downsampling module (HWD). In DSA image processing, while traditional downsampling methods (such as max pooling or stride convolution) can effectively reduce computational overhead and expand the receptive field, they often lead to the loss of important spatial information such as boundaries, textures, and details. These problems are particularly prominent in semantic segmentation tasks, as the lost information is crucial for pixel-level prediction. To address this challenge, the downsampling in this invention employs the Haar wavelet downsampling module (HWD), which reduces the spatial resolution of the feature map while preserving as much original image information as possible, thus solving the information loss problem caused by traditional downsampling methods to a certain extent.

[0049] H0(z1), H1(z1), H0(z2), and H1(z2)

[0050] See Figure 3The diagram illustrates the specific structure of the HWD module. In the diagram, "H0(·)" and "H1(·)" both refer to the z-domain transfer function of the Haar analysis filter; the difference lies in their passband type (low-pass / high-pass), while "z1" and "z2" within parentheses distinguish whether the filter operates along the first dimension (row, x-direction) or the second dimension (column, y-direction) of the two-dimensional signal. The core of the HWD module is the lossless feature coding block, which utilizes Haar wavelet transform to reduce the spatial resolution of the feature map while preserving all information. Haar wavelet transform is a compact and orthogonal wavelet transform widely used in image coding, edge extraction, and other fields. Its main advantage lies in its ability to decompose image signals into low-frequency and high-frequency components while maintaining information integrity. For details, see [link to documentation]. Figure 4 As shown, the first-order Haar wavelet transform decomposes an image into four components: low frequency (A) and high frequency (H, V, D). The spatial resolution of each component is reduced to half that of the original image, while the number of channels is increased to four times that of the original image. This transform not only reduces the spatial resolution of the feature map but also prevents information loss by encoding some spatial information into the channel dimension.

[0051] By performing a wavelet transform on the signal, the image is decomposed into four components. The spatial resolution of each component is reduced to half that of the original image, while the number of channels is increased to four times that of the original image. This transform not only reduces the spatial resolution of the feature map but also prevents information loss by encoding some spatial information into the channel dimension.

[0052] Subtle changes and dynamic features of vascular structures are crucial for accurate identification and segmentation of vascular lesions. Compared to traditional stride convolution, the HWD module demonstrates superior computational efficiency and retains more information compared to average pooling. It can directly replace existing downsampling methods without significant modifications to the network architecture, exhibiting excellent versatility and scalability. When processing DSA temporal images, using the HWD module can significantly improve the accuracy of intracranial vessel segmentation, especially when dealing with small vessels or vessels blurred between the background and foreground. Furthermore, the HWD module is better able to capture the dynamic changes of vessels over time, thus providing more comprehensive and accurate feature support for cerebral vessel segmentation tasks.

[0053] In this embodiment, the upsampling module employs an efficient upsampling convolutional block (EUCB). In DSA image segmentation tasks, the high resolution and complex spatiotemporal characteristics of image sequences pose challenges to traditional upsampling methods, making it difficult to effectively recover details and thus limiting segmentation accuracy. To better adapt to these characteristics and improve segmentation performance, an efficient upsampling convolutional block (EUCB) is introduced. The core objective of the efficient upsampling convolutional block (EUCB) is to optimize the upsampling process, progressively adjusting the size and resolution of the feature maps to ensure precise alignment with the skip connection feature maps in the next stage, thereby providing higher-quality input for subsequent feature fusion.

[0054] The operational flow of the efficient upsampling convolutional block (EUCB) is as follows: First, the upsampling operation Up(·) magnifies the feature map by a factor of two to recover some lost spatial details. Next, a depthwise convolution (DWC(·)) is applied to refine the upsampled feature map. Depthwise convolution significantly reduces computational complexity by performing convolutions per channel while preserving rich semantic information. Subsequently, batch normalization (BN(·)) and the ReLU activation function are applied to normalize the feature map and introduce non-linearity, thereby stabilizing the training process and enhancing feature representation. Finally, a 1×1 convolution C1×1 is used to adjust the number of channels to ensure compatibility with the feature map in the next stage. Figure 5 The structure diagram of the efficient upsampling convolutional block EUCB is shown.

