Liver anatomical structure segmentation method, device and minimally invasive surgery intelligent planning system
Patent Information
- Application Number
- CN202610785292.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-02
- Publication Date
- 2026-08-21
AI Technical Summary
即便现有分割模型已实现多维度特征优化，但其在肝脏复杂解剖结构的高精度、精细化分割任务中依旧存在诸多亟待解决的问题
本申请提出的频域注意力Transformer模型与现有的结合卷积网络与Transformer的分割模型不同，其一方面利用Swin-Transformer与Transformer级联的结构有效提取局部纹理以及全局语义特征，另一方面通过频域注意力增强特征中的纹理细节。比起现有的模型能够提取更准确的局部和全局特征。
Smart Images
Figure CN122618232A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of medical image processing technology, and specifically relates to methods, equipment and intelligent planning systems for liver anatomical structure segmentation and minimally invasive surgery. Background Technology
[0002] Existing methods for segmenting liver anatomy in preoperative CT / MR images can be broadly categorized into traditional segmentation methods and deep learning segmentation methods. Traditional Couinaud segmentation methods for the liver can be further divided into voxel-based and section-based approaches. Voxel-based segmentation methods, such as Yang X et al. (2018) in their paper "Segmentation of liver and vessels from CT images and classification of liver segments for preoperative liversurgical planning in living donor liver transplantation" published in *Computer Methods and Programs in Biomedicine*, assign liver segment labels by calculating the spatial distance between liver voxels and hepatic vein branches. However, this method does not strictly adhere to the Couinaud segmentation criterion. Section-based segmentation methods first extract key points, dividing lines, and other geometric anatomical elements from the liver surface morphology and vascular anatomy, and then construct anatomical sections based on these elements to reconstruct liver segment regions. A typical example is Alirr OI et al. (2020) in their paper "Automatic atlas-based liver segmental anatomy identification for hepaticsurgical" published in *International Journal of Computer Assisted Radiology and Surgery*. In the planning, the hepatic veins are processed into a skeleton, and functional liver segments are reconstructed based on the spatial orientation relationship between voxels and the vein skeleton. Combined with key point detection, standardized Couinaud liver segment division is achieved.
[0003] With the emergence of Convolutional Neural Networks (CNNs), deep learning-based medical image segmentation methods have gradually replaced traditional segmentation algorithms. Since Ronneberger O et al. (2015) published "U-net: Convolutional networks for biomedical image segmentation" at the International Conference on Medical Image Computing and Computer-assisted Intervention, proposing the U-shaped convolutional neural network (U-Net) for medical image segmentation, the encoder that progressively extracts and downsamples semantic features, the decoder that progressively upsamples, and the skip connection structure that can utilize multi-scale features to obtain segmentation details have become the mainstream style of medical image segmentation models. However, convolutional operations can only focus on local neighborhood information and have an inherent deficiency in the ability to model long-distance global dependencies.
[0004] In recent years, the Vision Transformer (ViT), proposed by Dosovitskiy A et al. (2020) in their preprint paper "An image is worth 16x16 words: Transformers for image recognition at scale" (arXiv:2010.11929), has been used to address the shortcomings of convolutional neural networks and has achieved good results in various vision tasks. ViT first segments the image into image patches and maps them to sequence tokens. It then models the global dependencies between tokens based on a multi-head self-attention mechanism to extract global deep features. In their paper "Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation" published at the Medical Imaging with Deep Learning conference in 2024, Rahman MM et al. (2024) further improved the accuracy of medical image segmentation by using the PyramidVision Transformer (PVT) for multi-scale global feature extraction. While Transformer has a powerful ability to capture global dependencies, its image block encoding method is prone to losing shallow, fine local details.
