Spine model generation method and system based on deep learning

Abstract

The present application relates to the field of spinal model generation, and specifically to a spinal model generation method and system based on deep learning. It collects multiple spinal local perspective CT images from multiple angles, and introduces image processing and analysis algorithms based on artificial intelligence and deep learning at the back end to analyze these spinal local perspective CT images, so as to learn and capture the key state feature information about the local perspective of the spine, and perform panoramic-based semantic fusion of these important features on the spinal state to reconstruct a more accurate spinal model. In this way, it is possible to realize intelligent recognition and reconstruction of the spinal model based on different spinal local perspective CT images, so as to assist doctors in diagnosing spinal diseases and formulating treatment plans.

Description

Technical Field

[0001] The present application relates to the field of spinal model generation, and specifically to a spinal model generation method and system based on deep learning. Background Art

[0002] Spinal diseases are a common disease that affects the quality of human life. With the changes in modern lifestyles, such as long periods of sitting and poor sitting posture, the incidence of spinal diseases has gradually increased, and is showing a trend of younger people. The spine, as the central axis of the human body, not only supports the trunk, protects the spinal cord and internal organs, but is also related to many physical diseases, such as cervical spondylosis and lumbar spondylosis. Therefore, the prevention, diagnosis and treatment of spinal diseases are of great medical significance.

[0003] In the medical field, the diagnosis and treatment planning of spinal diseases often rely on high-quality medical images. Existing medical images are often multiple two-dimensional images obtained through CT scans. These two-dimensional images are images of the patient's spine collected from different perspectives. They are used to help doctors identify changes in the spinal structure and make surgical plans. They can also help medical students and researchers better understand the spinal structure and related diseases.

[0004] In order to further help doctors more accurately identify abnormal changes in the spinal structure and assist in disease diagnosis and treatment planning, it is necessary to accurately reconstruct the three-dimensional structure of the spine from CT images, which is crucial for clinical diagnosis and surgical planning. An accurate three-dimensional spinal model can help doctors better understand the specific condition of the patient's spine, thereby making more accurate diagnoses and treatment plans. However, traditional spinal model generation schemes often rely on manual segmentation and reconstruction of CT scan images. This method has a low degree of intelligence, which affects the efficiency and quality of the three-dimensional structure reconstruction of the spine and is easily affected by the operator's technical level. In addition, due to the complex structure of the spine, traditional methods may have errors in identifying and reconstructing the spinal structure, especially when dealing with details and fuzzy boundaries.

[0005] Therefore, an optimized spine model generation scheme is desired. Summary of the invention

[0006] This application is made in consideration of the above problems. One purpose of this application is to provide a method and system for generating a spine model based on deep learning.

[0007] The embodiment of the present application provides a method for generating a spine model based on deep learning, which includes:

[0008] Acquire multiple spinal local view CT images collected from multiple angles;

[0009] Performing feature extraction based on the spinal state on each of the plurality of spinal local view CT images respectively to obtain a plurality of spinal local view state feature maps;

[0010] Each of the plurality of spinal local perspective state feature maps is respectively passed through a feature multi-scale perception enhancement module to obtain a plurality of enhanced spinal local perspective state feature maps;

[0011] Each of the plurality of enhanced spinal local perspective state feature maps is respectively subjected to a content differentiation module based on a spatial mask to obtain a plurality of spinal local perspective foreground highlighting state feature maps;

[0012] Inputting the plurality of spine local perspective foreground highlight state feature maps into a self-supervised feature sequence saliency fusion network to obtain a panoramic spine state fusion feature map;

[0013] A spinal model is generated based on the panoramic spinal state fusion feature map to obtain a generated spinal model.

[0014] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, wherein, feature extraction based on the spine state is performed on each of the multiple spine local view CT images to obtain multiple spine local view state feature maps, including:

[0015] Each of the multiple spinal local perspective CT images is passed through a spinal state feature extractor based on a void convolutional neural network model to obtain the multiple spinal local perspective state feature maps.

[0016] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, each of the multiple spine local perspective state feature maps is respectively subjected to a feature multi-scale perception enhancement module to obtain multiple enhanced spine local perspective state feature maps, including:

[0017] In the first branch, point convolution processing is performed on the spinal local perspective state feature map to obtain a first spinal local perspective state channel compression feature map; global mean pooling is performed on the first spinal local perspective state channel compression feature map to obtain a first spinal local perspective state channel compression feature vector; nonlinear activation processing is performed on the first spinal local perspective state channel compression feature vector to obtain a first spinal local perspective state channel local activation feature vector;

[0018] In the second branch, point convolution is performed on the spinal column local view state feature map to obtain a second spinal column local view state channel compression feature map;

[0019] In the third branch, the spinal local perspective state feature map is subjected to hole convolution encoding to obtain a spinal local perspective state receptive field expansion feature map; the spinal local perspective state receptive field expansion feature map is subjected to point convolution processing to obtain a spinal local perspective state global feature matrix with receptive field expansion; the spinal local perspective state global feature matrix with receptive field expansion is subjected to nonlinear activation to obtain a spinal local perspective state global activation feature matrix with receptive field expansion;

[0020] Multiply the receptive field expanded spine local perspective state global activation feature matrix by each feature matrix along the channel dimension of the second spine local perspective state channel compression feature map at each position point to obtain the second channel compression spine local perspective state global activation feature map;

[0021] Taking each position feature value in the local activation feature vector of the first spine local perspective state channel as a weighted weight, weighting each feature matrix along the channel dimension of the second spine local perspective state channel compression feature map to obtain the second spine local perspective state channel compression local activation feature map;

[0022] Adding the second channel compressed spine local view state global activation feature map and the second channel compressed spine local view state channel compressed local activation feature map by position to obtain the second channel compressed spine local view state multi-scale fusion activation feature map;

[0023] The second channel compressed spine local viewing state multi-scale fusion activation feature map is subjected to dilated convolution encoding to obtain the enhanced spine local viewing state feature map.

[0024] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, each of the multiple enhanced spine local perspective state feature maps is respectively subjected to a content differentiation module based on a spatial mask to obtain multiple spine local perspective foreground highlighting state feature maps, including:

[0025] Using the negative number of the characteristic value of each position of the enhanced spinal local perspective state characteristic map as the exponent of the natural constant to calculate the exponential function value based on the position with the natural constant as the base to obtain the spinal local perspective state class support characteristic map;

[0026] Calculate the sum of the characteristic values ​​of each position in the spinal local view state class support feature map and a constant one to obtain a spinal local view state modulation feature map;

[0027] Dividing the spinal local view state class support feature map and the spinal local view state modulation feature map by position to obtain a spinal local view state representation feature map;

[0028] Masking the spinal local view state representation feature map to obtain a spinal local view state mask weight feature map;

[0029] The spine local perspective state mask weight feature map and the spine local perspective state feature map are multiplied by position points to obtain the spine local perspective foreground highlighting state feature map.

