Lightweight wavelet convolution guide wire segmentation network model and double guide wire generation method

Abstract

The invention provides a lightweight wavelet convolution guide wire segmentation network model and a double guide wire generation method, and the construction of the model comprises the steps: obtaining an input image, carrying out the feature extraction of the input image through a Stem module, obtaining a bottom-layer feature, carrying out the processing of the bottom-layer feature through five wavelet encoders in an iteration manner, and carrying out the recognition of the processed bottom-layer feature through the five wavelet encoders. And inputting the features output by the wavelet encoders into a decoder in an iteration form to carry out decoding processing so as to obtain a segmented prediction map. The method comprises the following steps of: inputting the features output by the wavelet encoders into the decoder in an iteration form to carry out decoding processing on the features output by the wavelet encoders; the wavelet transform attention convolution provided by the invention is specially used for capturing complex edge and texture nuisance features of the guide wire. A channel attention mechanism is introduced after a wavelet decomposition process to enhance feature representation. The combination of channel attention can selectively emphasize key frequency components, significantly improve the ability of the network to capture complex texture modes, and improve the precision of guide wire segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and particularly relates to a lightweight wavelet convolutional wire segmentation network model and a dual-wire generation method. Background Art

[0002] The prevalence, disability rate, and mortality rate of cardiovascular and cerebrovascular diseases are all very high. Given the current technological progress, vascular interventional surgery is considered the most effective treatment method. To avoid damaging the blood vessel wall, the movement of the wire within the blood vessel must be very precise. Therefore, accurate wire segmentation technology is crucial for the success of vascular interventional treatment. However, wire segmentation poses significant challenges. In X-ray sequences, the slender wire usually exhibits characteristics of class imbalance and low signal-to-noise ratio, resulting in potential artifacts caused by the patient's heartbeat and respiration. In addition, the wire may be blocked by blood vessels and ribs, and the similar structures make the wire prone to misclassification.

[0003] Currently, wire segmentation methods are mainly divided into two types: traditional methods and deep learning-based methods. Traditional methods are similar to edge detection techniques. The focus of these methods is to extract the curve contour of the wire trajectory. For example, B-spline curve fitting technology is usually used to model the wire, and fluorophores can be used to locate the endpoints for tracking. There are two main types of deep learning methods: models based on convolutional neural networks (CNNs) and models based on Transformers. Convolutional neural networks provide a unique advantage of inductive bias in medical image segmentation, especially in the case of limited datasets. Thanks to the characteristics of the self-attention mechanism, Transformers can effectively capture global dependencies, but they also bring the problem of a large number of parameters. To further solve the computational overhead problem related to real-time wire segmentation tasks, the development of lightweight architectures is imperative. There is currently no dedicated lightweight network designed specifically for wire segmentation, but several general lightweight medical image segmentation networks have been proposed by scholars, such as UNeXt, CMUNeXT, etc. Although this method reduces parameters and improves efficiency by simplifying the model, it affects the preservation of key texture details. Summary of the Invention

[0004] In view of the above situation, the main purpose of the present invention is to propose a lightweight wavelet convolutional wire segmentation network model and a dual-wire generation method to solve the above technical problems.

[0005] The present invention proposes a lightweight wavelet convolutional wire segmentation network model, and the construction of the model includes the following steps: Step 1, constructing a Stem module based on two-dimensional convolution, constructing a wavelet encoder based on wavelet transform attention convolution, pointwise convolution, and two-dimensional convolution, and constructing a decoder based on grouped convolution and pointwise convolution. The Stem module, wavelet encoder, and decoder constitute the wire segmentation network model; Step 2: Obtain the input image, extract features from the input image using the Stem module to obtain the underlying features, and sequentially process the underlying features through five wavelet encoders in an iterative manner to respectively obtain the features output by the five wavelet encoders; Step 3: Input the features output by the wavelet encoder into the decoder in an iterative manner for decoding processing to obtain a segmentation prediction map; Step 4: Based on the segmentation prediction map, construct a binary cross-entropy loss function and a Dice loss function respectively, and use the binary cross-entropy loss function and the Dice loss function to optimize the wire segmentation network model to obtain the final lightweight wavelet convolutional wire segmentation network model.

