Pelvic image deformable registration method and system based on multi-scale dynamic convolution fusion

Abstract

The invention discloses a pelvic image deformable registration method and system based on multi-scale dynamic convolution fusion, and belongs to the technical field of medical image registration. Obtaining a pelvis CT image; inputting the pelvis CT image into a DKF-Block module for feature extraction, and obtaining the large-range features of the pelvis image; and inputting the large-range features of the pelvic image into an SCC-Block module for fusion to obtain a fusion result. According to the method, the problems of long consumed time, limited registration precision and the like of a traditional free form registration algorithm are effectively solved, and meanwhile, the limitation of a deep learning registration network in the aspects of receptive field range and large-displacement deformation feature capture is overcome. By capturing and fusing feature information on different scales, the method can better perceive and represent large-range deformation, thereby laying a solid foundation for subsequent automatic planning of a pelvic screw channel.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image registration, and in particular to a deformable registration method and system for pelvic images based on multi-scale dynamic convolution fusion. Background Art

[0002] Pelvic fracture is a serious orthopedic trauma, usually caused by high-energy trauma (such as traffic accidents, falls from heights, or industrial accidents), accounting for about 3% to 8% of all fractures. This type of fracture often involves multiple areas of the pelvis, especially the anterior and posterior pelvic rings, resulting in vertical instability of the pelvis. The treatment of pelvic fractures is divided into two categories based on stability: stable fractures are generally treated conservatively, while unstable fractures require surgical intervention. Surgical treatment usually includes internal plate fixation and percutaneous screw fixation. Percutaneous screw fixation is widely used in clinical treatment due to its advantages of minimal invasiveness, rapid recovery, and low surgical cost.

[0003] Pelvic percutaneous screw fixation surgery usually relies on the surgeon's experience and manual operation. This process is not only time-consuming and laborious, but also lacks standardization, resulting in differences in planning results among different doctors. With the continuous development of medical computer software, pelvic screw channel planning methods have been widely studied. Existing studies have proposed methods for screw channel planning using medical imaging software. These methods usually require manual extraction and adjustment of the pelvic model, and manual operation of the software to plan the screw position, which often leads to strong subjectivity in the planning results and difficulty in obtaining the best results. Recently, some studies have used computer programs to assist in screw channel planning by developing pelvic screw channel search algorithms. Although these computer-assisted methods effectively reduce subjectivity, they still require a large amount of manual annotation to run the program, which consumes a lot of time, which is extremely disadvantageous for emergency surgery.

[0004] In recent years, significant progress has been made in the study of automatic screw planning using registration technology. By analyzing 523 pelvic subacetabular channels using a free-form registration algorithm (Free Form), an average morphological model of the pelvis was constructed through the registration algorithm. They manually marked the screw entry and exit areas, mapped these areas to the rest of the pelvis, and then obtained the pelvic screw path through full traversal calculation. A multi-template-based automatic screw planning method was also proposed. This method manually planned the screw trajectories of 40 pelvic templates and used the active shape model (ASM) registration method to map the template path to the patient's pelvis. It plans a path suitable for a specific patient by optimizing the cumulative function.

[0005] After investigation and analysis, the current pelvic medical image registration technology has the following three problems: (1) Existing pelvic screw automatic planning technologies based on image registration usually rely on iterative functions based on energy optimization to achieve free-form registration. The main disadvantage is that the optimization parameters need to be recalibrated each time the registration is performed, which is time-consuming and limits the real-time performance of the pelvic screw automatic planning technology.

[0006] (2) Current free-form registration methods often rely on general mathematical models for global deformation when processing pelvic images, which makes it difficult to fully characterize the anatomical differences between individual pelvises. Since the specific features of each pelvis cannot be effectively identified, the registration accuracy of such methods is easily affected when processing data from different patients, thereby reducing the accuracy and reliability of automatic pelvic screw planning.