[0030] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, the spine local view state characterization feature map is masked to obtain a spine local view state mask weight feature map, including:

[0031] The feature values ​​greater than or equal to a predetermined threshold in each position of the spinal local view state characterization feature map are set to one, and the rest are set to zero to obtain the spinal local view state mask weight feature map.

[0032] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, the plurality of spine local perspective foreground highlight state feature maps are input into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spine state fusion feature map, including:

[0033] Performing maximum-based global pooling processing, random-value-based global pooling processing, and average-based global pooling processing on each feature matrix along the channel dimension of each spinal local perspective foreground highlighting state feature map in the multiple spinal local perspective foreground highlighting state feature maps to obtain multiple spinal local perspective foreground highlighting state feature global maximum pooling feature vectors, multiple spinal local perspective foreground highlighting state feature global average pooling feature vectors, and multiple spinal local perspective foreground highlighting state feature global random value pooling feature vectors;

[0034] Calculate the position-weighted sum of the global maximum pooling feature vectors of the multiple spinal local perspective foreground highlighting state features, the global mean pooling feature vectors of the multiple spinal local perspective foreground highlighting state features and the global random value pooling feature vectors of the multiple spinal local perspective foreground highlighting state features corresponding to each group, the global maximum pooling feature vector of the spinal local perspective foreground highlighting state features, the global mean pooling feature vector of the spinal local perspective foreground highlighting state features and the global random value pooling feature vector of the spinal local perspective foreground highlighting state features to obtain a multi-scale pooling representation vector of the multiple spinal local perspective foreground highlighting state features;

[0035] Calculate the mean vector of the plurality of multi-scale pooled representation vectors of the local perspective foreground salient state features of the spine as the center of the local perspective foreground state cluster of the spine to obtain a semantic feature vector of the center of the local perspective foreground state cluster of the spine;

[0036] Calculate the semantic similarity score between each of the plurality of multi-scale pooled representation vectors of the local perspective foreground salient state features of the spine and the central semantic feature vector of the local perspective foreground state cluster of the spine to obtain a sequence of central semantic similarity score values ​​of the local perspective foreground state cluster of the spine;

[0037] Normalizing the sequence of the semantic similarity score values ​​of the spine local perspective foreground state cluster center to obtain a sequence of the semantic similarity attention weight coefficients of the spine local perspective foreground state cluster center;

[0038] Using the sequence of semantic similarity attention weight coefficients of the center of the local perspective foreground state cluster of the spine as weights, the position-by-position weighted sum of the multiple local perspective foreground highlight state feature maps of the spine is calculated to obtain the panoramic spine state fusion feature map.

[0039] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, the semantic similarity score between each of the plurality of spine local perspective foreground salient state feature multi-scale pooling representation vectors and the spine local perspective foreground state cluster center semantic feature vector is calculated to obtain a sequence of spine local perspective foreground state cluster center semantic similarity score values, including:

[0040] Calculate the matrix product between each weight coefficient matrix and the corresponding multi-scale pooling representation vector of the local perspective foreground salient state of the spine to obtain a sequence of weighted feature vectors of the local perspective foreground salient state of the spine;

[0041] Adding a corresponding bias vector to each spinal local perspective foreground highlighting state weighted feature vector in the sequence of spinal local perspective foreground highlighting state weighted feature vectors to obtain a sequence of spinal local perspective foreground highlighting state weighted bias adjustment feature vectors;

[0042] Inputting the sequence of the weighted bias-adjusted feature vectors of the local perspective foreground highlighting state of the spine into a sigmoid function for activation processing to obtain a sequence of activated feature vectors of the local perspective foreground highlighting state of the spine;

[0043] The products of each spinal local perspective foreground highlighting state activation feature vector in the sequence of spinal local perspective foreground highlighting state activation feature vectors and the transposed vector of the spinal local perspective foreground state cluster center semantic feature vector are calculated respectively to obtain a sequence of spinal local perspective foreground state cluster center semantic similarity score values.

[0044] For example, according to the deep learning-based spine model generation method of an embodiment of the present application, the spine model generation is performed based on the panoramic spine state fusion feature map to obtain a generated spine model, including:

[0045] The panoramic spine state fusion feature map is input into a spine model generator based on a diffusion model to obtain a generated spine model.

[0046] The embodiment of the present application also provides a spine model generation system based on deep learning, which includes:

[0047] A multi-angle acquisition module is used to acquire multiple spinal local view CT images acquired from multiple angles;

[0048] A feature extraction module, used for performing feature extraction based on the spinal state on each of the plurality of spinal local view CT images to obtain a plurality of spinal local view state feature maps;

[0049] An enhancement module, used for respectively passing each of the plurality of spinal local perspective state feature maps through a feature multi-scale perception enhancement module to obtain a plurality of enhanced spinal local perspective state feature maps;

[0050] A distinguishing module, used for respectively passing each of the plurality of enhanced spinal local perspective state feature maps through a content distinguishing module based on a spatial mask to obtain a plurality of spinal local perspective foreground highlighting state feature maps;

[0051] A fusion module, used for inputting the plurality of spine local perspective foreground highlight state feature maps into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spine state fusion feature map;

[0052] The spine model generation module is used to generate a spine model based on the panoramic spine state fusion feature map to obtain a generated spine model.

[0053] For example, according to the deep learning-based spine model generation system of an embodiment of the present application, the feature extraction module is used to:

[0054] Each of the multiple spinal local perspective CT images is passed through a spinal state feature extractor based on a void convolutional neural network model to obtain the multiple spinal local perspective state feature maps.

[0055] According to the deep learning-based spinal model generation method and system of the embodiment of the present application, the spinal model is intelligently identified and reconstructed based on CT images of different local spinal angles to assist doctors in diagnosing spinal diseases and formulating treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application are briefly introduced below. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0057] Figure 1 A schematic diagram of the application architecture of the method for generating a spine model based on deep learning in an embodiment of the present application is shown;

[0058] Figure 2 A flowchart of a method for generating a spine model based on deep learning in an embodiment of the present application is shown;

[0059] Figure 3 A flowchart of sub-step S540 of the method for generating a spine model based on deep learning in an embodiment of the present application is shown;

[0060] Figure 4 A schematic diagram of the structure of a spine model generation system based on deep learning in an embodiment of the present application is shown;

[0061] Figure 5 An application scenario diagram of a spine model generation method based on deep learning in an embodiment of the present application is shown; and

[0062] Figure 6 A schematic diagram of a storage medium according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work also fall within the scope of protection of the present application.

[0064] The terms used in this specification are those common terms currently widely used in the art in consideration of the functions of the present application, but these terms may vary according to the intention of a person of ordinary skill in the art, precedents, or new technologies in the art. In addition, specific terms may be selected by the applicant, and in this case, their detailed meanings will be described in the detailed description of the present application. Therefore, the terms used in the specification should not be understood as simple names, but rather as a general description based on the meaning of the terms and the present application.