[0006] The present invention also proposes a method for generating double wires. Among them, the method applies the lightweight wavelet convolutional wire segmentation network model as described above. The method includes the following steps: S801: Obtain two single-wire images and their corresponding ground truth labels, and generate a binary mask based on the ground truth label corresponding to the second single-wire image; S802: Based on the binary mask, perform pixel replacement on the two single-wire images and their corresponding ground truth labels respectively to obtain a pseudo double-wire image and its corresponding ground truth label; S803: Input the pseudo double-wire image into the final lightweight wavelet convolutional wire segmentation network model for processing to obtain a segmentation prediction map.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A lightweight wire segmentation network model is proposed to adapt to the medical scenario of real-time wire segmentation. The present invention uses a fully convolutional architecture and designs a novel and lightweight network model specifically for real-time wire segmentation tasks, achieving the best trade-off between segmentation performance and computational consumption. Among many wire segmentation network models, it shows the smallest number of parameters and computational complexity, and the training method is fast, and it can also present the best segmentation effect; 2. The wavelet transform attention convolution proposed by the present invention is specifically used to capture the complex edge and texture subtle difference features of the wire. A channel attention mechanism is introduced after the wavelet decomposition process to enhance the feature representation. The combination of channel attention can selectively emphasize key frequency components, significantly improving the network's ability to capture complex texture patterns and improving the accuracy of wire segmentation; 3. The proposed double-guidewire generation algorithm effectively solves the dilemma of scarce double-guidewire data. During the previous guidewire segmentation training process, the shortage of guidewire samples, especially the scarcity of double-guidewire samples, restricted the training effect of the model. The present invention proposes a guidewire data generation method for converting single-guidewire images into double-guidewire images, aiming to overcome the problem of limited double-guidewire data and establish a more robust and general solution for the guidewire segmentation task.

[0008] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is the flowchart of the steps of the lightweight wavelet convolutional guidewire segmentation network model proposed by the present invention; Figure 2 is the model framework diagram of the lightweight wavelet convolutional guidewire segmentation network model proposed by the present invention; Figure 3 is the architecture diagram of the wavelet transform attention convolution in the specific embodiment of the present invention; Figure 4 is the processing flowchart of the wavelet transform attention convolution in the specific embodiment of the present invention; Figure 5 is the algorithm framework diagram of the double-guidewire generation method proposed by the present invention; Figure 6 is the comparison diagram between the double-guidewire data generated by the double-guidewire generation method proposed by the present invention and the actual double-guidewire data. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0011] Referring to the following description and drawings, these and other aspects of the embodiments of the present invention will become clear. In these descriptions and drawings, some specific embodiments of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0012] Please refer to Figure 1 , this embodiment provides a lightweight wavelet convolutional guidewire segmentation network model, and the construction of this model includes the following steps: Step 1: Construct a Stem module based on two-dimensional convolution, construct a wavelet encoder based on wavelet transform attention convolution, pointwise convolution, and two-dimensional convolution, and construct a decoder based on grouped convolution and pointwise convolution. The Stem module, wavelet encoder, and decoder constitute a guide wire segmentation network model.

[0013] Step 2: Obtain an input image, use the Stem module to extract features from the input image to obtain underlying features, and process the underlying features iteratively through five wavelet encoders in sequence to obtain the features output by the five wavelet encoders respectively.