[0007] (3) Deep learning registration algorithms have made up for the shortcomings of free-form registration to a certain extent due to their learnability and real-time inference speed. However, for human-specific anatomical structures such as the pelvis, they still face many challenges in practical applications. Since pelvic registration often involves large-scale, highly nonlinear deformations, existing networks are difficult to fully capture and characterize long-displacement feature information, resulting in registration accuracy and robustness that are difficult to meet clinical needs. In addition, the spatial and channel information of the image is not fully mined for deformation correlation information, which is particularly important for registration technology. Summary of the invention

[0008] The purpose of the present invention is to provide a deformable registration method and system for pelvic images based on multi-scale dynamic convolution fusion, to accurately register pelvic CT images and provide a prerequisite for subsequent automatic planning of pelvic screw channels.

[0009] In order to solve the above technical problems, the present invention provides a deformable registration method for pelvic images based on multi-scale dynamic convolution fusion, comprising the following steps: Obtain pelvic CT images; The pelvic CT image is input into the DKF-Block module for feature extraction to obtain large-scale features of the pelvic image; The large-scale features of the pelvic image are input into the SCC-Block module for fusion to obtain the fusion result.

[0010] Preferably, the pelvic CT image is input into the DKF-Block module for feature extraction to obtain large-scale features of the pelvic image, which specifically includes the following steps: Downsample the pelvic CT image features to obtain the input features of DKF-Block ; Input features of DKF-Block After the 3×3×3 standard convolution layer, the Leaky ReLu activation layer, and the instance regularization layer, the intermediate features are obtained. ; Intermediate Features Enter DKF-Layer to extract multi-scale features and output fused features ; The fused features After the 3×3×3 standard convolution layer and the original features of the residual branch are added element by element, the output features are obtained , as a large-scale feature of pelvic images.

[0011] Preferably, the intermediate feature Enter DKF-Layer to extract multi-scale features and output fused features , specifically including the following steps: DKF-Layer includes high-resolution branch and low-resolution branch; The high-resolution branch extracts image detail features through 3×3×3 depth separation convolution ; After downsampling, the low-resolution branch uses 5×5×5 and 7×7×7 large kernel depthwise separable convolutions to capture the contextual information of a larger receptive field in parallel, and obtains the first parallel feature and the first parallel feature ; The first parallel feature and the first parallel feature Splice along the channel dimension to obtain the spliced ​​features; The global maximum pooling layer is used to aggregate the concatenated features globally, and the dynamic weights are generated through the Sigmoid activation function. and ; Through dynamic weight and calibration and The importance of , so as to realize the dynamic selection and fusion of features obtained from different large kernel convolutions, and obtain the fused low-resolution features; The fused low-resolution features are restored to high resolution by upsampling and then added element by element to obtain the fused features. ; .

[0012] Preferably, the large-scale features of the pelvic image are input into the SCC-Block module for fusion to obtain a fusion result, which specifically includes the following steps: The large-scale features of the pelvic image As the input features of the lower encoder and the features of the same encoder, the lower decoder features are first upsampled as , and then combined with the encoder features at the same level as Perform concat operation to obtain the input features of SCC-Block ; Input characteristics of SCC-Block The core component SCC-Layer extracts the image space-channel information and fuses the feature map to obtain the output features of the residual branch. ; Output features of the residual branch It passes through a standard convolution layer with a convolution kernel size of 3, a LeakyReLu activation layer, an instance regularization layer, and another standard convolution layer with a convolution kernel size of 3, and the output features of the residual branch Add element by element to get the output features of SCC-Block , as the fusion result.

[0013] Preferably, the input features of SCC-Block The core component SCC-Layer extracts the image space-channel information and fuses the feature map to obtain the output features of the residual branch. , specifically including the following steps: SCC-Layer uses the cross attention mechanism to convert the input features of SCC-Block Process the spatial attention branch and channel attention branch separately, and dynamically generate spatial attention weights and channel attention weights ; Cross attention introduces spatial weights into the calculation of channel weights, and integrates channel weights into the generation of spatial weights to obtain the output features of the residual branch .