[0065] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0066] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0067] Figure 1 A schematic diagram of the application architecture of a spine model generation method based on deep learning in an embodiment of the present application is shown, including a server 100 and a terminal device 200.

[0068] The terminal device 200 and the server 100 can be connected via the Internet to achieve mutual communication. Optionally, the above-mentioned Internet uses standard communication technology and / or protocol. The Internet is usually the Internet, but it can also be any network, including but not limited to any combination of a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a mobile, wired or wireless network, a dedicated network or a virtual private network. In some embodiments, the data exchanged through the network is represented by technologies and / or formats including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc. In addition, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), Internet Protocol Security (IPsec) can also be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can also be used to replace or supplement the above-mentioned data communication technologies.

[0069] The server 100 can provide various network services for the terminal device 200, wherein the server 100 can be a single server, a server cluster consisting of several servers, or a cloud computing center. Specifically, the server 100 may include a processor 110 (Center Processing Unit, CPU), a memory 120, an input device 130, and an output device 140, etc. The input device 130 may include a keyboard, a mouse, a touch screen, etc., and the output device 140 may include a display device, such as a liquid crystal display (Liquid Crystal Display, LCD), a cathode ray tube (Cathode Ray Tube, CRT), etc.

[0070] The memory 120 may include a read-only memory (ROM) and a random access memory (RAM), and provides the processor 110 with program instructions and data stored in the memory 120. In an embodiment of the present application, the memory 120 may be used to store the program of the spine model generation method based on deep learning in an embodiment of the present application.

[0071] The processor 110 calls the program instructions stored in the memory 120, and the processor 110 is used to execute the steps of any deep learning-based spine model generation method in the embodiments of the present application according to the obtained program instructions.

[0072] In addition, the application architecture diagram in the embodiment of the present application is intended to more clearly illustrate the technical solution in the embodiment of the present application, and does not constitute a limitation on the technical solution provided in the embodiment of the present application. Of course, for other application architectures and business applications, the technical solution provided in the embodiment of the present application is also applicable to similar problems.

[0073] The following is a non-restrictive explanation of the deep learning-based spinal model generation method provided according to at least one embodiment of the present application through several examples or embodiments. As described below, different features in these specific examples or embodiments can be combined with each other without conflicting with each other to obtain new examples or embodiments, and these new examples or embodiments also fall within the scope of protection of the present application.

[0074] In response to the above technical problems, in the technical solution of the present application, a deep learning-based spine model generation method is proposed, which can use a deep learning algorithm to analyze and capture features of multiple spinal local perspective CT images acquired at multiple angles, so as to extract key spinal local perspective state features from these CT images and realize the recognition and reconstruction of the spinal structure.

[0075] Specifically, the technical concept of the present application is to analyze these local spinal CT images by acquiring multiple spinal local perspective CT images from multiple angles, and introducing an image processing and analysis algorithm based on artificial intelligence and deep learning at the back end, so as to learn and capture the key state feature information of the local perspective of the spine, and to perform a panoramic-based semantic fusion of these important features on the spine state to reconstruct a more accurate spinal model. In this way, the intelligent recognition and reconstruction of the spinal model can be realized based on different local spinal perspective CT images, so as to assist doctors in diagnosing spinal diseases and formulating treatment plans.

[0076] Figure 2 The flowchart of the method for generating a spine model based on deep learning in the embodiment of the present application is shown. For example, the method for generating a spine model based on deep learning can be executed by a server, which can be Figure 1 The server 100 shown in FIG. Figure 2 As shown, according to the deep learning-based spine model generation method of the embodiment of the present application, the steps include: S510, acquiring multiple spine local perspective CT images collected from multiple angles; S520, performing spine state-based feature extraction on each of the multiple spine local perspective CT images to obtain multiple spine local perspective state feature maps; S530, passing each of the multiple spine local perspective state feature maps through a feature multi-scale perception enhancement module to obtain multiple enhanced spine local perspective state feature maps; S540, passing each of the multiple enhanced spine local perspective state feature maps through a spatial mask-based content differentiation module to obtain multiple spine local perspective foreground highlighting state feature maps; S550, inputting the multiple spine local perspective foreground highlighting state feature maps into a self-supervised feature sequence saliency fusion network to obtain a panoramic spine state fusion feature map; S560, generating a spine model based on the panoramic spine state fusion feature map to obtain a generated spine model.

[0077] More specifically, in the technical solution of the present application, first, a plurality of spinal local perspective CT images collected from multiple angles are obtained. Then, each of the plurality of spinal local perspective CT images is subjected to feature mining in a spinal state feature extractor based on a dilated convolutional neural network model, so as to extract the implicit long-distance correlation feature distribution information about the spinal local perspective state in each spinal local perspective CT image, thereby obtaining a plurality of spinal local perspective state feature maps.

[0078] Correspondingly, in step S520, feature extraction based on the spine state is performed on each of the multiple spine local view CT images to obtain a plurality of spine local view state feature maps, including: each of the multiple spine local view CT images is passed through a spine state feature extractor based on a void convolutional neural network model to obtain the multiple spine local view state feature maps.

[0079] Then, considering that in the actual process of characterizing and identifying the local perspective state features of the spine under different perspectives, the spinal structure may have different features and morphological presentations at different scales, and the spinal CT image contains rich multi-scale information of the spinal structure, such as the edges of the vertebrae, the texture of the intervertebral disc, etc. Therefore, in order to be able to capture the spinal structure and state features more comprehensively and accurately, so as to better characterize the local perspective state of the spine in order to perform subsequent spinal model generation tasks, in the technical solution of the present application, each of the multiple local perspective state feature maps of the spine is further processed through a feature multi-scale perception enhancement module to obtain multiple enhanced local perspective state feature maps of the spine. Through the processing of the feature multi-scale perception enhancement module, the local and global multi-scale state features under each local perspective of the spine can be perceived and fused respectively, so that the network can more comprehensively understand the structure and state information of the local perspective of the spine, and improve the model's perception of spinal structures of different sizes.