[0014] Please refer to Figure 2 , in Step 2, obtain an input image, use the Stem module to extract features from the input image to obtain underlying features, and process the underlying features iteratively through five wavelet encoders in sequence to obtain the features output by the five wavelet encoders respectively, which specifically includes the following sub-steps: Obtain an input image, and sequentially pass the input image through two-dimensional convolution, ReLU activation function, and batch normalization to obtain underlying features. The following relational expressions exist in the corresponding process: ; Among them, represents the underlying features, represents passing through batch normalization, represents passing through the ReLU activation function, represents passing through two-dimensional convolution operation, represents the input image; S201: Sequentially pass the underlying features through wavelet transform attention convolution, GELU activation function, and batch normalization, and perform residual addition with the underlying features to obtain the intermediate features after wavelet transform attention convolution and residual addition in the first cycle. The following relational expressions exist in the corresponding process: ; Among them, represents the intermediate features after wavelet transform attention convolution and residual addition in the first cycle, represents passing through batch normalization, represents passing through the GELU activation function, represents passing through wavelet transform attention convolution operation, represents residual addition; S202: Sequentially pass the intermediate features after wavelet transform attention convolution and residual addition in the first cycle through pointwise convolution, GELU activation function, and batch normalization to obtain the intermediate features after the first pointwise convolution in the first cycle. The following relational expressions exist in the corresponding process: ; Among them, represents the intermediate feature passing through the first pointwise convolution in the first cycle, represents passing through the pointwise convolution operation; S203. The intermediate feature passing through the first pointwise convolution in the first cycle is successively passed through pointwise convolution, GELU activation function, and batch normalization to obtain the feature output in the first cycle. The following relational expressions exist in the corresponding process: ; Among them, represents the feature output in the first cycle; S204. Taking the feature output in the first cycle as the input, repeat the steps of S201 to S203 in an iterative form times, to obtain the feature output in the th cycle. The following relational expressions exist in the corresponding process: ; Among them, represents the intermediate feature passing through wavelet transform attention convolution and residual addition in the th cycle, represents the index of the number of cycles, and ; represents the total number of cycles, represents the feature output in the th cycle, represents the intermediate feature passing through the first pointwise convolution in the th cycle, represents the feature output in the th cycle; S205. The feature output in the th cycle is successively passed through two-dimensional convolution, GELU activation function, batch normalization, and max pooling to obtain the feature output by the first wavelet encoder. The following relational expressions exist in the corresponding process: ; Among them, represents the feature output by the first wavelet encoder, represents passing through the max pooling operation.

[0015] Taking the feature output by the first wavelet encoder as the input, repeat the steps of S201 to S205 in an iterative form 4 times, and respectively obtain the feature output by the second wavelet encoder, the feature output by the third wavelet encoder, the feature output by the fourth wavelet encoder, and the feature output by the fifth wavelet encoder. The feature output by the fifth wavelet encoder is used as the bottleneck feature.

[0016] Further, please refer to Figure 3 and Figure 4 , the wavelet transform attention convolution operation includes the following sub-steps: Pass the input features through a two-dimensional Haar wavelet transform and a two-dimensional convolution operation in sequence to obtain the low-frequency component and the high-frequency components. The following relational expressions exist in the corresponding process: ; Among them, represents the low-pass filter, both represent the high-pass filters, represents the low-frequency component of represents the index of the wavelet level, represents the high-frequency component in the horizontal direction, represents the high-frequency component in the vertical direction, represents the high-frequency component in the diagonal direction; Construct an SE module based on the channel attention mechanism, and use the SE module to process the three high-frequency components generated by the wavelet transform respectively to obtain three enhanced high-frequency components. The following relational expressions exist in the corresponding process: ; Among them, both represent the enhanced high-frequency components, represents being processed by the SE module; Perform a two-dimensional convolution operation on the low-frequency component and the enhanced high-frequency components to obtain the low-frequency component and the high-frequency components after the two-dimensional convolution operation. The following relational expressions exist in the corresponding process: ; Among them, represents the low-frequency component after the two-dimensional convolution operation, both represent the high-frequency components after the two-dimensional convolution operation, represents the weight matrix; Perform a transposed convolution operation on the low-frequency component and the high-frequency components after the two-dimensional convolution operation to obtain the output feature of the inverse wavelet transform at wavelet level . The following relational expressions exist in the corresponding process: ; Among them, represents the output feature of the inverse wavelet transform at wavelet level , represents being processed by the transposed convolution operation; Perform a two-dimensional convolution operation on the input features and add the result to the output features of the inverse wavelet transform at wavelet level 1 in a residual manner to obtain the features of the wavelet transform attention convolution output. The following relational expressions exist in the corresponding process: ; Among them, represents the features of the wavelet transform attention convolution output, represents the input features, represents the output features of the inverse wavelet transform at wavelet level 1.

[0017] It should be noted that the input image includes a single guide wire image and a pseudo double guide wire image generated by the double guide wire generation method; in Figure 3 , represents element-wise multiplication; Figure 4 In (a) of Figure 4 , the LH features with and without the SE module are compared: the left figure shows the LH features without the SE module, and the right figure shows the features weighted by the SE module. It can be seen that the application of the SE module significantly enhances the details of the high-frequency components, thus more effectively highlighting the key features and suppressing irrelevant information. In addition,

[0018] Step 3: Input the features output by the wavelet encoder into the decoder in an iterative manner for decoding processing to obtain a segmentation prediction map.