[0055] Compared to traditional 3×3 convolution methods, the depthwise convolution technique employed in the Efficient Upsampling Convolutional Block (EUCB) significantly reduces computational overhead while maintaining excellent segmentation performance. This optimized design not only improves the model's computational efficiency but also enhances its ability to recover details, thereby achieving higher accuracy and efficiency in DSA image segmentation tasks. The functionality of the Efficient Upsampling Convolutional Block (EUCB) can be expressed by the following mathematical formula:

[0056] EUCB(x)=C 1×1 (ReLU(BN(DWC(Up(x)))))

[0057] Among them, C 1×1 (·) represents a 1×1 convolution Conv; ReLU(·) represents an activation function; BN(·) represents batch normalization; DWC(·) represents depthwise convolution; Up(·) represents upsampling; x represents the input feature map.

[0058] By replacing traditional convolution operations with depthwise convolution, the efficient upsampling convolutional block (EUCB) significantly improves computational efficiency while maintaining high segmentation performance. This design not only optimizes the utilization of computational resources but also enhances the model's adaptability to high-resolution image segmentation tasks. Especially when processing DSA images with complex spatiotemporal characteristics, the efficient upsampling convolutional block EUCB can effectively recover details, providing more accurate feature support for blood vessel segmentation tasks.

[0059] In this embodiment, the downsampling module, the sequence feature extraction module SFEM, and the convolutions used in the downsampling module all use the WTConv module (WT convolution module).

[0060] In cerebral vessel segmentation, the temporal characteristics of DSA images cause vascular structures to dynamically change at different time points. Due to the locality of traditional convolution operations, they struggle to capture this global spatiotemporal information. To better handle this complexity, a method is needed that can expand the receptive field without significantly increasing the number of parameters. The WTConv module achieves this by utilizing the multi-scale properties of wavelet transform. Wavelet transform can decompose an image into components of different frequencies while preserving spatial information at each scale. This allows convolution operations to be performed over a larger area, thus more effectively capturing low-frequency information, which is particularly important for cerebral vessel segmentation. The three-step operation of WTConv is presented in the figure as a complete vertical chain of "downward capture arrow → SFEM → upward capture arrow". It combines a large receptive field, low-frequency global semantics, and high-frequency edge details while keeping the number of parameters controllable, thereby significantly improving the accuracy of cerebral vessel segmentation. The workflow of the WTConv module can be represented by the following steps: (1) Wavelet transform: decompose the input image into multiple frequency components; (2) Small convolution kernel convolution: apply a small convolution kernel to each frequency component; (3) Inverse wavelet transform: recombine the convolved components into the output image.

[0061] The core of the WTConv module is wavelet transform, which expands the receptive field of convolution by decomposing the input image into low-frequency and high-frequency components. Specifically, the WTConv module first performs a wavelet transform on the input image, decomposing it into multiple frequency components. Then, a small convolution kernel is applied to each frequency component. Finally, an inverse wavelet transform is used to recombine these convolutional components into the output image. This process not only expands the receptive field of convolution but also preserves important low-frequency information through multi-scale processing.

[0062] See Figure 6The diagram illustrates the complete WTConv workflow. Using this method, the WTConv module can significantly expand the receptive field of convolution without significantly increasing the number of parameters, enabling it to better capture low-frequency information. This is highly advantageous for vascular structure recognition in brain vessel segmentation tasks.

[0063] S102, Wavelet-enhanced spatiotemporal connectivity preservation network training steps: Train the wavelet-enhanced spatiotemporal connectivity preservation network to obtain the trained wavelet-enhanced spatiotemporal connectivity preservation network.

[0064] The wavelet-enhanced spatiotemporal connectivity preservation network can be trained using any existing method, which will not be specifically described in this embodiment.

[0065] S103, DSA image segmentation step, uses a trained wavelet-enhanced spatiotemporal connectivity preservation network to segment the DSA image to be processed.

[0066] The following experiments validate the intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity preservation networks proposed in this embodiment, showcasing experimental results to verify the performance of the proposed WESCP-Net. First, the experimental setup, including implementation details and evaluation metrics, is introduced. Then, the performance of WESCP-Net is compared with other methods, and an ablation study is conducted to analyze the contribution of each component.