[0005] Given the respective advantages and disadvantages of CNN and ViT, a large number of hybrid architecture medical image segmentation models that integrate CNN and ViT have emerged at this stage, which can be divided into three research approaches. The first type uses CNN or window-type ViT (such as Swin-Transformer) to extract local features, then introduces ViT to mine global features, and finally achieves segmentation prediction by fusing the two features through a decoder. An example is the TransUNet model proposed by Chen J et al. (2024) in their paper "Transunet: Rethinking the u-net architecture design for medical imagesegmentation through the lens of transformers" published in Medical Image Analysis. The second type designs a convolutional and Transformer fusion interaction module, using the two structures to collaboratively and synchronously extract local detail features and global contextual features. An example is the CFATransUNet model proposed by Wang C et al. (2024) in their paper "Cfatransunet: Channel-wise cross fusion attention and transformer for 2d medical image segmentation" published in Computers in Biology and Medicine. The third type uses a CNN and ViT dual-branch parallel encoding structure to independently extract local and global features, and completes cross-branch feature interaction fusion during the encoding or decoding stage. An example is the model proposed by Li Z et al. (2024) in IEEE... The ScribFormer model was proposed in the paper "Scribformer: Transformer makes cnnwork better for scribble-based medical image segmentation" published in IEEE Transactions on Medical Imaging.
[0006] Building upon such CNN-ViT fusion models, some studies have further introduced frequency domain transformation strategies to enhance the model's ability to represent image texture and structural features from the frequency domain dimension. For example, Li M et al. (2025) proposed PFESA, a parameter-free edge and structure attention network based on Fast Fourier Transform, in their paper "Pfesa: Fft-based parameter-free edge and structure attention for medical image segmentation" published at the International Conference on Medical Image Computation and Computer-Aided Intervention. Even though existing segmentation models have achieved multi-dimensional feature optimization, many problems still need to be solved in the high-precision and refined segmentation of complex anatomical structures of the liver. Summary of the Invention
[0007] To address the aforementioned issues, this application provides a method, equipment, and a digital intelligent planning system for liver anatomical structure segmentation, used for preoperative CT / MR imaging liver anatomical structure segmentation. This overcomes the difficulties in preoperative imaging segmentation of liver tumors, hepatic veins, portal veins, and the eight segments of the liver, and improves the accuracy of liver anatomical structure segmentation.
[0008] This application provides a method for segmenting liver anatomical structures, including: Input the three-dimensional liver medical image into the trained frequency domain attention Transformer model; The liver anatomical structure segmentation results are obtained by segmenting using a trained frequency domain attention Transformer model; The trained frequency domain attention Transformer model includes: The image block encoding module is used to segment the input three-dimensional liver medical image into three-dimensional blocks, and convert them into feature tokens that can be processed by Transformer through flattening and linear projection. The pyramid-shaped Transformer encoder consists of a shallow encoder composed of multiple stacked Swin-Transformer modules or DenseNet modules, and a deep encoder composed of multiple stacked standard Transformer modules. Each encoder layer achieves feature downsampling through image patch fusion, and simultaneously extracts and outputs feature maps at various scales. The feature maps at each scale include: local texture detail feature maps and global semantic feature maps. The global frequency domain attention module is connected to each level of the encoder to enhance the structural boundaries of the feature maps at each scale and output the enhanced feature maps at each scale. The multi-scale feature fusion module adopts a bottom-up, stepwise upsampling and channel splicing fusion method to fuse enhanced feature maps at various scales, taking into account both global semantics and local details. The prediction head module receives the fused enhanced features and outputs the final 3D segmentation result through convolution operations, achieving multi-class segmentation of liver anatomical structures.
[0009] Furthermore, medical images used to train the frequency domain attention Transformer model are obtained in the following manner: Identify the liver mask; The hepatic vein and portal vein structures of three-dimensional liver medical images were segmented using a frequency domain attention Transformer model; then, the center lines of the main trunks of the hepatic vein and portal vein were extracted using VTK. The enhanced three-dimensional liver medical image is obtained by fusing three-dimensional liver medical images, liver mask and vascular centerline information and performing grayscale mixing enhancement. Unify the voxel spatial resolution of all enhanced 3D liver medical images, and normalize and crop out irrelevant areas outside the liver.