[0019] In Step 3, input the features output by the wavelet encoder into the decoder in an iterative manner for decoding processing to obtain a segmentation prediction map, which specifically includes the following sub-steps: S301: Pass the bottleneck features through upsampling, two-dimensional convolution, ReLU activation function, and batch normalization processing in sequence to obtain the preprocessed bottleneck features. The following relational expressions exist in the corresponding process: ; Among them, represents the preprocessed bottleneck features, represents the bottleneck features, represents the upsampling operation; S302. After concatenating the preprocessed bottleneck features with the output features of the corresponding encoder, perform grouped convolution, GELU activation function, and batch normalization in sequence to obtain the fused features of the first encoder. The following relational expressions exist in the corresponding process: ; Among them, represents the fused features of the first encoder, represents the operation of grouped convolution, represents the operation of feature concatenation, represents the features output by the fourth wavelet encoder; S303. Pass the fused features of the first encoder through pointwise convolution, GELU activation function, and batch normalization in sequence to obtain the intermediate features of the first encoder. The following relational expressions exist in the corresponding process: ; Among them, represents the intermediate features of the first encoder; S304. Pass the intermediate features of the first encoder through pointwise convolution, GELU activation function, and batch normalization in sequence to obtain the features output by the first decoder. The following relational expressions exist in the corresponding process: ; Among them, represents the features output by the first decoder; In an iterative form, use the features output by the first decoder as the input and repeat the steps of S301 to S304 three times to obtain the features output by the second decoder, the features output by the third decoder, and the features output by the fourth decoder respectively. Use the features output by the fourth encoder as the segmentation prediction map. The following relational expressions exist in the corresponding process: ; Among them, represents the features output by the th decoder after preprocessing, represents the features output by the th decoder, represents the operation of upsampling, represents the index of the encoder, and ; represents the fused features of the th encoder, represents the features output by the th wavelet encoder, represents the output features of the previous decoder, represents the The intermediate features of an encoder, representing the features output by the

[0020] Step 4: Based on the segmentation prediction map, construct a binary cross-entropy loss function and a Dice loss function respectively, and use the binary cross-entropy loss function and the Dice loss function to optimize the guide wire segmentation network model to obtain the final lightweight wavelet convolution guide wire segmentation network model.

[0021] In Step 4, based on the segmentation prediction map, construct a binary cross-entropy loss function and a Dice loss function respectively, and use the binary cross-entropy loss function and the Dice loss function to optimize the guide wire segmentation network model to obtain the final lightweight wavelet convolution guide wire segmentation network model. Among them, the expression of the binary cross-entropy loss function is: ; where represents the binary cross-entropy loss, represents the total number of pixel points, represents the pixel label value, represents the pixel predicted value; The expression of the Dice loss function is: ; where represents the Dice loss.

[0022] Furthermore, the total training loss function is: ; where represents the total training loss.

[0023] Please refer to Figure 5 and Figure 6 , this embodiment also provides a method for generating a double guide wire. Among them, the method applies the lightweight wavelet convolution guide wire segmentation network model as described above, and the method includes the following steps: S801: Obtain two single guide wire images and their corresponding ground truths, and generate a binary mask based on the ground truth corresponding to the second single guide wire image.

[0024] In S801, obtain two single guide wire images and their corresponding ground truths, and generate a binary mask based on the ground truth corresponding to the second single guide wire image. There are the following relational expressions in the corresponding process: ; where represents the binary mask at the pixel point value represents pixel coordinates represents the value at the pixel point corresponding to the true label of the second single guide wire image value

[0025] S802. Based on the binary mask, perform pixel replacement on the two single guide wire images and their corresponding true labels respectively to obtain a pseudo double guide wire image and its corresponding true label.

[0026] In S802, based on the binary mask, perform pixel replacement on the two single guide wire images and their corresponding true labels respectively to obtain a pseudo double guide wire image and its corresponding true label. The following relational expressions exist during the corresponding process: ; where represents the value of the pseudo double guide wire image at the pixel point value represents the value of the true label corresponding to the pseudo double guide wire image at the pixel point value represents the value of the first single guide wire image at the pixel point value represents the value of the second single guide wire image at the pixel point value represents the value of the true label corresponding to the first single guide wire image at the pixel point value; S803. Input the pseudo double guide wire image into the final lightweight wavelet convolutional guide wire segmentation network model for processing to obtain a segmentation prediction map.