[0080] Correspondingly, in step S530, each of the multiple spinal local perspective state feature maps is respectively subjected to a feature multi-scale perception enhancement module to obtain a plurality of enhanced spinal local perspective state feature maps, including: in a first branch, performing point convolution processing on the spinal local perspective state feature map to obtain a first spinal local perspective state channel compression feature map; performing global mean pooling on the first spinal local perspective state channel compression feature map to obtain a first spinal local perspective state channel compression feature vector; performing nonlinear activation processing on the first spinal local perspective state channel compression feature vector to obtain a first spinal local perspective state channel local activation feature vector; in a second branch, performing point convolution processing on the spinal local perspective state feature map to obtain a second spinal local perspective state channel compression feature map; in a third branch, performing atrous convolution encoding on the spinal local perspective state feature map to obtain a spinal local perspective state receptive field expansion feature map; performing point convolution processing on the spinal local perspective state receptive field expansion feature map to obtain a spinal local perspective state global feature matrix with expanded receptive field; The global feature matrix of the expanded local perspective state of the spine with receptive field is nonlinearly activated to obtain the global activation feature matrix of the expanded local perspective state of the spine with receptive field; the global activation feature matrix of the expanded local perspective state of the spine with receptive field is point-multiplied by position with each feature matrix along the channel dimension of the channel compression feature map of the second local perspective state of the spine to obtain the global activation feature map of the second channel compressed local perspective state of the spine; each position feature value in the channel local activation feature vector of the first local perspective state of the spine is used as a weighted weight to weight each feature matrix along the channel dimension of the channel compression feature map of the second local perspective state of the spine to obtain the channel compression local activation feature map of the second local perspective state of the spine; the global activation feature map of the second channel compressed local perspective state of the spine and the channel compression local activation feature map of the second local perspective state of the spine are added by position to obtain the multi-scale fusion activation feature map of the second channel compressed local perspective state of the spine; the multi-scale fusion activation feature map of the second channel compressed local perspective state of the spine is subjected to void convolution encoding to obtain the enhanced local perspective state feature map of the spine.

[0081] In a specific example, each of the multiple spinal local perspective state feature maps is respectively processed through a feature multi-scale perception enhancement module to obtain a plurality of enhanced spinal local perspective state feature maps, including: each of the multiple spinal local perspective state feature maps is respectively processed through the feature multi-scale perception enhancement module using the following multi-scale perception enhancement formula to obtain the multiple enhanced spinal local perspective state feature maps; wherein the multi-scale perception enhancement formula is:

[0082]

[0083] Wherein, F is the local visual state feature map of each spine, C 3,3 and C 3,2 is a 3×3 dilated convolution operation with dilation numbers of 3 and 2, C1 is a 1×1 convolution operation, Avg(·) represents the global mean pooling of each feature matrix along the channel dimension in the feature map, δ(·) is a nonlinear activation process, ⊙ represents the positional point multiplication, It represents the weighted multiplication processing of the feature map along the channel dimension using the feature vector, and F1 is each enhanced spinal column local perspective state feature map among the multiple enhanced spinal column local perspective state feature maps.

[0084] Further, since each of the multiple enhanced spinal local perspective state feature maps contains multi-scale state features of the spine under different local perspectives, some of these features have a high contribution to the recognition of the spinal structure and the subsequent spinal model generation task, while some are background interference and some feature information irrelevant to the subsequent task. That is to say, in the spinal CT image, in addition to the spine itself, there are many other tissues and structures. Based on this, in order to distinguish and highlight the structural features of the spine from other background features, so as to improve the model's recognition ability of the spinal structure and state, thereby achieving better spinal model reconstruction, in the technical solution of the present application, each of the multiple enhanced spinal local perspective state feature maps is further processed by a content distinction module based on a spatial mask to obtain multiple spinal local perspective foreground highlighting state feature maps. Through the processing of the content distinction module based on the spatial mask, the key semantic features related to the spinal structure in the spinal local perspective state features under each perspective can be automatically learned and captured, so as to help the model distinguish the foreground and background in the image, especially for foreground areas such as the spine, by enhancing the foreground features, the model can pay more attention to the spinal structure itself. This processing approach can help the model focus more on the spinal structure and state semantics that are most important for the task of spinal model reconstruction.

[0085] Accordingly, in step S540, if Figure 3As shown, each of the multiple enhanced spinal local perspective state feature maps is respectively subjected to a content differentiation module based on a spatial mask to obtain a plurality of spinal local perspective foreground highlighting state feature maps, including: S541, using the negative number of each position feature value of the enhanced spinal local perspective state feature map as the exponent of the natural constant to calculate the position-based exponential function value with the natural constant as the base to obtain a spinal local perspective state class support feature map; S542, calculating the sum of each position feature value in the spinal local perspective state class support feature map and a constant one to obtain a spinal local perspective state modulation feature map; S543, dividing the spinal local perspective state class support feature map by the spinal local perspective state modulation feature map by position to obtain a spinal local perspective state representation feature map; S544, masking the spinal local perspective state representation feature map to obtain a spinal local perspective state mask weight feature map; S545, point multiplying the spinal local perspective state mask weight feature map and the spinal local perspective state feature map by position to obtain the spinal local perspective foreground highlighting state feature map.

[0086] Among them, in step S544, the spinal local perspective state characterization feature map is masked to obtain a spinal local perspective state mask weight feature map, including: setting the feature values ​​greater than or equal to a predetermined threshold in each position of the spinal local perspective state characterization feature map to one, and setting the rest to zero to obtain the spinal local perspective state mask weight feature map.

[0087] In a specific example, each of the multiple enhanced spinal local perspective state feature maps is processed through a content differentiation module based on a spatial mask to obtain multiple spinal local perspective foreground highlighting state feature maps, including: each of the multiple enhanced spinal local perspective state feature maps is processed through the content differentiation module based on a spatial mask with the following content differentiation enhancement formula to obtain the multiple spinal local perspective foreground highlighting state feature maps; wherein the content differentiation enhancement formula is:

[0088]

[0089] F1′=F s ⊙F1

[0090] Wherein, F1(i,j,k) represents the feature value of the (i,j,k)th position of the enhanced spine local view state feature map, mask(·) represents masking processing, exp(·) represents an exponential function with the natural constant e as the base, and F s(i, j, k) represents the feature value of the (i, j, k)th position of the spine local view state mask weight feature map, ε is a hyperparameter, F s represents the spine local perspective state mask weight feature map, F1 represents the enhanced spine local perspective state feature map, ⊙ is the position point multiplication, and F1′ is the spine local perspective foreground highlighting state feature map.

[0091] It should be understood that each of the multiple spinal local perspective foreground highlighting state feature maps contains information about the spinal local perspective state feature after the foreground feature highlighting expression under different local perspectives, and these different local perspective spinal states have panoramic-based correlation relationships and associated features. Analyzing this panoramic spinal state and structure helps to better understand the overall structure of the spine, which is conducive to the generation of a spinal model. Based on this, in the technical solution of the present application, the multiple spinal local perspective foreground highlighting state feature maps are further input into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spinal state fusion feature map. Through the processing based on the self-supervised feature sequence significant fusion network, the correlation relationship and implicit interaction information between the spinal state features under each local perspective can be automatically learned, which helps to identify which local perspectives have important spinal foreground highlighting state semantics for subsequent spinal model generation and reconstruction, and which spinal local perspective state features have a smaller contribution. In this way, in the process of fusing the semantic features of the local perspective foreground highlighting state of the spine under different perspectives, the features with different contributions to the subsequent spine model generation task can be adaptively weighted and fused, so as to better fuse the spine structure feature representations under different angles, so that the network can more comprehensively understand the overall spine structure and features. This helps to improve the model's ability to understand the spine structure, enhance the richness and diversity of feature expression, and thus improve the accuracy and robustness of spine model generation.