[0010] Furthermore, the frequency-domain attention Transformer model is trained and updated using a region re-detection loss function that violates the target topology constraints. for:
[0011]
[0012]
[0013] Where R and G represent the predicted result and the true annotation, respectively, if If False, it means the prediction result violates the target topological constraints; if False, it means there is no violation. For regions that violate the target topological constraints, λ is used to control the additional penalty coefficient for omitted regions. For cross-entropy loss and Dice coefficient loss Mixed loss composition Here are the voxel coordinates of the three-dimensional liver medical image, representing the values in the width direction, height direction, and slice direction, respectively.
[0014] Furthermore, To employ a convolution-based topological interaction module, all adjacent voxel pairs in the liver segmentation prediction results are checked and located based on a predefined adjacency graph of the liver eight-segment structure.
[0015] Furthermore, the shallow encoder is constructed by stacking two multi-layer Swin-Transformer modules or DenseNet modules, while the deep encoder is constructed by stacking two multi-layer standard Transformer modules.
[0016] Furthermore, the global frequency domain attention module is specifically used to enhance the structural boundaries of feature maps at each scale through three-dimensional discrete Fourier transform, high-pass band-pass filtering, three-dimensional discrete inverse Fourier transform, and convolution operations, and output enhanced feature maps at each scale.
[0017] This application provides an electronic device, including: Memory, which stores execution instructions; and The processor executes the instructions stored in memory, causing the processor to perform the methods described above.
[0018] This application also provides a minimally invasive surgical digital intelligent planning system, including: the aforementioned electronic device.
[0019] This application also provides a readable storage medium storing executable instructions, which are executed by a processor to implement the above-described method.
[0020] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0021] Compared with the prior art, this application has the following advantages: The frequency-domain attention Transformer model proposed in this application differs from existing segmentation models that combine convolutional networks and Transformers. On one hand, it effectively extracts local texture and global semantic features using a cascaded structure of Swing-Transformer and Transformer; on the other hand, it enhances texture details in the features through frequency-domain attention. Compared to existing models, it can extract more accurate local and global features.
[0022] This application proposes an anatomical structure-aware preprocessing method that incorporates prior knowledge of the hepatic vein and portal vein, bringing the advantages of traditional section-based segmentation methods into a deep learning model. Compared to existing methods for segmenting eight liver segments, the framework proposed in this application can obtain more accurate segmentation results for the eight liver segments, especially the boundaries of the eight liver segments.
[0023] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 A diagram of a frequency domain attention Transformer model according to an embodiment of this application is shown; Figure 2 An anatomical structure sensing preprocessing diagram according to an embodiment of this application is shown; Figure 3 A schematic diagram of a minimally invasive surgical digital intelligent planning system according to an embodiment of this application is shown. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0027] This application explores and analyzes the performance of Convolutional Neural Networks (CNNs) and Transformers in medical image segmentation tasks. CNN modules can effectively extract local texture and structural features, but their ability to model global semantic information is limited. Conversely, Transformers, through multi-head self-attention mechanisms, can capture long-range dependencies, thereby extracting richer global semantic features, while sliding window-based attention mechanisms can also capture sufficient local features. Therefore, this application designs a feature extraction model based on a pyramid-shaped Transformer, combining the Swin-Transformer with the standard Transformer. This approach not only extracts both local and global features simultaneously but also is applicable to input images with higher resolution, avoiding excessive computation. Furthermore, this model enhances the edge structure information of the feature map through a frequency domain attention module, enabling the network to pay more attention to vascular structures when processing CT / MR data containing anatomical priors, thereby obtaining more accurate liver segmentation boundaries.