[0027] It should be noted that in Figure 6 , the three groups of figures on the left show some examples of the generated pseudo double guide wire data, and two groups of real double guide wire samples are also given on the right. The double guide wire generation method proposed by the present invention can not only expand the single guide wire to the double guide wire, but also realize various forms of data augmentation. This method supports simultaneously inputting multiple guide wire information to generate synthetic images containing three or more guide wires, significantly expanding the coverage range of the training data, providing more diverse training samples for the guide wire segmentation network model, and significantly enhancing the generalization ability.

[0028] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0029] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0030] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0031] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A lightweight wavelet convolutional guide wire segmentation network model, characterized in that, The construction of the model includes the following steps: Step 1: Construct a Stem module based on two-dimensional convolution, construct a wavelet encoder based on wavelet transform attention convolution, pointwise convolution, and two-dimensional convolution, and construct a decoder based on grouped convolution and pointwise convolution. The Stem module, wavelet encoder, and decoder constitute a guidewire segmentation network model; Step 2: Obtain an input image, use the Stem module to extract features from the input image to obtain underlying features, and iteratively process the underlying features through five wavelet encoders in sequence to respectively obtain the features output by the five wavelet encoders; Step 3: Iteratively input the features output by the wavelet encoder into the decoder for decoding processing to obtain a segmentation prediction map; Step 4: Based on the segmentation prediction map, construct a binary cross-entropy loss function and a Dice loss function respectively, and use the binary cross-entropy loss function and the Dice loss function to optimize the guidewire segmentation network model to obtain the final lightweight wavelet convolution guidewire segmentation network model.

2. The lightweight wavelet convolution guide wire segmentation network model according to claim 1, wherein In Step 2, when obtaining the input image, using the Stem module to extract features from the input image to obtain underlying features, and iteratively processing the underlying features through five wavelet encoders in sequence to respectively obtain the features output by the five wavelet encoders, it specifically includes the following sub-steps: Obtain the input image, and sequentially process the input image through two-dimensional convolution, ReLU activation function, and batch normalization to obtain underlying features; S201: Sequentially process the underlying features through wavelet transform attention convolution, GELU activation function, and batch normalization, and perform residual addition with the underlying features to obtain the intermediate features after wavelet transform attention convolution and residual addition in the first cycle; S202: Sequentially process the intermediate features after wavelet transform attention convolution and residual addition in the first cycle through pointwise convolution, GELU activation function, and batch normalization to obtain the intermediate features after the first pointwise convolution in the first cycle; S203: Sequentially process the intermediate features after the first pointwise convolution in the first cycle through pointwise convolution, GELU activation function, and batch normalization to obtain the features output in the first cycle; S204. After repeating the steps of S201 to S203 a certain number of times with the features output in the first loop as the input in an iterative form, the features output in the th loop are obtained; S205. Sequentially pass the features output in the th cycle through two-dimensional convolution, GELU activation function, batch normalization, and max pooling to obtain the features output by the first wavelet encoder; Taking the features output by the first wavelet encoder as input and repeating the steps of S201 to S205 a certain number of times in an iterative manner, respectively obtain the features output by the second wavelet encoder, the features output by the third wavelet encoder, the features output by the fourth wavelet encoder, and the features output by the fifth wavelet encoder, and take the features output by the fifth wavelet encoder as bottleneck features.

