Spine-pelvis joint segmentation method based on frequency-space cooperation and adaptive fusion

Abstract

The application discloses a spine-pelvis joint segmentation method based on frequency-space cooperation and adaptive fusion, and belongs to the technical field of medical image processing. A joint segmentation dataset containing spine and pelvis structures is constructed, and standardization preprocessing is completed by supplementing pelvis annotation to public data and combining private clinical data; a multi-scale three-dimensional segmentation network is built, a frequency-space cooperation feature modeling HFSM module is introduced in the coding stage, local details and global semantic features are extracted through joint extraction in the spatial domain and the frequency domain; a selective cross adaptive fusion SCAF module is connected between the encoder and the decoder, boundary enhancement, structure perception and gate weight modulation are performed on different levels of coding features and decoding features, adaptive feature aggregation is realized, the network adopts an end-to-end training mode, a weighted combination loss function is used to optimize model parameters, and the joint segmentation result of the spine and the pelvis is output. The application can effectively improve the segmentation accuracy and integrity of the connection region, complex edge and small structure of the spine-pelvis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion. Background Technology

[0002] Precise segmentation of the spinal-pelvic anatomy is the core foundation for surgical navigation, preoperative planning, and intraoperative registration in spinal surgery, directly determining the positioning accuracy and safety of procedures such as pedicle screw placement, corrective reduction, and decompression fusion. CT images can clearly present the vertebral bodies, appendages, and bony contours of the pelvis, providing a reliable basis for 3D modeling and spatial positioning. Clinical surgical navigation often requires a balance between global positioning and precise local matching. Relying solely on spinal segmentation can easily lead to segmental confusion, initial positioning drift, and local registration deviations. The pelvis, with its high rigidity, anatomical stability, and spatial proximity to the lumbosacral segment, can serve as a global reference anchor point. Most existing methods focus only on the segmentation of the spine alone, failing to incorporate the pelvis into a unified modeling framework. This makes it difficult to provide complete anatomical priors and cannot meet the practical needs of navigation systems for global-local collaborative registration.

[0003] Current academic research on spinal CT segmentation has developed into multiple technical approaches, and related achievements continue to drive clinical applications. Zheng Guoyan's team proposed Spine-Transformers, which uses 3DTransformer to achieve vertebral body labeling and segmentation in arbitrary field-of-view spinal CT, transforming vertebral body detection into an object prediction task and improving localization robustness in complex fields of view. Aqsa Dastgir et al. proposed Attention LinkNet-152, which achieves high segmentation accuracy on the VerSe dataset by improving the encoder and attention mechanism. Chao Yang et al. combined high-resolution segmentation of the main spine with Vision Transformer to achieve vertebral body variation recognition and fine segmentation. However, these methods mostly focus on the vertebral body as the core target, emphasizing the extraction of the spine's own structure. None of them consider the pelvis as a co-segmentation object, lack joint modeling in the frequency and spatial domains, and have insufficient representation capabilities for the spine-pelvis connection region, complex edges, and fine structures. Intramodal feature redundancy and cross-structural information discontinuity are prominent issues.

[0004] Chinese patent CN114359168A discloses a method for measuring and planning spine-pelvic force lines, establishing a three-dimensional force line relationship between the spine and pelvis for preoperative orthopedic angle planning. However, it only performs force line calculation rather than joint segmentation and does not involve image feature extraction or pixel-level segmentation modeling. Chinese patent CN120031800A proposes a three-dimensional spine image segmentation method based on deep learning, employing a two-stage coarse-fine segmentation strategy to improve vertebral contour accuracy. However, it only processes the single structure of the spine, completely excluding the pelvic region, and the segmentation network lacks frequency domain enhancement and cross-level adaptive fusion.

[0005] Existing literature and patents have improved the accuracy of spine segmentation, but they generally have four major shortcomings: lack of joint annotation of spine and pelvis and end-to-end joint segmentation modeling; extraction of features only in the spatial domain, with a lack of frequency domain detail enhancement leading to blurred boundaries; shallow encoder-decoder fusion method and unreasonable weight allocation of key structures; and lack of an integrated architecture for frequency-space collaboration and adaptive fusion. Summary of the Invention

[0006] To address the aforementioned technical issues, this application discloses a spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion, specifically including:

[0007] S1: Construct a joint segmentation dataset containing the spinal and pelvic structures;

[0008] S2: Construct a 3D segmentation network framework consisting of a multi-scale encoder, a multi-scale decoder, and a feature fusion module;

[0009] S3: Introduce the frequency-space co-modeling (HFSM) module in the network coding stage to complete feature extraction through the collaboration of the spatial and frequency domains;

[0010] S4: A Selective Cross-Adaptive Fusion (SCAF) module is introduced at the connection between the encoder and decoder to selectively aggregate and adaptively fuse encoded and decoded features;

[0011] S5: Input the constructed dataset into the network for end-to-end training, and the network outputs the joint segmentation results of the spine and pelvis.