[0092] Accordingly, in step S550, the multiple spinal local perspective foreground highlighting state feature maps are input into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spinal state fusion feature map, including: performing maximum-based global pooling processing, random-value-based global pooling processing and average-based global pooling processing on each feature matrix along the channel dimension of each spinal local perspective foreground highlighting state feature map in the multiple spinal local perspective foreground highlighting state feature maps to obtain multiple spinal local perspective foreground highlighting state feature global maximum pooling feature vectors, multiple spinal local perspective foreground highlighting state feature global mean pooling feature vectors and multiple spinal local perspective foreground highlighting state feature global random value pooling feature vectors; calculating the spinal local perspective foreground highlighting state feature global maximum pooling feature vectors, spinal local perspective foreground highlighting state feature global mean pooling feature vectors and spinal local perspective foreground highlighting state feature global random value pooling feature vectors among the multiple spinal local perspective foreground highlighting state feature global maximum pooling feature vectors, spinal local perspective foreground highlighting state feature global mean pooling feature vectors and spinal local perspective foreground highlighting state feature global maximum pooling feature vectors, spinal local perspective foreground highlighting state feature global mean pooling feature vectors and spinal local perspective foreground highlighting state feature global random value pooling feature vectors. The position-weighted sum of the global random value pooling feature vectors of the local perspective foreground salient state feature is obtained to obtain multiple multi-scale pooling representation vectors of the local perspective foreground salient state feature of the spine; the mean vector of the multiple multi-scale pooling representation vectors of the local perspective foreground salient state feature of the spine is calculated as the center of the local perspective foreground state cluster of the spine to obtain the semantic feature vector of the center of the local perspective foreground state cluster of the spine; the semantic similarity score between each multi-scale pooling representation vector of the local perspective foreground salient state feature of the spine in the multiple multi-scale pooling representation vectors of the local perspective foreground state feature of the spine and the semantic feature vector of the center of the local perspective foreground state cluster of the spine is calculated to obtain a sequence of semantic similarity score values ​​of the center of the local perspective foreground state cluster of the spine; the sequence of semantic similarity score values ​​of the center of the local perspective foreground state cluster of the spine is normalized to obtain a sequence of semantic similarity attention weight coefficients of the center of the local perspective foreground state cluster of the spine; the sequence of semantic similarity attention weight coefficients of the center of the local perspective foreground state cluster of the spine is used as the weight to calculate the position-weighted sum of the multiple local perspective foreground salient state feature maps of the spine to obtain the panoramic spine state fusion feature map.

[0093] Among them, the semantic similarity score between each spinal local perspective foreground salient state feature multi-scale pooling representation vector of the multiple spinal local perspective foreground salient state feature multi-scale pooling representation vectors and the spinal local perspective foreground state cluster center semantic feature vector is calculated to obtain a sequence of spinal local perspective foreground state cluster center semantic similarity score values, including: respectively calculating the matrix product between each weight coefficient matrix and the corresponding spinal local perspective foreground salient state feature multi-scale pooling representation vector to obtain a sequence of spinal local perspective foreground salient state weighted feature vectors; each spinal local perspective foreground salient state feature multi-scale pooling representation vector in the sequence of spinal local perspective foreground salient state weighted feature vectors The weighted feature vector of the foreground highlighting state is added with the corresponding bias vector to obtain a sequence of weighted bias-adjusted feature vectors of the foreground highlighting state of the local perspective of the spine; the sequence of weighted bias-adjusted feature vectors of the foreground highlighting state of the local perspective of the spine is input into the sigmoid function for activation processing to obtain a sequence of activated feature vectors of the foreground highlighting state of the local perspective of the spine; the product of each activated feature vector of the foreground highlighting state of the local perspective of the spine in the sequence of activated feature vectors of the foreground highlighting state of the local perspective of the spine and the transposed vector of the semantic feature vector of the center of the foreground state of the local perspective of the spine cluster is calculated respectively to obtain a sequence of semantic similarity score values ​​of the center of the foreground state of the local perspective of the spine cluster.

[0094] In a specific example, the plurality of spine local perspective foreground highlighting state feature maps are input into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spine state fusion feature map, including: inputting the plurality of spine local perspective foreground highlighting state feature maps into the significant fusion network based on a self-supervised feature sequence to process the panoramic spine state fusion feature map using the following significant fusion formula; wherein the significant fusion formula is:

[0095] v p =α·AvgPool(F1′ (p) )+β·MaxPool(F1′ (p) )+γ·StoPool(F1′ (p) )

[0096]

[0097] Among them, F1′ (p) is the pth spine local perspective foreground highlighting state feature map among the multiple spine local perspective foreground highlighting state feature maps, AvgPool(·), MaxPool(·) and StoPool(·) are respectively the global mean pooling, maximum value pooling and random value pooling of each feature matrix along the channel dimension of the feature map, α, β and γ are all preset trainable weight values, v pis the multi-scale pooling representation vector of the spine local perspective foreground highlighting state corresponding to the p-th spine local perspective foreground highlighting state feature map, N is the number of feature maps in the multiple spine local perspective foreground highlighting state feature maps, v r is the semantic feature vector of the spine local view foreground state cluster center, W p represents the weight coefficient matrix, B p is the bias vector, sigmoid(·) represents the sigmoid function, e p is the semantic similarity score between the pth multi-scale pooled representation vector of the spine local perspective foreground highlight state and the central semantic feature vector of the spine local perspective foreground state cluster, softmax(·) represents the softmax function, and F′ is the panoramic spine state fusion feature map.

[0098] Then, the panoramic spine state fusion feature map is input into the spine model generator based on the diffusion model to obtain the generated spine model. In other words, the panoramic state feature fusion representation information of the spine state is used to generate the spine model to reconstruct a more accurate spine model. In this way, the intelligent recognition and reconstruction of the spine model can be realized based on different local spinal angle CT images to assist doctors in diagnosing spinal diseases and formulating treatment plans.

[0099] Accordingly, in step S560, a spinal model is generated based on the panoramic spinal state fusion feature map to obtain a generated spinal model, including: inputting the panoramic spinal state fusion feature map into a spinal model generator based on a diffusion model to obtain a generated spinal model.