3. The lightweight wavelet convolution guide wire segmentation network model according to claim 2, wherein In the step of obtaining the input image, and sequentially processing the input image through two-dimensional convolution, ReLU activation function, and batch normalization to obtain underlying features, there is the following relationship: ； Among them, represents the underlying feature, represents being processed by batch normalization, represents being processed by the ReLU activation function, represents being processed by a two-dimensional convolution operation, represents the input image; In the step of sequentially processing the underlying features through wavelet transform attention convolution, GELU activation function, and batch normalization, and performing residual addition with the underlying features to obtain the intermediate features after wavelet transform attention convolution and residual addition in the first cycle, there is the following relationship: ； Among them, represents the intermediate feature after wavelet transform attention convolution and residual addition in the first loop, represents after batch normalization processing, represents after processing by the GELU activation function, represents the wavelet transform attention convolution operation, represents residual addition; In the step of sequentially passing the intermediate features that have undergone wavelet transform attention convolution and residual addition in the first loop through pointwise convolution, GELU activation function, and batch normalization to obtain the intermediate features that have passed through the first pointwise convolution in the first loop, the following relational expressions exist: ； Among them, represents the intermediate feature after the first pointwise convolution in the first loop, represents the pointwise convolution operation; In the step of sequentially passing the intermediate features that have passed through the first pointwise convolution in the first loop through pointwise convolution, GELU activation function, and batch normalization to obtain the features output in the first loop, the following relational expressions exist: ； Among them, represents the feature output in the first loop; After repeating steps S201 to S203 a certain number of times with the features output in the first loop as input in an iterative form to obtain the features output in the nth loop, the following relational expression exists: ； Among them, represents the intermediate feature after wavelet transform attention convolution and residual addition in the th cycle, represents the index of the number of cycles, represents the total number of cycles, represents the feature output in the th cycle, represents the intermediate feature after the first pointwise convolution in the th cycle, represents the feature output in the th cycle; In the step of successively passing the features output in the th cycle through two-dimensional convolution, GELU activation function, batch normalization, and max pooling to obtain the features output by the first wavelet encoder, there is the following relational expression: ； Among them, represents the feature output by the first wavelet encoder, indicating that a max pooling operation has been performed.

4. The lightweight wavelet convolution guide wire segmentation network model according to claim 3, characterized in that, The wavelet transform attention convolution includes the following sub-steps: Sequentially pass the input features through two-dimensional Haar wavelet transform and two-dimensional convolution operation to obtain low-frequency components and high-frequency components. The following relational expressions exist during the corresponding process: ； Among them, represents a low-pass filter, both represent high-pass filters, represents the low-frequency component of represents the index of the wavelet level, represents the high-frequency component in the horizontal direction, represents the high-frequency component in the vertical direction, represents the high-frequency component in the diagonal direction; Construct an SE module based on the channel attention mechanism, and use the SE module to process the three high-frequency components generated by wavelet transform respectively to obtain three enhanced high-frequency components. The following relational expressions exist during the corresponding process: ； Among them, both represent the enhanced high-frequency components, indicating being processed by the SE module; Perform two-dimensional convolution operation on the low-frequency component and the enhanced high-frequency components to obtain the low-frequency component and high-frequency components after two-dimensional convolution operation. The following relational expressions exist during the corresponding process: ； Among them, represents the low-frequency component after two-dimensional convolution operation, both represent the high-frequency components after two-dimensional convolution operation, represents the weight matrix; Perform a transposed convolution operation on the low-frequency component and the high-frequency component after the two-dimensional convolution operation to obtain the output feature of the inverse wavelet transform with the wavelet level of There is the following relational expression in the corresponding process: ； Among them, represents the output feature of the inverse wavelet transform at wavelet level , indicates that a transposed convolution operation has been performed; Perform two-dimensional convolution operation on the input features, and perform residual addition with the output features of the inverse wavelet transform at wavelet level 1 to obtain the features output by the wavelet transform attention convolution. The following relational expressions exist during the corresponding process: ； Among them, represents the feature output by the wavelet transform attention convolution, represents the input feature, represents the output feature of the inverse wavelet transform with a wavelet level of 1.

5. The lightweight wavelet convolutional guide wire segmentation network model according to claim 4, wherein In the step 3, the features output by the wavelet encoder are input into the decoder in an iterative form for decoding processing to obtain the segmentation prediction map, which specifically includes the following sub-steps: S301. Sequentially pass the bottleneck features through upsampling, two-dimensional convolution, ReLU activation function, and batch normalization to obtain the preprocessed bottleneck features; S302. After feature concatenation of the preprocessed bottleneck features and the output features of the corresponding encoder, sequentially pass through grouped convolution, GELU activation function, and batch normalization to obtain the fused features of the first encoder; S303. Sequentially pass the fused features of the first encoder through pointwise convolution, GELU activation function, and batch normalization to obtain the intermediate features of the first encoder; S304. Sequentially pass the intermediate features of the first encoder through pointwise convolution, GELU activation function, and batch normalization to obtain the features output by the first decoder; In an iterative form, use the features output by the first decoder as the input to repeat the steps of S301 to S304 a certain number of times to obtain the features output by the second decoder, the features output by the third decoder, and the features output by the fourth decoder respectively, and use the features output by the fourth encoder as the segmentation prediction map.