[0067] (1) Experimental setup

[0068] The method proposed in this invention was implemented using Python 3.7, PyTorch 1.12.1, and Ubuntu 20.04.6LTS. All experiments were conducted on a machine equipped with an RTX 3090 (24GB VRAM) GPU. The initial learning rate was set to 5×10(-4), and the learning rate was adjusted using a cosine annealing algorithm. The DIAS dataset was divided into 30 training samples, 10 validation samples, and 20 test samples according to patient level. The DSA sequences were resampled to length 8 before being input into the model. Training was performed for a total of 200 epochs, with 100 iterations per epoch to ensure fairness in comparisons.

[0069] To evaluate the segmentation performance of tubular and linear structures, a novel connectivity preservation metric, clDice, was employed. This metric calculates segmentation performance based on the intersection of the skeleton and the mask.

[0070] The skeleton SP and SL are extracted from the predicted segmentation mask (VP) and the ground truth mask (VL), respectively. Then, the proportion of SP within VL is calculated, called the topology accuracy (Tprec(SP,VL)), and the topology sensitivity (Tsens(SL,VP)) is calculated. clDice is defined as the harmonic mean of these two metrics. The specific formulas are as follows:

[0071]

[0072] (2) Performance Comparison

[0073] The following table compares the performance of different methods across multiple evaluation metrics. Table 1 shows the performance of several network models, including VSS-Net, on a range of metrics.

[0074] Table 1

[0075]

[0076] WESCP-Net demonstrates a significant advantage over 2D methods. Regarding the Dice similarity coefficient (DSC), all 2D methods typically score lower than the method presented in this invention. For example, UNet++ has a DSC of 0.7569, Res-UNet has 0.7494, while WESCP-Net achieves a DSC of 0.7812, surpassing these methods. Similarly, in terms of accuracy (Acc), 2D methods such as Res-UNet and Att-UNet have Accuities of 0.9609 and 0.9622, respectively, while WESCP-Net's Accuity is 0.9668, showing a significant improvement. The Intersection over Union (IOU), an important indicator of segmentation performance, also favors WESCP-Net, with an IOU of 0.6420, significantly higher than UNet++'s 0.6120 and Res-UNet's 0.6030.

[0077] In 3D methods, WESCP-Net also performs exceptionally well. Compared to common 3D methods (such as UNet, Res-UNet, and FR-UNet), WESCP-Net demonstrates significant advantages across multiple metrics. In DSC, UNet scores 0.7450, while WESCP-Net achieves 0.7812, a significant difference. FR-UNet's Acc score is 0.9648, while WESCP-Net's is 0.9668, showing even better performance. WESCP-Net also outperforms many traditional 3D methods in Area Under the Curve (AUC) and Class Dice coefficient (clDice). For example, FR-UNet's AUC is 0.9816, while WESCP-Net's AUC is 0.9842. Similarly, FR-UNet's clDice is 0.6904, while WESCP-Net's clDice is 0.7075. WESCP-Net also shows some improvements in comparisons with 3D-to-2D methods. 3D-to-2D methods (such as SVS-Net and PSC) performed relatively weakly on several metrics. For example, SVS-Net's DSC was 0.7449 and Acc was 0.9629, while WESCP-Net's DSC was 0.7812 and Acc was 0.9668, significantly outperforming these methods. Figure 7 The image shows the segmentation details of the same case using different networks.

[0078] Overall, WESCP-Net performs exceptionally well across 2D, 3D, and 3D-to-2D methods, particularly outperforming all comparable methods in key metrics such as DSC, Acc, and IOU. Compared to 2D and 3D methods, WESCP-Net demonstrates significant improvements in segmentation accuracy and stability, especially in IOU and AUC. Compared to 3D-to-2D methods, WESCP-Net also surpasses these traditional methods, exhibiting stronger segmentation capabilities and stability.

[0079] Compared to the VSS-Net model, WESCP-Net shows a significant improvement in the DSC (Dice similarity coefficient) metric. VSS-Net's DSC is 0.7805, while WESCP-Net reaches 0.7812, indicating a slight advantage in segmentation performance. In terms of Accuracy, WESCP-Net also slightly outperforms VSS-Net, scoring 0.9668, while VSS-Net scores 0.9667. Regarding IOU (Intersection over Union), WESCP-Net leads with an IOU of 0.6420, while VSS-Net's IOU is 0.6411, reflecting WESCP-Net's advantage in segmentation accuracy. In AUC (Area Under the Curve), VSS-Net slightly surpasses VSS-Net with a score of 0.9849, while WESCP-Net scores 0.9842. WESCP-Net also outperforms VSS-Net in the clDice metric. The clDice of VSS-Net is 0.7073, while that of WESCP-Net is 0.7075, indicating that WESCP-Net maintains better vessel connectivity and generates more accurate topology in the segmentation results.