[0012] Preferably, S1 specifically includes:

[0013] Based on spinal CT data, the 3D Slicer tool was used to manually supplement the annotations of the pelvic region where the original annotations were missing.

[0014] We introduced private spinal-pelvic CT data from the hospital's clinical practice, and professional physicians completed the complete manual annotation of the spinal vertebrae and pelvic structures.

[0015] Preprocessing operations were performed on all raw CT images, including resampling the images to a uniform voxel spacing, normalizing the gray values, and cropping the region of interest containing the spinal and pelvic structures.

[0016] The preprocessed samples are divided into training set, validation set and test set according to a preset ratio to form a standardized spine-pelvis joint segmentation dataset.

[0017] Preferably, in step S2, the three-dimensional segmentation network framework takes a three-dimensional CT image as input, assuming the input three-dimensional medical image is... Its dimensions are represented as: ,in, Input a 3D CT image; Input the number of channels; The depth dimension of the image; The height dimension of the image; The width dimension of the image;

[0018] The input image first undergoes initial feature mapping through the Patch Embedding layer, and then enters the four-level coding structure to extract features;

[0019] The encoding process is represented as follows: , ,in, For Patch Embedding operations; This is a downsampling operation; For the first Each encoding stage outputs features.

[0020] Preferably, in S2, the decoding stage adopts a structure symmetrical to the encoder, upsampling deep features layer by layer to restore resolution, and aggregating encoded and decoded features at different levels through the SCAF module;

[0021] The decoding process is represented as follows:

[0022]

[0023]

[0024]

[0025] in, This is an upsampling operation; Operations on residual blocks; Features fused to SCAF modules; For the first Each decoding stage outputs features; This is the final segmentation result.

[0026] Preferably, in S3, the HFSM module processes the input features. Processing:

[0027] The basic features are obtained by channel mapping through 1×1×1 convolution and layer normalization:

[0028]

[0029] in, Input features for the HFSM module; Characteristics of the initial linear transformation; For layer normalization operation; It is a 1×1×1 three-dimensional convolution;

[0030] Local spatial features are extracted using depthwise convolution and the GELU activation function, resulting in:

[0031]

[0032] in, Enhance features for the spatial domain; The activation function for the Gaussian error linear unit; It is a 3×3×3 three-dimensional depthwise convolution;

[0033] Performing a Fast Fourier Transform on the local spatial features yields the frequency domain representation:

[0034]

[0035] in, Frequency domain characteristics; For Fast Fourier Transform;

[0036] Decomposing the frequency domain features into real and imaginary parts yields:

[0037]

[0038]

[0039] in, The real part of the frequency domain is the characteristic; Features of the imaginary part in the frequency domain; To perform the real part operation; This is an operation to extract the imaginary part.

[0040] Preferably, the HFSM module continues to perform the following on the frequency domain features:

[0041] The real and imaginary parts are concatenated along the channel dimension, and then subjected to convolution, normalization, and activation to obtain the joint frequency domain features:

[0042]

[0043] in, This represents the cross-channel fusion characteristic in the frequency domain; It is a splicing of the real and imaginary features in the frequency domain;

[0044] By enhancing the real and imaginary parts separately using joint frequency domain features, we obtain:

[0045]

[0046]

[0047] in, To enhance the real part characteristics in the post-frequency domain; To enhance the imaginary part characteristics in the post-frequency domain; For channel attention modules;

[0048] Perform an inverse Fourier transform on the enhanced frequency domain features to obtain the enhanced frequency domain features in the spatial domain:

[0049]

[0050] in, This represents the frequency domain enhancement feature of the inverse Fourier transform output. This is the Fast Inverse Fourier Transform.

[0051] Preferably, the HFSM module simultaneously performs basic feature analysis. Perform spatial context modeling to obtain context features ;

[0052] The frequency domain enhancement features and contextual features are fused using residual blocks, then output through a 1×1×1 convolution, and residual connections are added:

[0053]

[0054] in, This is the final output feature of the HFSM module; Operations on residual blocks; Model features for spatial context.