6. The lightweight wavelet convolution guide wire segmentation network model according to claim 5, wherein In the step of sequentially passing the bottleneck features through upsampling, two-dimensional convolution, ReLU activation function, and batch normalization to obtain the preprocessed bottleneck features, the following relational expressions exist: ； Among them, represents the bottleneck feature after preprocessing, represents the bottleneck feature, represents the upsampling operation; In the step of performing feature concatenation of the preprocessed bottleneck features and the output features of the corresponding encoder, and then sequentially passing through grouped convolution, GELU activation function, and batch normalization to obtain the fused features of the first encoder, the following relational expressions exist: ； Among them, represents the fused feature of the first encoder, represents after the grouped convolution operation, represents after the feature concatenation operation, represents the feature output by the fourth wavelet encoder; In the step of sequentially passing the fusion features of the first encoder through pointwise convolution, the GELU activation function, and batch normalization to obtain the intermediate features of the first encoder, the following relational expression exists: ； Among them, represents the intermediate feature of the first encoder; In the step of sequentially passing the intermediate features of the first encoder through pointwise convolution, the GELU activation function, and batch normalization to obtain the features output by the first decoder, the following relational expression exists: ； Among them, represents the feature output by the first decoder; In the step of taking the features output by the first decoder as the input and repeating the steps S301 to S304 a certain number of times in an iterative form, and respectively obtaining the features output by the second decoder, the features output by the third decoder, and the features output by the fourth decoder, and taking the features output by the fourth encoder as the segmentation prediction map, the following relational expression exists: ； Among them, represents the feature output by the th decoder after preprocessing, represents the feature output by the th decoder, represents the index of the encoder, represents the th fused feature of the encoder, represents the feature output by the th wavelet encoder, represents the output feature of the previous decoder, represents the th intermediate feature of the encoder, represents the th feature output by the decoder.

7. The lightweight wavelet convolution guide wire segmentation network model according to claim 6, wherein In the step 4, a binary cross-entropy loss function and a Dice loss function are respectively constructed based on the segmentation prediction map, and the wire segmentation network model is optimized by using the binary cross-entropy loss function and the Dice loss function to obtain the final lightweight wavelet convolution wire segmentation network model, where the expression of the binary cross-entropy loss function is: ； Among them, represents the binary cross-entropy loss, represents the total number of pixel points, represents the pixel label value, represents the pixel predicted value; The expression of the Dice loss function is: ； Among them, represents the Dice loss.

8. A method for generating double guide wires, characterized in that, The method applies the lightweight wavelet convolution wire segmentation network model according to any one of claims 1 to 7, and the method includes the following steps: S801. Obtain two single-wire images and their corresponding ground truth labels, and generate a binary mask based on the ground truth label corresponding to the second single-wire image; S802. Based on the binary mask, perform pixel replacement on the two single-wire images and their corresponding ground truth labels respectively to obtain a pseudo double-wire image and its corresponding ground truth label; S803. Input the pseudo double-wire image into the final lightweight wavelet convolution wire segmentation network model for processing to obtain a segmentation prediction map.

9. The double guide wire generation method according to claim 8, wherein In the step of obtaining two single-wire images and their corresponding ground truth labels, and generating a binary mask based on the ground truth label corresponding to the second single-wire image, the following relational expression exists: ； Among them, represents the value of the binary mask at the pixel point , represents the pixel coordinates, represents the value of the ground truth label corresponding to the second single guide wire image at the pixel point .

10. The double guide wire generation method according to claim 9, characterized in that, In the step of performing pixel replacement on the two single-wire images and their corresponding ground truth labels respectively based on the binary mask to obtain a pseudo double-wire image and its corresponding ground truth label, the following relational expression exists: ； Among them, represents the value of the pseudo double guide wire image at the pixel point . represents the value of the true label corresponding to the pseudo double guide wire image at the pixel point . represents the value of the first single guide wire image at the pixel point . represents the value of the second single guide wire image at the pixel point . represents the value of the true label corresponding to the first single guide wire image at the pixel point .

Images