[0080] (3) Ablation test

[0081] To verify the effectiveness of different components in WESCP-Net, this invention designed ablation experiments. The impact on model performance was observed by removing or replacing some key modules. Table 2 shows the experimental results under several different configurations, including: the complete WESCP-Net model, WESCP-Net with EUCB and HWD modules retained but WTConv module removed, WESCP-Net containing only the WTConv module, and WESCP-Net containing only the EUCB module. By comparing the experimental results of each version on multiple metrics, the contribution of different modules to performance was further verified.

[0082] Table 2

[0083]

[0084] In WESCP-Net, the model still performs well on most metrics. DSC slightly decreases from 0.7812 to 0.7806, while Acc increases from 0.9668 to 0.9670. Although these differences are small, the slight decrease in DSC indicates that the WT convolutional module helps improve the model's segmentation accuracy. clDice decreases significantly from 0.7075 to 0.7038. Notably, AUC (0.9846) is similar to the original model, indicating that removing these two modules has little impact on the model's overall discriminative ability. Therefore, it can be inferred that the EUCB and HWD modules mainly play a role in refining segmentation boundaries and enhancing sensitivity to local features, with a relatively small impact on global classification ability.

[0085] In WESCP-Net with only the WTConv module retained, model performance significantly decreased. DSC dropped from 0.7812 to 0.7791, and Acc slightly decreased from 0.9668 to 0.9672. The change in AUC (from 0.9842 to 0.9845) was more pronounced. clDice also decreased significantly, from 0.7075 to 0.6988. This indicates that the WTConv module plays a crucial role in capturing details, and removing it leads to a decrease in segmentation accuracy, especially affecting the model's robustness when handling complex cases.

[0086] For WESCP-Net containing only the EUCB module, DSC and Acc were observed to be 0.7804 and 0.9673, respectively, a slight decrease compared to the original model's 0.7812 and 0.9668. Compared to the impact of removing the WTConv module, the effect of the EUCB module appears to be more concentrated on segmentation accuracy metrics such as DSC and clDice, with its removal leading to a significant decrease in segmentation performance. This suggests that the primary function of the EUCB module is to enhance the fine-grained recognition of local structures, particularly for identifying small or difficult-to-distinguish arterial or vascular regions.

[0087] Ablation studies show that the EUCB, WTConv, and HWD modules in WESCP-Net exhibit significant synergistic effects. When using the WTConv or EUCB modules alone, the clDice metric shows a significant decrease due to differences in feature maps. Only when the EUCB, WTConv, and HWD modules are applied to the network simultaneously can the stability of the clDice metric and the improvement in segmentation accuracy be effectively ensured. This further validates the complementarity of the various modules in WESCP-Net.

[0088] The effectiveness and superiority of WESCP-Net were verified through comparative and ablation experiments as described above. In the comparative experiments, researchers conducted detailed comparisons with multiple benchmark methods, demonstrating WESCP-Net's advantages in several key metrics. The ablation experiments explored the contributions of different modules in depth, particularly highlighting the importance of the clDice metric in evaluating model performance. The following is a summary and discussion of the experimental results.

[0089] Compared to existing methods such as VSS-Net, WESCP-Net demonstrates superior performance across multiple key metrics, including DSC, Acc, IOU, AUC, and clDice. Particularly noteworthy is the significant improvement in clDice, reflecting its outstanding capabilities in fine-grained segmentation tasks. clDice, as a metric for segmentation quality, is particularly important in medical image segmentation because it comprehensively evaluates the model's accuracy in detailed regions. In practical applications, especially for the precise segmentation of lesion regions, an improved clDice metric indicates that the model can better capture small but clinically significant lesion areas, thus providing stronger support for medical diagnosis.