[0055] Preferably, in step S4, the SCAF module performs encoding on the current level features of the encoder. deal with:

[0056] Boundary enhancement features are obtained through depthwise convolution, normalization, and activation:

[0057]

[0058] in, For encoder number Level output characteristics; Enhance features for encoding feature branches;

[0059] The adjacent deep-coded features are concatenated with the adjacent decoded feature channels to obtain the fused input:

[0060]

[0061] in, For cross-stage feature splicing; For encoder number Level output characteristics; For decoder number Level output characteristics;

[0062] By performing convolution, normalization, activation, and depthwise convolution on the concatenated features, we obtain structure-aware features:

[0063]

[0064] in, Enhance features for cross-stage feature branches.

[0065] Preferably, the SCAF module continues to perform feature interaction and weight modulation:

[0066] The boundary enhancement features and structure-aware features are multiplied element-wise to obtain the intermediate fused features:

[0067]

[0068] in, The feature interaction matrix; This is an element-wise multiplication operation;

[0069] Global pooling, 1×1×1 convolution, and sigmoid activation are applied to the decoder features to generate gated weights.

[0070]

[0071] in, The global attention weights for the decoder features; Use the Sigmoid activation function; This is a global average pooling operation;

[0072] The gating weights are added element-wise to the intermediate fused features to obtain the modulation features:

[0073]

[0074] in, This is the feature interaction matrix after fusion; This is an element-wise addition operation;

[0075] After convolution, normalization, and activation of the modulation features, the output is obtained by fusing them with the feature residuals of the current layer of the encoder.

[0076]

[0077] in, Output features for the SCAF module.

[0078] Preferably, in step S5, a joint loss function is constructed for end-to-end training:

[0079] The total loss function is a weighted combination of Dice loss and cross-entropy loss:

[0080]

[0081] in, Total loss; , These are the loss weighting coefficients; For Dice's loss; Cross-entropy loss;

[0082] The parameters of each module in the network are updated using the backpropagation algorithm, including the parameters of the HFSM module, SCAF module, and the convolutional, normalization, and activation layers of the encoder and decoder.

[0083] After training, the test set samples are input into the network, and the joint segmentation results of the spine and pelvic structure are output.

[0084] Compared with the prior art, the technical solution of this application has the following technical effects:

[0085] This invention constructs a spine-pelvis joint segmentation dataset that better meets the needs of surgical navigation applications by supplementing the pelvic region with annotations based on public datasets and combining it with private CT data from hospitals for joint training, thereby improving the model's ability to learn the overall anatomical structure of the spine-pelvis.

[0086] This invention breaks through the limitation of existing CT segmentation methods that mainly segment the spine itself, by incorporating the pelvic structure into the same segmentation framework. This provides more complete anatomical information for subsequent navigation registration, and the segmentation results have both global positioning significance and local fine registration value.

[0087] This invention introduces an HFSM module in the encoding stage and an SCAF module at the connection between the encoder and decoder, which can enhance the model's ability to represent features of edge regions, fine structures and complex connection regions, thereby improving the segmentation effect of spinal boundaries, complex pelvic contours and spine-pelvic connection regions, and improving the completeness and accuracy of the segmentation results.

[0088] The segmentation results obtained by this invention can simultaneously meet the requirements of global positioning and local fine matching. The pelvis provides global positioning reference, while the spine and single vertebral structures provide local structural constraints, thereby helping to improve the accuracy and robustness of subsequent real-time registration of single vertebral bodies and having strong engineering application value.

[0089] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0090] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0091] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0092] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0093] Figure 1 A schematic diagram of the overall process of the spine-pelvis joint segmentation method based on frequency-space collaboration and adaptive fusion;

[0094] Figure 2 A schematic diagram of the construction and preprocessing process of the spine-pelvis joint segmentation dataset;

[0095] Figure 3 , three Schematic diagram of the overall structure of the 3D segmentation network encoder and decoder;

[0096] Figure 4 A schematic diagram of the internal structure and operation process of the HFSM module for frequency-space collaborative feature modeling;

[0097] Figure 5 A schematic diagram of the feature calculation and fusion process of the Selective Cross-Adaptive Fusion (SCAF) module. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0099] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0100] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0101] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0102] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0103] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0104] Example 1

[0105] This embodiment mainly describes a spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion, such as... Figure 1 As shown, it specifically includes:

[0106] S1: Construct a joint segmentation dataset containing the spinal and pelvic structures;

[0107] S2: Construct a 3D segmentation network framework consisting of a multi-scale encoder, a multi-scale decoder, and a feature fusion module;

[0108] S3: Introduce the frequency-space co-modeling (HFSM) module in the network coding stage to complete feature extraction through the collaboration of the spatial and frequency domains;

[0109] S4: A Selective Cross-Adaptive Fusion (SCAF) module is introduced at the connection between the encoder and decoder to selectively aggregate and adaptively fuse encoded and decoded features;

[0110] S5: Input the constructed dataset into the network for end-to-end training, and the network outputs the joint segmentation results of the spine and pelvis.

[0111] Furthermore, such as Figure 2 As shown in S1, the public datasets VerSe2019 and VerSe2020 are selected as the basic data. The pelvic regions that are not included or are not fully included in the original annotations are manually annotated using 3D Slicer to form extended labels that include the spine and pelvic structure.

[0112] At the same time, private CT data from hospital clinical cases were introduced, and the spine and pelvic regions in the private data were manually labeled.

[0113] The publicly available data and privately labeled data after supplementary annotation are uniformly sorted and filtered. Then, the original image is resampled, grayscale normalized, and region of interest cropped to reduce the data differences between different devices, different acquisition parameters, and different cases.

[0114] The processed samples are divided into training, validation and test sets to construct a dataset for spine-pelvis joint segmentation.

[0115] Furthermore, such as Figure 3 As shown in S2, the network framework includes a multi-scale encoder, a multi-scale decoder, and a feature fusion module disposed between the encoder and the decoder.

[0116] Let the input three-dimensional medical image be... ,in, Indicates the number of input channels. , , These represent the depth, height, and width of the image, respectively.

[0117] During the encoding stage, the input image first undergoes initial feature mapping through the Patch Embedding layer, and then sequentially passes through four layers of feature extraction modules;

[0118] Each layer includes downsampling operations and an HFSM module. As the network deepens, the spatial resolution of the feature map decreases layer by layer, while the number of channels increases layer by layer, thereby achieving the gradual capture of multi-scale contextual information.

[0119] During the decoding stage, a decoding structure symmetrical to the encoder is used to upsample deep features layer by layer in order to gradually restore the spatial resolution of the feature map.

[0120] To fully integrate the shallow structural information of the encoder with the deep semantic information of the decoder, a SCAF module is introduced in each decoding layer to perform deep fusion of the corresponding layer's encoded features with the current decoded features, thereby improving the segmentation ability of boundary regions and fine structural regions.

[0121] The high-resolution features output by the decoder are convolutionally mapped to obtain the spine segmentation result.

[0122] Furthermore, the encoding process in S2 can be represented as:

[0123]

[0124]

[0125] in, This indicates the Patch Embedding operation. This indicates a downsampling operation. Indicates the first Each encoding stage outputs features.

[0126] Furthermore, the decoding process in S2 can be represented as:

[0127]

[0128]

[0129]

[0130] in, Indicates an upsampling operation. This indicates residual block operations. This indicates the features after fusion by the SCAF module. This indicates the final segmentation result.

[0131] Furthermore, such as Figure 4As shown, in S3, the HFSM module includes an input mapping branch, a local space modeling branch, a frequency domain enhancement branch, and a context modeling branch;

[0132] Let the characteristics input to the HFSM module be: First through Convolution and layer normalization perform channel mapping on the input features to obtain the basic features:

[0133]

[0134] Local spatial features are extracted using depthwise convolution and activation functions, resulting in:

[0135]

[0136] Furthermore, a fast Fourier transform is performed on the local spatial features to obtain a frequency domain representation:

[0137]

[0138] The frequency domain representation is then decomposed into a real part and an imaginary part:

[0139]

[0140]

[0141] The real and imaginary parts are concatenated along the channel dimension, and joint frequency domain features are extracted through convolutional mapping, layer normalization, and activation functions.

[0142]

[0143] By enhancing the real and imaginary parts using the joint frequency domain features, we obtain:

[0144]

[0145]

[0146] in, This indicates a channel recalibration operation.