[0028] This application provides a method for segmenting liver anatomical structures, including: Input the three-dimensional liver medical image into the trained frequency domain attention Transformer model; The liver anatomical structure segmentation result is obtained by segmenting the liver using the trained frequency domain attention Transformer model. The overall structure of the trained frequency domain attention Transformer model is as follows: Figure 1 As shown, it includes: The image block encoding module is used to segment the input three-dimensional liver medical image into three-dimensional blocks, and convert them into feature tokens that can be processed by Transformer through flattening and linear projection. The pyramid-shaped Transformer encoder consists of a shallow encoder composed of multiple stacked Swin-Transformer modules or DenseNet modules, and a deep encoder composed of multiple stacked standard Transformer modules. Each encoder layer achieves feature downsampling through image patch fusion, and simultaneously extracts and outputs feature maps at various scales. The feature maps at each scale include: local texture detail feature maps and global semantic feature maps. The global frequency domain attention module is connected to each level of the encoder to enhance the structural boundaries of the feature maps at each scale and output the enhanced feature maps at each scale. The multi-scale feature fusion module adopts a bottom-up, stepwise upsampling and channel splicing fusion method to fuse enhanced feature maps at various scales, taking into account both global semantics and local details. The prediction head module receives the fused enhanced features and outputs the final 3D segmentation result through convolution operations, achieving multi-class segmentation of liver anatomical structures.
[0029] The global frequency domain attention module is specifically used to enhance the structural boundaries of feature maps at each scale through three-dimensional discrete Fourier transform, high-pass band-pass filtering, three-dimensional discrete inverse Fourier transform, and convolution operations, and output enhanced feature maps at each scale.
[0030] Specifically, the pyramid-shaped Transformer encoder consists of four stages. The first stage begins by inputting the training data. Divided into three-dimensional volume data blocks (size: ), where the volume data block size Set to 2, This indicates the number of volume data blocks. Each 3D volume data block is then flattened and mapped via a learnable linear projection. Dimensional space is transformed into tokens. The first phase uses... A Swin-Transformer layer containing moving window multi-head self-attention (SW-MSA) and a multilayer perceptron (MLP) is used to extract local texture features. Subsequently, these local features will be reshaped into a feature map similar to that of the image. This data is used for feature extraction in the next stage or by the subsequent decoder. The structure of the second to fourth stages is similar to the first stage. First, a voxel patch fusion module (Patch-Merging) sets the voxel patch size to 2 to downsample the feature map from the previous stage, and then performs feature extraction. It is worth noting that the second stage still uses the Swin-Transformer to extract local features to reduce the computational cost of the model, while the third and fourth stages use the standard Transformer, including multi-head self-attention (MSA) and multilayer perceptron (MLP), to extract multi-scale global features. Finally, the model obtains two sets of local feature maps and two sets of global feature maps through the encoder: .
[0031] Next, a frequency domain attention module is applied to each feature map to further highlight key boundary structure information. Then, all feature maps are fused through a hierarchical feature fusion module (from coarse to fine), and finally, the Head module outputs the segmentation result. The frequency domain attention module acts on the feature maps. The process can be represented as:
[0032]
[0033]
[0034] in, and These represent the three-dimensional discrete Fourier transform and the three-dimensional inverse Fourier transform, respectively, applied to each channel of the feature map; The designed bandpass filter weighting map is used to suppress low-frequency information and preserve high-frequency edge structures in the frequency domain; , , The center position of the feature map; , Hyperparameters for controlling low-frequency attenuation; " indicates dot product; and They represent and Convolution operation is used to obtain the frequency domain attention map. .
[0035] After completing the Transformer encoder and frequency domain attention enhancement, the enhanced feature map can be obtained. Subsequently, the four feature fusion modules fuse this information step by step:
[0036]
[0037] in, Indicates channel splicing. This represents the residual convolution module. This indicates an upsampling operation. Ultimately, this is achieved by including a residual convolution module, upsampling, and... The prediction head module of the convolution obtains the final segmentation result. .