[0090] Among all 2D methods, WESCP-Net achieved the best performance. While 2D methods excel in simpler tasks, they often struggle with complex medical image segmentation tasks due to a lack of contextual information about the spatial structure of the image. On the other hand, 3D methods significantly improve segmentation accuracy by incorporating spatial structure information. FR-UNet is a strong 3D benchmark method, performing well on multiple metrics, with a DSC of 0.7648, an AUC of 0.9816, and a clDice score of 0.6904. However, despite FR-UNet's excellent performance in 3D segmentation, VSS-Net and WESCP-Net maintain their leading positions, particularly in DSC and AUC metrics. VSS-Net and WESCP-Net achieve DSCs of 0.7805 and 0.7812, respectively, demonstrating their advantage in handling complex detail segmentation tasks. For 3D-to-2D segmentation methods, WESCP-Net also performs exceptionally well. It continuously improves segmentation accuracy in 3D-to-2D medical image segmentation tasks, particularly excelling in detail processing and complex structure recognition.

[0091] Ablation studies further revealed the specific roles of each module in WESCP-Net. In the ablation experiments, researchers removed the EUCB, HWD, and WT modules respectively and observed the impact of these changes on model performance. The results showed that the clDice value decreased significantly after removing some modules. This indicates that clDice not only measures the model's accuracy at the level of detail but also provides sufficient feedback on whether the model can comprehensively identify the target region.

[0092] In practical medical applications, particularly in vascular segmentation tasks in neuroimaging, accurately capturing each blood vessel and lesion area is crucial. Models that improve clDice values ​​demonstrate enhanced capabilities in detecting subtle changes in images and identifying potential lesion areas. This ability is of great benefit in assisting physicians in diagnosis and providing precise treatment.

[0093] This invention presents an intracranial artery segmentation method based on a wavelet-enhanced spatiotemporal connectivity preservation network, which significantly improves the segmentation accuracy of intracranial vessels in DSA images. This breakthrough is achieved by integrating a sequence feature extraction module, an efficient upsampling convolutional block, a Haar wavelet downsampling module, and a wavelet transform convolutional module. Experiments on the DIAS dataset demonstrate that WESCP-Net outperforms existing 2D, 3D, and 3D-to-2D methods on several key metrics, particularly excelling in fine-grained segmentation tasks. Furthermore, ablation studies confirm the effectiveness of each module and their synergistic contribution to performance improvement. Future research could explore integrating more advanced feature extraction techniques to incorporate richer global contextual information, thereby further enhancing the model's ability to segment complex vascular structures.

[0094] like Figure 8 As shown, this embodiment also discloses an intracranial artery segmentation device based on a wavelet-enhanced spatiotemporal connectivity-preserving network, comprising:

[0095] The wavelet-enhanced spatiotemporal connectivity preservation network construction module 801 is used to construct a wavelet-enhanced spatiotemporal connectivity preservation network WESCP-Net based on U-Net. The wavelet-enhanced spatiotemporal connectivity preservation network includes an encoder and a decoder. The encoder includes several stages, with the input of the first stage being the original DSA image, and the input of other stages being the output of the previous stage. Each stage of the encoder includes a downsampling module and a sequence feature extraction module (SFEM) connected together, and the output of the downsampling module is used as the input of the sequence feature extraction module (SFEM). The sequence feature extraction module (SFEM) extracts multi-scale feature maps from the image, which includes a 2D image sequence and a time dimension. The decoder includes several stages corresponding to the encoder, and each stage includes an upsampling module. The sequence feature extraction module (SFEM) of each stage is residually connected to the upsampling module.

[0096] The wavelet-enhanced spatiotemporal connectivity preservation network training module 802 is used to train the wavelet-enhanced spatiotemporal connectivity preservation network to obtain the trained wavelet-enhanced spatiotemporal connectivity preservation network.

[0097] DSA image segmentation module 803 is used to segment the DSA image to be processed using a trained wavelet-enhanced spatiotemporal connectivity preservation network.