[0147] Furthermore, an inverse fast Fourier transform is performed on the enhanced frequency domain features to restore them to the spatial domain, yielding the enhanced frequency domain features:

[0148]

[0149] At the same time, regarding basic features Perform context modeling to obtain context features:

[0150]

[0151] The frequency domain enhancement features are fused with the contextual features, and residual connections are used to preserve input information to obtain the output of the HFSM module:

[0152]

[0153] The HFSM module achieves feature enhancement for spinal boundaries, fine structures, and complex regions through the synergistic effect of spatial domain local modeling and frequency domain global enhancement.

[0154] Furthermore, such as Figure 5 As shown in S4, the SCAF module is used to achieve adaptive fusion between encoding features and decoding features at different levels;

[0155] Let the current layer encoder feature be The characteristics of adjacent deep encoders are Adjacent layer decoder features ;

[0156] The current layer encoder features are subjected to depthwise convolution, layer normalization, and activation mapping to obtain boundary enhancement features:

[0157]

[0158] The features of adjacent deep encoder layers and adjacent decoder layers are concatenated along the channel dimension to obtain the fused input:

[0159]

[0160] Furthermore, by extracting structure-aware features through convolutional mapping, normalized activation, and deep convolution, we obtain:

[0161]

[0162] Element-wise multiplication of boundary enhancement features and structure-aware features yields intermediate fused features:

[0163]

[0164] in, This indicates element-wise multiplication.

[0165] At the same time, global average pooling is performed on the features of adjacent decoder layers, and then... Convolution and the Sigmoid function generate gated weights:

[0166]

[0167] in, This indicates a global average pooling operation. This represents the Sigmoid activation function.

[0168] Furthermore, the gating weights are injected into the intermediate fusion features to obtain the modulated fusion features:

[0169]

[0170] in, This indicates element-wise addition.

[0171] The modulated fused features are convolutionally mapped and then residually fused with the features of the current layer encoder to obtain the output of the SCAF module:

[0172]

[0173] The SCAF module achieves effective fusion between shallow structural information of the encoder and deep semantic information of the decoder through the synergistic effect of boundary enhancement, structural coupling, and gating guidance, thereby improving the integrity of the spine segmentation boundary and the accuracy of the segmentation results.

[0174] This implementation achieves integrated modeling of the spine and pelvis by constructing a joint segmentation dataset. A frequency-space collaborative feature modeling module is employed to simultaneously extract spatial details and global frequency domain information, enhancing the representation of edges and fine structures. A selective cross-adaptive fusion module is used to efficiently aggregate encoded and decoded features, improving segmentation integrity and boundary accuracy. The output joint segmentation results provide stable global positioning and fine local constraints for surgical navigation, effectively improving registration accuracy and robustness, and possessing significant clinical application value.

[0175] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion, characterized in that, Includes the following steps: S1: Construct a joint segmentation dataset containing the spinal and pelvic structures; S2: Construct a 3D segmentation network framework consisting of a multi-scale encoder, a multi-scale decoder, and a feature fusion module; S3: Introduce the frequency-space co-modeling (HFSM) module in the network coding stage to complete feature extraction through the collaboration of the spatial and frequency domains; S4: A Selective Cross-Adaptive Fusion (SCAF) module is introduced at the connection between the encoder and decoder to selectively aggregate and adaptively fuse encoded and decoded features; S5: Input the constructed dataset into the network for end-to-end training, and the network outputs the joint segmentation results of the spine and pelvis.

2. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 1, characterized in that, S1 specifically includes: Based on spinal CT data, the 3D Slicer tool was used to manually supplement the annotations of the pelvic region where the original annotations were missing. We introduced private spinal-pelvic CT data from the hospital's clinical practice, and professional physicians completed the complete manual annotation of the spinal vertebrae and pelvic structures. Preprocessing operations were performed on all raw CT images, including resampling the images to a uniform voxel spacing, normalizing the gray values, and cropping the region of interest containing the spinal and pelvic structures. The preprocessed samples are divided into training set, validation set and test set according to a preset ratio to form a standardized spine-pelvis joint segmentation dataset.

3. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 1, characterized in that, In S2, the three-dimensional segmentation network framework takes a three-dimensional CT image as input. Let the input three-dimensional medical image be... Its dimensions are represented as: ,in, Input a 3D CT image; Input the number of channels; The depth dimension of the image; The height dimension of the image; The width dimension of the image; The input image first undergoes initial feature mapping through the Patch Embedding layer, and then enters the four-level coding structure to extract features; The encoding process is represented as follows: , ,in, For Patch Embedding operations; This is a downsampling operation; For the first Each encoding stage outputs features.

4. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 3, characterized in that, In S2, the decoding stage adopts a structure symmetrical to the encoder, upsampling deep features layer by layer to restore resolution, and aggregating encoded and decoded features at different levels through the SCAF module; The decoding process is represented as follows: in, This is an upsampling operation; Operations on residual blocks; Features fused to SCAF modules; For the first Each decoding stage outputs features; This is the final segmentation result.

5. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 4, characterized in that, In S3, the HFSM module processes the input features. Processing: The basic features are obtained by channel mapping through 1×1×1 convolution and layer normalization: in, Input features for the HFSM module; Characteristics of the initial linear transformation; For layer normalization operation; It is a 1×1×1 three-dimensional convolution; Local spatial features are extracted using depthwise convolution and the GELU activation function, resulting in: in, Enhance features for the spatial domain; The activation function for the Gaussian error linear unit; It is a 3×3×3 three-dimensional depthwise convolution; Performing a Fast Fourier Transform on the local spatial features yields the frequency domain representation: in, Frequency domain characteristics; For Fast Fourier Transform; Decomposing the frequency domain features into real and imaginary parts yields: in, The real part of the frequency domain is the characteristic; Features of the imaginary part in the frequency domain; To perform the real part operation; This is an operation to extract the imaginary part.

6. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 5, characterized in that, The HFSM module continues to execute on the frequency domain features: The real and imaginary parts are concatenated along the channel dimension, and then subjected to convolution, normalization, and activation to obtain the joint frequency domain features: in, This represents the cross-channel fusion characteristic in the frequency domain; It is a splicing of the real and imaginary features in the frequency domain; By enhancing the real and imaginary parts separately using joint frequency domain features, we obtain: in, To enhance the real part characteristics in the post-frequency domain; To enhance the imaginary part characteristics in the post-frequency domain; For channel attention modules; Perform an inverse Fourier transform on the enhanced frequency domain features to obtain the enhanced frequency domain features in the spatial domain: in, This represents the frequency domain enhancement feature of the inverse Fourier transform output. This is the Fast Inverse Fourier Transform.

7. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 6, characterized in that, The HFSM module simultaneously performs basic feature analysis. Perform spatial context modeling to obtain context features ; Frequency domain enhancement features and contextual features are fused using residual blocks, then output via 1×1×1 convolution, and residual connections are added: in, This refers to the final output characteristics of the HFSM module; Operations on residual blocks; Model features for spatial context.

8. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 7, characterized in that, In step S4, the SCAF module performs feature processing on the encoder at the current level. deal with: Boundary enhancement features are obtained through depthwise convolution, normalization, and activation: in, For encoder number Level output characteristics; Enhance features for encoding feature branches; The adjacent deep-coded features are concatenated with the adjacent decoded feature channels to obtain the fused input: in, For cross-stage feature splicing; For encoder number Level output characteristics; For decoder number Level output characteristics; By performing convolution, normalization, activation, and depthwise convolution on the concatenated features, we obtain structure-aware features: in, Enhance features for cross-stage feature branches.

9. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 8, characterized in that, The SCAF module continues to perform feature interaction and weight modulation: The boundary enhancement features and structure-aware features are multiplied element-wise to obtain the intermediate fused features: in, The feature interaction matrix; This is an element-wise multiplication operation; Global pooling, 1×1×1 convolution, and sigmoid activation are applied to the decoder features to generate gated weights. in, The global attention weights for the decoder features; Use the Sigmoid activation function; This is a global average pooling operation; The gating weights are added element-wise to the intermediate fused features to obtain the modulation features: in, This is the feature interaction matrix after fusion; This is an element-wise addition operation. Gating weights generated for decoder features; After convolution, normalization, and activation of the modulation features, the output is obtained by fusing them with the feature residuals of the current layer of the encoder. in, Output features for the SCAF module.

10. The spine-pelvis joint segmentation method based on frequency-space coordination and adaptive fusion according to claim 9, characterized in that, In S5, a joint loss function is constructed for end-to-end training: The total loss function is a weighted combination of Dice loss and cross-entropy loss: in, Total loss; , These are the loss weighting coefficients; For Dice's loss; Cross-entropy loss; The backpropagation algorithm is used to update the parameters of each module of the network, including the HFSM module, SCAF module, and the parameters of the convolutional layers, normalization layers and activation layers of the encoder and decoder; After training, the test set samples are input into the network, and the joint segmentation results of the spine and pelvic structure are output.

Images