[0014] Preferably, SCC-Layer uses the cross attention mechanism to convert the input features of SCC-Block into Process the spatial attention branch and channel attention branch separately, and dynamically generate spatial attention weights and channel attention weights ; Cross attention introduces spatial weights into the calculation of channel weights, and integrates channel weights into the generation of spatial weights, specifically including the following steps: Channel attention weight branch on input features , Perform 1×1×1 standard convolution and element-wise addition respectively, and then reduce the number of channels Perform element-wise multiplication and use the SoftMax normalization function to get the channel result ; The spatial attention weight branch uses a global maximum pooling layer to highlight input features. The spatial features of the convolution kernel size of 3×3×3 are used to further extract the features, and the number of channels is expanded. Perform element-wise multiplication and obtain spatial results through the SoftMax normalization function ; In SCC-Layer, the input feature x of SCC-Block is sequentially compared with the spatial result , Channel Results After fusion, the output features of the residual branch are obtained .

[0015] Preferably, the loss function of the DKF-Block module and the SCC-Block module includes: the first part is the similarity calculation between the deformed moving image and the fixed image; the second part is the spatial regularization of the predicted deformation field to make it as smooth as possible and to avoid image folding to the greatest extent; ; Where: represents the image similarity loss function; represents the deformation field regularization; λ is the regularization hyperparameter; Represents a moving image; Represents a fixed image; represents the regularization term.

[0016] Preferably, the similarity calculation between the deformed moving image and the fixed image specifically includes the following steps: Using local normalized cross-correlation as image similarity loss function ; ; ; Where: Ω represents the set of all points in the three-dimensional voxel space; and Represents the average voxel value within a local window of size 9×9×9 centered on voxel P.

[0017] Preferably, the predicted deformation field is spatially regularized, specifically comprising the following steps: Use regularization penalty to achieve smooth regularization : ; Where: Represents the gradient value of voxel p in the three-dimensional registration space. The larger the gradient, The possibility of non-smoothness at point P is greater.

[0018] The present invention also provides a pelvic image deformable registration system based on multi-scale dynamic convolution fusion, comprising: An acquisition module, used for acquiring a pelvic CT image; A feature extraction module is used to input the pelvic CT image into the DKF-Block module for feature extraction to obtain large-scale features of the pelvic image; The fusion module is used to input the large-scale features of the pelvic image into the SCC-Block module for fusion to obtain the fusion result.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The deformable registration algorithm for pelvic images based on multi-scale dynamic convolution fusion proposed in the present invention improves the traditional deep learning registration model, designs a multi-scale dynamic convolution fusion module and a spatial channel coupling module, and realizes fast and accurate registration of pelvic CT images. While effectively solving the problems of long time consumption and limited registration accuracy existing in the traditional free-form registration algorithm, it also overcomes the limitations of the deep learning registration network in terms of receptive field range and large displacement deformation feature capture. By capturing and fusing feature information at different scales, this method can better perceive and characterize large-scale deformations, thus laying a solid foundation for the subsequent automatic planning of pelvic screw channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0021] Figure 1 This is the DMSC-Reg architecture diagram; Figure 2 It is the internal structure of DKF-Block and DKF-Layer; Figure 3 It is the internal architecture of SCC-Block and SCC-Layer. DETAILED DESCRIPTION

[0022] Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present invention, so the present invention is not limited to the specific implementation disclosed below.