[0098] The specific implementation of each module of the intracranial artery segmentation device based on wavelet-enhanced spatiotemporal connectivity preservation network is the same as the intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity preservation network, and will not be described again in this embodiment.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intracranial artery segmentation based on wavelet-enhanced spatiotemporal connectivity-preserving networks, used for intracranial artery segmentation in digital subtraction angiography (DSA) images, characterized in that, include: S101, Wavelet-enhanced Spatiotemporal Connectivity Preservation Network Construction Steps: Construct a wavelet-enhanced spatiotemporal connectivity preservation network (WESCP-Net) based on U-Net. The wavelet-enhanced spatiotemporal connectivity preservation network includes an encoder and a decoder. The encoder includes several stages. The input of the first stage is the original DSA image, and the input of other stages is the output of the previous stage. Each stage of the encoder includes a downsampling module and a sequence feature extraction module (SFEM) connected together. The output of the downsampling module is used as the input of the sequence feature extraction module (SFEM). The sequence feature extraction module (SFEM) extracts multi-scale feature maps from images including 2D image sequences and time dimensions. The decoder includes several stages corresponding to the encoder. Each stage includes an upsampling module. The sequence feature extraction module (SFEM) and the upsampling module of each stage are residually connected. S102, Wavelet-enhanced spatiotemporal connectivity preservation network training steps: Train the wavelet-enhanced spatiotemporal connectivity preservation network to obtain the trained wavelet-enhanced spatiotemporal connectivity preservation network; S103, DSA image segmentation step, uses a trained wavelet-enhanced spatiotemporal connectivity preservation network to segment the DSA image to be processed.

2. The intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network according to claim 1, characterized in that, The representation of the multi-scale feature maps extracted at each stage of the encoder is as follows: Among them, F j represents the fused multi-scale feature map; j represents the downsampling ratio; N represents the sequence length; W represents a 1×1 convolution kernel; This represents the input sequence of the i-th frame; This represents the input feature map of the i-th frame; * indicates a convolution operation; ⊕ indicates a concatenation operation.

3. The intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network according to claim 1, characterized in that, The downsampling module uses the Haar wavelet downsampling module (HWD). By performing wavelet transform on the input image, the image is decomposed into four components, the spatial resolution of each component is reduced to half that of the original image, and the number of channels is increased to four times that of the input image.

4. The intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network according to claim 1, characterized in that, The upsampling module employs an efficient upsampling convolutional block (EUCB); the efficient upsampling convolutional block (EUCB) comprises, in sequence, an upsampling operation Up, a depthwise convolution (DWC), a batch normalization (BN), an activation function ReLU, and a convolution Conv. The upsampling operation Up enlarges the feature map by a factor of two. The depthwise convolution DWC refines the upsampled feature map. Batch normalization (BN) and ReLU activation functions normalize the feature map and introduce non-linearity. The 1×1 convolution Conv adjusts the number of channels to ensure compatibility with the feature map in the next stage.

5. The intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network according to claim 4, characterized in that, The efficient upsampling convolutional block (EUCB) is represented as follows: EUCB(x)=C 1×1 (ReLU(BN(DWC(Up(x))))) Among them, C 1×1 (·) represents a 1×1 convolution Conv; ReLU(·) represents an activation function; BN(·) represents batch normalization; DWC(·) represents depthwise convolution; Up(·) represents upsampling; x represents the input feature map.

6. The intracranial artery segmentation method based on wavelet-enhanced spatiotemporal connectivity-preserving network according to claim 1, characterized in that, The downsampling module, the sequence feature extraction module SFEM, and the convolutions used in the downsampling module all use the WTConv module.

7. An intracranial artery segmentation device based on wavelet-enhanced spatiotemporal connectivity-preserving network, used for intracranial artery segmentation in digital subtraction angiography (DSA) images, characterized in that, include: A wavelet-enhanced spatiotemporal connectivity preservation network construction module is used to construct a wavelet-enhanced spatiotemporal connectivity preservation network (WESCP-Net) based on U-Net. The wavelet-enhanced spatiotemporal connectivity preservation network includes an encoder and a decoder. The encoder comprises several stages, with the input of the first stage being the original DSA image, and the input of other stages being the output of the previous stage. Each stage of the encoder includes a connected downsampling module and a sequence feature extraction (SFEM) module, and the output of the downsampling module serves as the input of the SFEM module. The SFEM module extracts multi-scale feature maps from images including 2D image sequences and the time dimension. The decoder comprises several stages corresponding to the encoder, with each stage including an upsampling module. The SFEM module and the upsampling module in each stage are residually connected. The wavelet-enhanced spatiotemporal connectivity preservation network training module is used to train the wavelet-enhanced spatiotemporal connectivity preservation network to obtain the trained wavelet-enhanced spatiotemporal connectivity preservation network. The DSA image segmentation module is used to segment the DSA image to be processed using a trained wavelet-enhanced spatiotemporal connectivity preservation network.

Images