[0038] Furthermore, this application also obtains three-dimensional liver medical images for training based on anatomical structure-aware preprocessing. Specifically, according to the Couinaud segmentation system, the boundaries of liver segments are determined by five key planes, including the left umbilical fissure and the four main vascular trunks (middle hepatic vein, right hepatic vein, left portal vein, and right portal vein) formed by the hepatic veins and portal veins. Therefore, the segmentation framework proposed in this application aims to extract these main vascular trunk structures and explicitly reinforce these key information during the training phase of CT / MR data, enabling the segmentation network to focus more attention on relevant regions, thereby obtaining more accurate liver segment boundaries, such as... Figure 2 As shown.
[0039] For the input CT / MR images This preprocessing method first utilizes existing work to segment the liver mask. Based on the proposed model, the hepatic vein and portal vein structures were first segmented. Subsequently, they used the vascular centerline extraction method from VTK to extract the centerline of the main hepatic vessels from the segmented vessels. It is important to note that all points located on the centerline are filtered based on their corresponding vessel diameters; points with excessively small diameters are discarded. After this processing, three types of intermediate results are obtained: the input image, the liver mask, and the centerline of the intrahepatic vessels. Subsequently, a hybrid enhancement operation is used to fuse the original input with the liver mask and main vessel information to generate new CT / MR data. :
[0040] in Indicates the input image The grayscale range. After completing the blending enhancement, all voxels were uniformly resampled to ensure a spatial resolution of (1.0, 1.0, 1.0), and then... Z-score normalization was performed to reduce the intensity differences among different patients, resulting in... .
[0041] Finally, to reduce the size of the input data and minimize interference from irrelevant regions, region clipping was performed on all cases. Regions that do not contain the liver are removed to obtain the final image data used for training. :
[0042] The aforementioned technical approach of introducing missed re-detection loss in the eight-segment liver segmentation can effectively reduce missegmentation.
[0043] Furthermore, in multi-class segmentation tasks, the topological relationships between different foreground classes typically fall into two categories: containment and exclusion. In the Couinaud liver segmentation task, the relationships between different liver segments are exclusionary. Therefore, this application employs a convolution-based topological interaction module to refine the liver segmentation results based on a predefined adjacency graph of the eight liver segments. Check all adjacent voxel pairs and locate the positions that violate the target topological constraints. :
[0044] in, and These represent the predicted result and the actual annotation, respectively. If If the result is False, it indicates that the prediction result violates the target topological constraints; if it is False, it indicates that there is no violation. This application uses this strategy to re-detect missed regions (or regions that violate the rules) in the segmentation of the eight segments of the liver. And calculate the corresponding losses accordingly.
[0045] This application employs cross-entropy loss. With Dice coefficient loss The combined loss function:
[0046] Finally, the missed areas were re-examined for losses. Defined as:
[0047] in, Used to control the additional penalty coefficient for missed areas.
[0048] In summary, this application proposes an anatomically perceptive frequency-domain attention Transformer framework for preoperative liver anatomical structure segmentation in CT / MR images. This framework comprises three key components: a frequency-domain attention Transformer model, anatomically perceptive preprocessing, and a missed re-detection loss function. This framework overcomes the challenges of preoperative segmentation of the hepatic vein, portal vein, and eight liver segments, achieving more accurate liver anatomical structure segmentation compared to existing methods.
[0049] Based on the same inventive concept as the above disclosure, this disclosure also provides an electronic device. The electronic device of this disclosure includes at least one processor and at least one memory electrically connected to the processor. The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.
[0050] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between lines. The indirect connection method can be applied to the embodiments of this disclosure as long as it achieves the purpose of this disclosure.
[0051] Based on the same inventive concept disclosed above, this disclosure also provides a minimally invasive surgical digital intelligent planning system, including the aforementioned electronic device. The liver anatomical structure segmentation method of this application can be used in… Figure 3 The digital intelligent planning system for minimally invasive surgery shown is as follows. Figure 3 As shown, preoperative medical images such as CT or MRI are introduced, and a surgical planning visualization software system is constructed using the liver anatomical structure segmentation method of this application. This planning software system can directly observe the three-dimensional digital model of the three-dimensional anatomical structure of human skin (fat), ribs, liver, liver tumor, hepatic veins and portal vein, and the eight segments of the liver, thereby guiding and planning the surgical tool approach and operation path used for liver tumor ablation or resection.