[0023] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0024] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0025] The present invention is further described in detail below in conjunction with the accompanying drawings: In order to better illustrate the technical effect of the present invention, the present invention provides the following specific embodiments to illustrate the above technical process: Embodiment 1: A pelvic image deformable registration algorithm based on multi-scale dynamic convolution fusion (Dynamic Multi-Scale Convolutional Registration Network, DMSC-Reg), the improvements of which are as follows: (1) To solve the problem of distant correspondence between features in pelvic CT images, the DMSC-Reg encoder is designed based on multi-scale information fusion. The dynamic kernel fusion block (DKF-Block) designed in this chapter is used in the upper half of the encoder to extract large-scale features of pelvic images in a high-resolution dimension, which can retain more related detail information between pelvic images. It also introduces large kernel depthwise separable convolution, downsampling, upsampling and residual connection processes. These designs not only deepen the overall structure of the algorithm, but also control the computational complexity and memory consumption of the algorithm, minimize the number of parameters in the registration model, and facilitate the deployment of the algorithm.

[0026] (2) In order to solve the problem of insufficient extraction of image spatial and channel feature information, the DMSC-Reg decoder is designed based on the concept of spatial-channel feature fusion and introduces the spatial-channel coupling block (SCC-Block) proposed in this chapter at the decoder layer. This module can fully integrate the spatial and channel information of the decoder features and the upsampled image features, effectively enhancing the deformation and displacement perception ability between pelvic images, thereby further improving the feature expression and decoding performance.

[0027] This paper discusses how to accurately register pelvic CT images to provide a prerequisite for the subsequent automatic planning of pelvic screw channels. A deformable registration algorithm for pelvic images based on multi-scale dynamic convolution fusion is proposed, and an experiment is set up for detailed analysis to verify the feasibility of the proposed method. The model architecture of the present invention is shown in the figure Figure 1 shown.

[0028] (1) Description of the multi-scale dynamic convolution fusion module (DKF-Block), such as Figure 2 As shown; Due to the large deviation of the human pelvic shape, it is difficult for traditional deep learning registration networks to effectively capture long-distance dependencies, so a broader perception is needed to extract features from pelvic CT images. To solve this problem, the present invention proposes a DKF-Block module. Figure 2 The internal architecture of the DKF-Block module is shown. In general, the DKF-Block module is a residual connection structure that effectively transmits feature information through a deeper network depth. Its core component is the DKF-Layer, which is specifically used to extract long-distance feature association information in images.

[0029] Down Sampling is composed of two sets of convolution structures (which can reduce the feature size and expand the dimension). DKF-Block itself does not change the size and dimension of the feature dimension, so the feature dimension changes in Down Sampling. The feature at this time is the input feature of DKF-Block. The pelvic CT image features are first downsampled and then input into the DKF-Block.

[0030] Given the input features of DKF-Block First, after a 3×3×3 standard convolution layer and a Leaky ReLu activation layer (LR) and an instance regularization layer (IN), the intermediate feature representation is obtained. .

[0031] then, Enter DKF-Layer, extract multi-scale features, and output fused features .

[0032] final, After another 3×3×3 standard convolutional layer, the original features of the residual branch (the input features of the DKF-Block) are added element by element to obtain the output features The input-output feature dimension of DKF-Block is unchanged, and the feature dimension of the left branch is the same as that of the right branch.

[0033] The characteristic calculation formula of the DKF-Block module is as follows: ; ; ; DKF-Layer is the key to enabling DKF-Block to extract multi-scale features of images. Its core idea is to extract multi-scale features of images through large kernel convolutions of different sizes, significantly improving the network receptive field. In order to reduce the number of parameters and computational complexity of large kernel convolutions, DKF-Layer uses depthwise separable convolutions to replace the original standard convolutions. Specifically, DKF-Layer includes two feature branches with different resolutions. The high-resolution branch extracts image detail features through 3×3×3 depthwise separable convolutions. .

[0034] ; After downsampling, the low-resolution branch uses 5×5×5 and 7×7×7 large kernel depthwise separable convolutions to capture contextual information with a larger receptive field in parallel, and obtains and .

[0035] ; ; The two parallel features are then concatenated along the channel dimension. Next, the global maximum pooling layer (MAP) is used to perform global feature aggregation on the concatenated features, and the dynamic weights are generated through the Sigmoid activation function. and .