[0100] Preferably, the panoramic spine state fusion feature map is input into a spine model generator based on a diffusion model to obtain a generated spine model, including: multiplying the panoramic spine state fusion feature map by the scale of the panoramic spine state fusion feature map and the square root of the scale of the panoramic spine state fusion feature map to obtain a panoramic spine state fusion full-width feature map and a panoramic spine state fusion half-width feature map, wherein the scale of the panoramic spine state fusion feature map is the width of the feature matrix of the panoramic spine state fusion feature map multiplied by the height and then multiplied by the number of channels of the panoramic spine state fusion feature map; performing point subtraction on the sum of the absolute values ​​of each eigenvalue of the panoramic spine state fusion feature map and the panoramic spine state fusion feature map, and calculating the square root of the absolute value of each position of the point subtraction result to obtain a panoramic spine state fusion full-width semantic change map; The spine state fusion half-width feature map and the square root of the sum of the squares of each eigenvalue of the panoramic spine state fusion feature map are point-subtracted, and the square root of the absolute value of each position of the point-subtraction result is calculated to obtain a panoramic spine state fusion half-width semantic change map; the logarithm with base 2 of each eigenvalue of the panoramic spine state fusion full-width semantic change map and the panoramic spine state fusion half-width semantic change map are respectively calculated to obtain a panoramic spine state fusion full-width semantic change information map and a panoramic spine state fusion half-width semantic change information map; the weighted sum of the panoramic spine state fusion full-width semantic change information map and the panoramic spine state fusion half-width semantic change information map is calculated with a balanced hyperparameter as a weight to obtain an optimized panoramic spine state fusion feature map; the optimized panoramic spine state fusion feature map is input into a spine model generator based on a diffusion model to obtain a generated spine model.

[0101] That is, in the technical solution of the present application, the step of inputting the panoramic spine state fusion feature map into the spine model generator based on the diffusion model to obtain the spine model generation step includes: optimizing the panoramic spine state fusion feature map with the following optimization formula to obtain the optimized panoramic spine state fusion feature map; wherein the optimization formula is:

[0102]

[0103] Among them, F is the panoramic spine state fusion feature map, L is the scale of the panoramic spine state fusion feature map, that is, the width of each feature matrix along the channel dimension of the panoramic spine state fusion feature map multiplied by the height and then multiplied by the number of channels of the panoramic spine state fusion feature map, ‖·‖1 is the norm of the feature map, ⊙ is the point multiplication by position, is the point-by-point subtraction, ‖·‖2 is the bi-norm of the feature map, ω is the hyperparameter, It is added by position, log represents the logarithmic function value with base 2, |·| ⊙1 / 2represents the square root of the absolute value of the feature value at each position in the calculated feature map, F′ is the optimized panoramic spine state fusion feature map; the optimized panoramic spine state fusion feature map is input into the spine model generator based on the diffusion model to obtain the generated spine model.

[0104] Here, multiple spinal local perspective state feature maps represent the image semantic features in the local image semantic space domain under the global image semantic space domain for each spinal local perspective CT image. When performing feature multi-scale perception enhancement and content differentiation based on spatial masks, the difference in image semantic feature distribution in the local image semantic space domain will cause differences in scale perception enhancement weights and spatial semantic association mask weights, resulting in the multiple spinal local perspective foreground highlighting state feature maps input based on the self-supervised feature sequence saliency fusion network to obtain the panoramic spine state fusion feature map. There will also be a problem of insufficient aggregation of image semantic feature fusion distribution, thereby affecting the generation efficiency of the spine model generator based on the diffusion model and the generation quality of the generated spine model.

[0105] Based on this, the applicant of the present application takes the panoramic spine state fusion feature map as a feature set, wherein the change semantics in units of the feature value of the panoramic spine state fusion feature map is expressed, in order to dynamically aggregate the semantic set composed of different change semantics of the panoramic spine state fusion feature map as a whole without ignoring the individual semantic change information, the individual features of the panoramic spine state fusion feature map and the aggregated scale of the panoramic spine state fusion feature map are collectively expressed as feature full amplitude and half amplitude, and the low-rank negative correlation of different dimensions of the overall semantics of the feature set of the panoramic spine state fusion feature map is used as the phase and scaled to dynamically adjust the change relationship of the semantic content of the panoramic spine state fusion feature map, thereby improving the overall semantic information expression aggregation of the feature set of the panoramic spine state fusion feature map, improving the generation efficiency of the panoramic spine state fusion feature map through the spine model generator based on the diffusion model, and improving the generation quality of the generated spine model. In this way, a more accurate spine model can be reconstructed based on different local perspective CT images of the spine to assist doctors in diagnosing spinal diseases and formulating treatment plans.

[0106] Based on the above embodiments, see Figure 4As shown, it is a structural diagram of a spinal model generation system 800 based on deep learning in an embodiment of the present application. The spine model generation system 800 based on deep learning includes: a multi-angle acquisition module 810, which is used to acquire multiple spine local perspective CT images acquired from multiple angles; a feature extraction module 820, which is used to perform feature extraction based on the spine state on each of the multiple spine local perspective CT images to obtain multiple spine local perspective state feature maps; an enhancement module 830, which is used to pass each of the multiple spine local perspective state feature maps through a feature multi-scale perception enhancement module to obtain multiple enhanced spine local perspective state feature maps; a distinction module 840, which is used to pass each of the multiple enhanced spine local perspective state feature maps through a content distinction module based on spatial masks to obtain multiple spine local perspective foreground highlighting state feature maps; a fusion module 850, which is used to input the multiple spine local perspective foreground highlighting state feature maps into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spine state fusion feature map; a spine model generation module 860, which is used to generate a spine model based on the panoramic spine state fusion feature map to obtain a generated spine model.

[0107] In one example, in the above-mentioned deep learning-based spine model generation system 800, the feature extraction module 820 is used to: respectively pass each of the multiple spinal local perspective CT images through a spinal state feature extractor based on a void convolutional neural network model to obtain the multiple spinal local perspective state feature maps.

[0108] Here, those skilled in the art will appreciate that the specific functions and operations of each module in the above-mentioned deep learning-based spine model generation system 800 have been described in detail above. Figure 2 to Figure 3 The description of the deep learning-based spine model generation method has been introduced in detail, and therefore, its repeated description will be omitted.

[0109] Figure 5 FIG. 1 is an application scenario diagram of a method for generating a spine model based on deep learning according to an embodiment of the present application. Figure 5 As shown, in this application scenario, first, multiple spinal local view CT images acquired from multiple angles are acquired (for example, Figure 5 Then, the plurality of spinal local view CT images are input to a server (for example, Figure 5 In S) shown in , the server is capable of using the deep learning-based spine model generation algorithm to process the multiple spine local perspective CT images to generate a spine model.

[0110] Based on the above embodiments, another exemplary embodiment of an electronic device is also provided in the embodiments of the present application. In some possible implementations, the electronic device in the embodiments of the present application may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the method for generating a spinal model based on deep learning in the above embodiments can be implemented when the processor executes the program.