[0052] Based on the same inventive concept, this disclosure also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the above-described method.
[0053] Although this application 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for segmenting the anatomical structure of the liver, characterized in that, include: Input the three-dimensional liver medical image into the trained frequency domain attention Transformer model; The liver anatomical structure segmentation result is obtained by segmenting the liver using the trained frequency domain attention Transformer model. The trained frequency domain attention Transformer model includes: The image block encoding module is used to segment the input three-dimensional liver medical image into three-dimensional blocks, and convert them into feature tokens that can be processed by Transformer through flattening and linear projection. The pyramid-shaped Transformer encoder consists of a shallow encoder composed of multiple stacked Swin-Transformer modules or DenseNet modules, and a deep encoder composed of multiple stacked standard Transformer modules. Each encoder layer achieves feature downsampling through image patch fusion, and simultaneously extracts and outputs feature maps at various scales. The feature maps at each scale include: local texture detail feature maps and global semantic feature maps. The global frequency domain attention module is connected to each level of the encoder to enhance the structural boundaries of the feature maps at each scale and output the enhanced feature maps at each scale. The multi-scale feature fusion module adopts a bottom-up, stepwise upsampling and channel splicing fusion method to fuse enhanced feature maps at various scales, taking into account both global semantics and local details. The prediction head module receives the fused enhanced features and outputs the final 3D segmentation result through convolution operations, achieving multi-class segmentation of liver anatomical structures.
2. The method according to claim 1, characterized in that, The medical images used to train the frequency domain attention Transformer model are obtained through the following methods: Identify the liver mask; The hepatic vein and portal vein structures of the three-dimensional liver medical image were segmented using the frequency domain attention Transformer model; then the center lines of the main trunks of the hepatic vein and portal vein were extracted using VTK. The enhanced three-dimensional liver medical image is obtained by fusing three-dimensional liver medical images, liver mask and vascular centerline information and performing grayscale mixing enhancement. Unify the voxel spatial resolution of all enhanced 3D liver medical images, and normalize and crop out irrelevant areas outside the liver.
3. The method according to claim 1, characterized in that, The frequency-domain attention Transformer model is trained and updated using a region re-detection loss function that violates the target topology constraints. for: Where R and G represent the predicted result and the true annotation, respectively, if If False, it means the prediction result violates the target topological constraints; if False, it means there is no violation. For regions that violate the target topological constraints, λ is used to control the additional penalty coefficient for omitted regions. For cross-entropy loss and Dice coefficient loss Mixed loss composition Here are the voxel coordinates of the three-dimensional liver medical image, representing the values in the width direction, height direction, and slice direction, respectively.
4. The method according to claim 3, characterized in that, The To employ a convolution-based topological interaction module, all adjacent voxel pairs in the liver segmentation prediction results are checked and located based on a predefined adjacency graph of the liver eight-segment structure.
5. The method according to claim 1, characterized in that, Shallow encoders are constructed by stacking two multi-layer Swin-Transformer modules or DenseNet modules, while deep encoders are constructed by stacking two multi-layer standard Transformer modules.
6. The method according to claim 1, characterized in that, The global frequency domain attention module is specifically used to enhance the structural boundaries of feature maps at each scale through three-dimensional discrete Fourier transform, high-pass band-pass filtering, three-dimensional discrete inverse Fourier transform, and convolution operations, and output enhanced feature maps at each scale.
7. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes execution instructions stored in the memory, causing the processor to perform the method of any one of claims 1 to 6.
8. A digital intelligent planning system for minimally invasive surgery, characterized in that, include: The electronic device according to claim 7.
9. A readable storage medium, characterized in that, The readable storage medium stores execution instructions, which, when executed by a processor, are used to implement the method of any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.