[0036] ; Weight , Able to perform adaptive adjustments based on the contextual information of the input features, through weight calibration and The importance of , thereby realizing the dynamic selection and fusion of features obtained from different large kernel convolutions. This mechanism further improves the ability of multi-scale feature expression, enabling the network to more effectively model long-distance dependencies and complex structural information in pelvic CT images. Finally, the fused low-resolution features are restored to high resolution through upsampling (UP) and added element by element to obtain the final output feature The overall feature calculation formula of DKF-Layer is as follows: ; (2) Description of the spatial channel coupling module (SCC-Block), such as Figure 2 As shown; In order to better adaptively fuse spatial-channel information into multi-scale local feature maps, this paper proposes SCC-Block. Figure 3 The internal architecture of the SCC-Block module is shown. Similar to DKF-Block, SCC-Block is also a residual connection structure. The difference is that the input of SCC-Block is the features of the encoder at the same level and the features sampled by the decoder at the next level. Features after splicing along the channel; output features of DKF-Block There are two directions: one is to be used as the input feature of the lower encoder, and the other is to be used as the feature of the same encoder as one of the input features of the SCC-Block. UpSampling consists of two sets of convolutional layers. Upsampling will enlarge the feature size and reduce the dimension to be consistent with the encoder feature dimension, so that the concat operation can be performed. (For example, the feature dimension of the same encoder is ×D×C, the lower decoder characteristics are / 2×D / 2×2C, after upsampling, the feature dimension becomes ×D×C), the features of the lower-level decoder are first upsampled, and then concat operation is performed with the features of the encoder at the same level (the blue block on the left in the figure below).

[0037] Given the input features of SCC-Block .

[0038] ; first The core component SCC-Layer extracts the image space-channel information and fuses the feature map to obtain the output result. .Then It passes through a standard convolution layer with a convolution kernel size of 3, a LeakyReLu activation layer (LR), an instance regularization layer (IN), and another standard convolution layer with a convolution kernel size of 3, and the output feature of the residual branch Add element by element to get the output features of SCC-Block The characteristic calculation formula of SCC-Block is as follows: ; ; The core of SCC-Block is to adaptively extract and fuse the spatial information and channel information in the input features through SCC-Layer. Specifically, SCC-Layer uses the cross-attention mechanism to process the input features along two branches separately and dynamically generate spatial attention weights. and channel attention weights In this process, cross attention introduces spatial weights into the calculation of channel weights and integrates channel weights into the generation of spatial weights, thus achieving deep interaction and optimization of spatial and channel dimension features. Ultimately, this fusion mechanism outputs an optimized feature representation. Through the design of the cross-attention mechanism, the model effectively constructs the deep association of multi-dimensional features and effectively enhances the coupling of spatial and channel information.

[0039] Specifically, the channel attention weight branch firstly , Perform 1×1×1 standard convolution and element-wise addition respectively, and then reduce the number of channels Perform element-wise multiplication and use the SoftMax normalization function to get the channel result .

[0040] ; The spatial attention weight branch is obtained by processing x. First, the global maximum pooling layer (MAP) is used to highlight the spatial features, and the depthwise separable convolution with a convolution kernel size of 3×3×3 is used to further extract the features. The subsequent operations are similar to the channel weight branch, and the expansion of the number of channels is similar to the expansion of the channel weight branch. Perform element-wise multiplication and finally obtain the spatial result through the SoftMax normalization function .

[0041] ; In SCC-Layer, the input feature x is sequentially compared with , After fusion, the output features of the residual branch are obtained The overall feature calculation formula of SCC-Layer is as follows: ; (3) Loss function; The overall loss function of the DMSC-Reg network is derived from the energy function of the traditional image registration algorithm. The loss function consists of two parts: the first part is the similarity calculation between the deformed moving image and the fixed image, and the second part is the spatial regularization of the predicted deformation field to make it as smooth as possible and minimize image folding: ; in represents the image similarity loss function, represents the deformation field regularization, and λ is the regularization hyperparameter.