[0111] For example, in the case of electronic equipment Figure 1 Taking the server 100 in the example as an example, the processor in the electronic device is the processor 110 in the server 100, and the memory in the electronic device is the memory 120 in the server 100.

[0112] An embodiment of the present application also provides a computer-readable storage medium. Figure 6 1 shows a schematic diagram of a computer-readable storage medium 1000 according to an embodiment of the present application. Figure 6 As shown, the computer-readable storage medium 1000 stores computer-executable instructions 1001. When the computer-executable instructions 1001 are executed by the processor, the method for generating a spine model based on deep learning according to the embodiment of the present application described with reference to the above figures can be executed. The computer-readable storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. The volatile memory may, for example, include a random access memory (RAM) and / or a cache memory (cache), etc. The non-volatile memory may, for example, include a read-only memory (ROM), a hard disk, a flash memory, etc.

[0113] The embodiment of the present application also provides a computer program product or a computer program, which includes computer executable instructions, and the computer executable instructions are stored in a computer readable storage medium. The processor of the computer device reads the computer executable instructions from the computer readable storage medium, and the processor executes the computer executable instructions, so that the computer device executes the spine model generation method based on deep learning according to the embodiment of the present application.

[0114] Those skilled in the art will appreciate that the contents disclosed in this application may be subject to various variations and improvements. For example, the various devices or components described above may be implemented by hardware, or by software, firmware, or a combination of some or all of the three.

[0115] In addition, although the present application makes various references to certain units in the system according to embodiments of the present application, any number of different units can be used and run on the client and / or server. The units are only illustrative, and different aspects of the system and method can use different units.

[0116] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present application is not limited to any particular form of combination of hardware and software.

[0117] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those of ordinary skill in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an idealized or extremely formal sense, unless explicitly defined as such herein.

[0118] The above is an explanation of the present application and should not be considered as a limitation thereof. Although several exemplary embodiments of the present application have been described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present application. Therefore, all of these modifications are intended to be included within the scope of the present application as defined by the claims. It should be understood that the above is an explanation of the present application and should not be considered to be limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present application is defined by the claims and their equivalents.

Claims

1. A method for generating a spine model based on deep learning, characterized in that: include: Acquire multiple spinal local view CT images collected from multiple angles; Performing feature extraction based on the spinal state on each of the plurality of spinal local view CT images respectively to obtain a plurality of spinal local view state feature maps; Each of the plurality of spinal local perspective state feature maps is respectively passed through a feature multi-scale perception enhancement module to obtain a plurality of enhanced spinal local perspective state feature maps; Each of the plurality of enhanced spinal local perspective state feature maps is respectively subjected to a content differentiation module based on a spatial mask to obtain a plurality of spinal local perspective foreground highlighting state feature maps; Inputting the plurality of spine local perspective foreground highlight state feature maps into a self-supervised feature sequence saliency fusion network to obtain a panoramic spine state fusion feature map; Generate a spinal model based on the panoramic spinal state fusion feature map to obtain a generated spinal model; Among them, each of the multiple spinal local perspective state feature maps is respectively passed through a feature multi-scale perception enhancement module to obtain multiple enhanced spinal local perspective state feature maps, including: In the first branch, point convolution processing is performed on the spinal local perspective state feature map to obtain a first spinal local perspective state channel compression feature map; global mean pooling is performed on the first spinal local perspective state channel compression feature map to obtain a first spinal local perspective state channel compression feature vector; nonlinear activation processing is performed on the first spinal local perspective state channel compression feature vector to obtain a first spinal local perspective state channel local activation feature vector; In the second branch, point convolution is performed on the spinal column local view state feature map to obtain a second spinal column local view state channel compression feature map; In the third branch, the spinal local perspective state feature map is subjected to hole convolution encoding to obtain a spinal local perspective state receptive field expansion feature map; the spinal local perspective state receptive field expansion feature map is subjected to point convolution processing to obtain a spinal local perspective state global feature matrix with receptive field expansion; the spinal local perspective state global feature matrix with receptive field expansion is subjected to nonlinear activation to obtain a spinal local perspective state global activation feature matrix with receptive field expansion; Multiply the receptive field expanded spine local perspective state global activation feature matrix by each feature matrix along the channel dimension of the second spine local perspective state channel compression feature map at each position point to obtain the second channel compression spine local perspective state global activation feature map; Taking each position feature value in the local activation feature vector of the first spine local perspective state channel as a weighted weight, weighting each feature matrix along the channel dimension of the second spine local perspective state channel compression feature map to obtain the second spine local perspective state channel compression local activation feature map; The second channel compressed spine local view state global activation feature map and the second spine local view state channel compressed local activation feature map are added according to the position to obtain the second channel compressed spine local view state multi-scale fusion activation feature map; The second channel compressed spine local viewing state multi-scale fusion activation feature map is subjected to hole convolution encoding to obtain the enhanced spine local viewing state feature map.

2. The method for generating a spine model based on deep learning according to claim 1, characterized in that: Performing feature extraction based on the spinal state on each of the plurality of spinal local view CT images to obtain a plurality of spinal local view state feature maps, including: Each of the multiple spinal local perspective CT images is passed through a spinal state feature extractor based on a void convolutional neural network model to obtain the multiple spinal local perspective state feature maps.

3. The method for generating a spine model based on deep learning according to claim 2, characterized in that: Each of the plurality of enhanced spinal local perspective state feature maps is respectively subjected to a content differentiation module based on a spatial mask to obtain a plurality of spinal local perspective foreground highlighting state feature maps, including: Using the negative number of the characteristic value of each position of the enhanced spinal local perspective state characteristic map as the exponent of the natural constant to calculate the exponential function value based on the natural constant according to the position to obtain the spinal local perspective state class support characteristic map; Calculate the sum of the characteristic values ​​of each position in the spinal local view state class support feature map and a constant one to obtain a spinal local view state modulation feature map; Dividing the spinal local view state class support feature map and the spinal local view state modulation feature map by position to obtain a spinal local view state representation feature map; Masking the spinal local view state representation feature map to obtain a spinal local view state mask weight feature map; The spine local perspective state mask weight feature map and the spine local perspective state feature map are multiplied by position points to obtain the spine local perspective foreground highlighting state feature map.

4. The method for generating a spine model based on deep learning according to claim 3, characterized in that: The spinal local view state characterization feature map is subjected to masking processing to obtain a spinal local view state mask weight feature map, including: The feature values ​​greater than or equal to a predetermined threshold in each position of the spinal local view state characterization feature map are set to one, and the rest are set to zero to obtain the spinal local view state mask weight feature map.