[0042] Use Local Normalized Cross-Correlation (LNCC) as the image similarity loss function For moving images and fixed images For two images, if they are highly similar in the local window, the LNCC value approaches 1; when the grayscale distribution of the two images in the window is unrelated or weakly correlated, the LNCC value is close to 0; if the local window shows an inverse grayscale distribution (strong negative correlation), the LNCC value will be close to -1. Since the higher the LNCC value, the more similar the images are, in order to transform the optimization goal of "maximizing similarity" into "minimizing loss", a negative sign is added to the LNCC during implementation, that is, , and thus use it as the loss function. Specifically, LNCC( , ) is as follows: ; Where Ω represents the set of all points in the three-dimensional voxel space, and It represents the average voxel value in a local window of size 9×9×9 centered on voxel P.

[0043] Minimum image similarity loss function Prompt fixed image Approximation to moving images Therefore, a regularization penalty term is used to achieve a smooth regularization term : ; in Represents the gradient value of voxel p in the three-dimensional registration space. The larger the gradient, The more likely it is that the smoothness will be at point P. Therefore, the complete loss function of the DMSC-Reg network is: ; In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules, modules or units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units, modules or components can be combined or integrated into another device, or some features can be ignored or not executed.

[0044] The units may or may not be physically separated, and the components displayed as units may be one physical unit or multiple physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0045] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0046] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part, and / or installed from a removable medium. When the computer program is executed by the central processing unit (CPU), the above-mentioned functions defined in the method of the present invention are executed. It should be noted that the above-mentioned computer-readable medium of the present invention can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of an electrical, magnetic, optical, electromagnetic, infrared segment, or semiconductor, or any combination of the above.

[0047] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0048] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A deformable pelvic image registration method based on multi-scale dynamic convolution fusion, characterized in that: The following steps are involved: Obtain pelvic CT images; The pelvic CT image is input into the DKF-Block module for feature extraction to obtain large-scale features of the pelvic image; The large-scale features of the pelvic image are input into the SCC-Block module for fusion to obtain the fusion result.

2. The deformable pelvic image registration method based on multi-scale dynamic convolution fusion according to claim 1, characterized in that: The pelvic CT image is input into the DKF-Block module for feature extraction to obtain large-scale features of the pelvic image, which specifically includes the following steps: Downsample the pelvic CT image features to obtain the input features of DKF-Block ; Input features of DKF-Block After the 3×3×3 standard convolution layer, the Leaky ReLu activation layer, and the instance regularization layer, the intermediate features are obtained. ; Intermediate Features Enter DKF-Layer to extract multi-scale features and output fused features ; The fused features After the 3×3×3 standard convolution layer and the original features of the residual branch are added element by element, the output features are obtained , as a large-scale feature of pelvic images.

3. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 2, characterized in that: Intermediate Features Enter DKF-Layer to extract multi-scale features and output fused features , specifically including the following steps: DKF-Layer includes high-resolution branch and low-resolution branch; The high-resolution branch extracts image detail features through 3×3×3 depth separation convolution ; After downsampling, the low-resolution branch uses 5×5×5 and 7×7×7 large kernel depthwise separable convolutions to capture contextual information with a larger receptive field in parallel, and obtains the first parallel feature and the first parallel feature ; The first parallel feature and the first parallel feature Splice along the channel dimension to obtain the spliced ​​features; The global maximum pooling layer is used to aggregate the concatenated features globally, and the dynamic weights are generated through the Sigmoid activation function. and ; Through dynamic weight and calibration and The importance of , so as to realize the dynamic selection and fusion of features obtained from different large kernel convolutions, and obtain the fused low-resolution features; The fused low-resolution features are restored to high resolution by upsampling and then added element by element to obtain the fused features. ; 。 4. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 3, characterized in that: The large-scale features of the pelvic image are input into the SCC-Block module for fusion to obtain the fusion result, which specifically includes the following steps: The large-scale features of the pelvic image As the input features of the lower encoder and the features of the same encoder, the lower decoder features are first upsampled as , and then combined with the encoder features at the same level as Perform concat operation to obtain the input features of SCC-Block ; Input characteristics of SCC-Block The core component SCC-Layer extracts the image space-channel information and fuses the feature map to obtain the output features of the residual branch. ; Output features of the residual branch It passes through a standard convolution layer with a convolution kernel size of 3, a Leaky ReLu activation layer, an instance regularization layer, and another standard convolution layer with a convolution kernel size of 3, and the output feature of the residual branch Add element by element to get the output features of SCC-Block , as the fusion result.

5. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 4, characterized in that: Input characteristics of SCC-Block The core component SCC-Layer extracts the image space-channel information and fuses the feature map to obtain the output features of the residual branch. , specifically including the following steps: SCC-Layer uses the cross attention mechanism to convert the input features of SCC-Block Process the spatial attention branch and channel attention branch separately, and dynamically generate spatial attention weights and channel attention weights ; Cross attention introduces spatial weights into the calculation of channel weights, and integrates channel weights into the generation of spatial weights to obtain the output features of the residual branch .

6. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 5, characterized in that: SCC-Layer uses the cross attention mechanism to convert the input features of SCC-Block Process the spatial attention branch and channel attention branch separately, and dynamically generate spatial attention weights and channel attention weights ; Cross attention introduces spatial weights into the calculation of channel weights, and integrates channel weights into the generation of spatial weights, specifically including the following steps: Channel attention weight branch on input features , Perform 1×1×1 standard convolution and element-wise addition respectively, and then reduce the number of channels Perform element-wise multiplication and use the SoftMax normalization function to get the channel result ; The spatial attention weight branch uses a global maximum pooling layer to highlight input features. The spatial features of the convolution kernel size of 3×3×3 are used to further extract the features, and the number of channels is expanded. Perform element-wise multiplication and obtain spatial results through the SoftMax normalization function ; In SCC-Layer, the input feature x of SCC-Block is sequentially compared with the spatial result , Channel Results After fusion, the output features of the residual branch are obtained .

7. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 6, characterized in that: The loss functions of the DKF-Block module and the SCC-Block module include: the first part is the similarity calculation between the deformed moving image and the fixed image; the second part is to perform spatial regularization on the predicted deformation field to make it as smooth as possible and to avoid image folding to the greatest extent; ； Where: represents the image similarity loss function; represents the regularization of the deformation field; λ is the regularization hyperparameter; Represents a moving image; Represents a fixed image; represents the regularization term.

8. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 7, characterized in that: The similarity calculation between the deformed moving image and the fixed image specifically includes the following steps: Using local normalized cross-correlation as image similarity loss function ; ； ； Where: Ω represents the set of all points in the three-dimensional voxel space; and Represents the average voxel value within a local window of size 9×9×9 centered on voxel P.

9. The deformable registration method for pelvic images based on multi-scale dynamic convolution fusion according to claim 8, characterized in that: The predicted deformation field is spatially regularized, which includes the following steps: Use regularization penalty to achieve smooth regularization : ； Where: Represents the gradient value of voxel p in the three-dimensional registration space. The larger the gradient, The possibility of non-smoothness at point P is greater.

10. A deformable registration system for pelvic images based on multi-scale dynamic convolution fusion, used to implement the deformable registration method for pelvic images based on multi-scale dynamic convolution fusion as described in any one of claims 1 to 9, characterized in that: include: An acquisition module, used for acquiring a pelvic CT image; A feature extraction module is used to input the pelvic CT image into the DKF-Block module for feature extraction to obtain large-scale features of the pelvic image; The fusion module is used to input the large-scale features of the pelvic image into the SCC-Block module for fusion to obtain the fusion result.

Images