5. The method for generating a spine model based on deep learning according to claim 4, characterized in that: Inputting the plurality of spine local perspective foreground highlight state feature maps into a significant fusion network based on self-supervised feature sequence to obtain a panoramic spine state fusion feature map, including: Performing maximum-based global pooling processing, random-value-based global pooling processing, and average-based global pooling processing on each feature matrix along the channel dimension of each spinal local perspective foreground highlighting state feature map in the multiple spinal local perspective foreground highlighting state feature maps to obtain multiple spinal local perspective foreground highlighting state feature global maximum pooling feature vectors, multiple spinal local perspective foreground highlighting state feature global average pooling feature vectors, and multiple spinal local perspective foreground highlighting state feature global random value pooling feature vectors; Calculate the position-weighted sum of the global maximum pooling feature vectors of the multiple spinal local perspective foreground highlighting state features, the global mean pooling feature vectors of the multiple spinal local perspective foreground highlighting state features and the global random value pooling feature vectors of the multiple spinal local perspective foreground highlighting state features corresponding to each group, the global maximum pooling feature vector of the spinal local perspective foreground highlighting state features, the global mean pooling feature vector of the spinal local perspective foreground highlighting state features and the global random value pooling feature vector of the spinal local perspective foreground highlighting state features to obtain a multi-scale pooling representation vector of the multiple spinal local perspective foreground highlighting state features; Calculate the mean vector of the plurality of multi-scale pooled representation vectors of the local perspective foreground salient state features of the spine as the center of the local perspective foreground state cluster of the spine to obtain a semantic feature vector of the center of the local perspective foreground state cluster of the spine; Calculate the semantic similarity score between each of the plurality of multi-scale pooled representation vectors of the local perspective foreground salient state features of the spine and the central semantic feature vector of the local perspective foreground state cluster of the spine to obtain a sequence of central semantic similarity score values ​​of the local perspective foreground state cluster of the spine; Normalizing the sequence of the semantic similarity score values ​​of the spine local perspective foreground state cluster center to obtain a sequence of the semantic similarity attention weight coefficients of the spine local perspective foreground state cluster center; Using the sequence of semantic similarity attention weight coefficients of the center of the local perspective foreground state cluster of the spine as weights, the position-by-position weighted sum of the multiple local perspective foreground highlight state feature maps of the spine is calculated to obtain the panoramic spine state fusion feature map.

6. The method for generating a spine model based on deep learning according to claim 5, characterized in that: Calculating the semantic similarity score between each of the plurality of multi-scale pooled representation vectors of the local perspective foreground salient state features of the spine and the semantic feature vector of the center of the local perspective foreground state cluster of the spine to obtain a sequence of semantic similarity score values ​​of the center of the local perspective foreground state cluster of the spine, including: Calculate the matrix product between each weight coefficient matrix and the corresponding multi-scale pooling representation vector of the local perspective foreground salient state of the spine to obtain a sequence of weighted feature vectors of the local perspective foreground salient state of the spine; Adding a corresponding bias vector to each spinal local perspective foreground highlighting state weighted feature vector in the sequence of spinal local perspective foreground highlighting state weighted feature vectors to obtain a sequence of spinal local perspective foreground highlighting state weighted bias adjustment feature vectors; The sequence of the weighted bias-adjusted feature vectors of the local perspective foreground highlight state of the spine is input The function performs activation processing to obtain a sequence of activation feature vectors of the foreground highlighting state of the local perspective of the spine; The products of each spinal local perspective foreground highlighting state activation feature vector in the sequence of spinal local perspective foreground highlighting state activation feature vectors and the transposed vector of the spinal local perspective foreground state cluster center semantic feature vector are calculated respectively to obtain a sequence of spinal local perspective foreground state cluster center semantic similarity score values.

7. The method for generating a spine model based on deep learning according to claim 6, characterized in that: Generating a spinal model based on the panoramic spinal state fusion feature map to obtain a generated spinal model includes: The panoramic spine state fusion feature map is input into a spine model generator based on a diffusion model to obtain a generated spine model.

8. A spine model generation system based on deep learning, characterized in that: include: A multi-angle acquisition module is used to acquire multiple spinal local view CT images acquired from multiple angles; A feature extraction module, used for performing feature extraction based on the spinal state on each of the plurality of spinal local view CT images to obtain a plurality of spinal local view state feature maps; An enhancement module, used for respectively passing each of the plurality of spinal local perspective state feature maps through a feature multi-scale perception enhancement module to obtain a plurality of enhanced spinal local perspective state feature maps; A distinguishing module, used for respectively passing each of the plurality of enhanced spinal local perspective state feature maps through a content distinguishing module based on a spatial mask to obtain a plurality of spinal local perspective foreground highlighting state feature maps; A fusion module, used for inputting the plurality of spine local perspective foreground highlight state feature maps into a significant fusion network based on a self-supervised feature sequence to obtain a panoramic spine state fusion feature map; A spine model generation module, used for generating a spine model based on the panoramic spine state fusion feature map to obtain a generated spine model; Among them, each of the multiple spinal local perspective state feature maps is respectively passed through a feature multi-scale perception enhancement module to obtain multiple enhanced spinal local perspective state feature maps, including: In the first branch, point convolution processing is performed on the spinal local perspective state feature map to obtain a first spinal local perspective state channel compression feature map; global mean pooling is performed on the first spinal local perspective state channel compression feature map to obtain a first spinal local perspective state channel compression feature vector; nonlinear activation processing is performed on the first spinal local perspective state channel compression feature vector to obtain a first spinal local perspective state channel local activation feature vector; In the second branch, point convolution is performed on the spinal column local view state feature map to obtain a second spinal column local view state channel compression feature map; In the third branch, the spinal local perspective state feature map is subjected to hole convolution encoding to obtain a spinal local perspective state receptive field expansion feature map; the spinal local perspective state receptive field expansion feature map is subjected to point convolution processing to obtain a spinal local perspective state global feature matrix with receptive field expansion; the spinal local perspective state global feature matrix with receptive field expansion is subjected to nonlinear activation to obtain a spinal local perspective state global activation feature matrix with receptive field expansion; Multiply the receptive field expanded spine local perspective state global activation feature matrix by each feature matrix along the channel dimension of the second spine local perspective state channel compression feature map at each position point to obtain the second channel compression spine local perspective state global activation feature map; Taking each position feature value in the local activation feature vector of the first spine local perspective state channel as a weighted weight, weighting each feature matrix along the channel dimension of the second spine local perspective state channel compression feature map to obtain the second spine local perspective state channel compression local activation feature map; The second channel compressed spine local view state global activation feature map and the second spine local view state channel compressed local activation feature map are added according to the position to obtain the second channel compressed spine local view state multi-scale fusion activation feature map; The second channel compressed spine local viewing state multi-scale fusion activation feature map is subjected to hole convolution encoding to obtain the enhanced spine local viewing state feature map.

9. The deep learning-based spine model generation system according to claim 8, characterized in that: The feature extraction module is used to: Each of the multiple spinal local perspective CT images is passed through a spinal state feature extractor based on a void convolutional neural network model to obtain the multiple spinal local perspective state feature maps.

Images