A spine vertebrae segmentation method based on frequency enhanced multi-granularity context

Abstract

The application discloses a spine vertebra segmentation method based on frequency enhanced multi-granularity context, and belongs to the fields of deep learning, computer vision and image segmentation. The method comprises the following steps: firstly, collecting data and performing division to obtain image data of spine vertebra CT and MRI in a three-dimensional imaging mode, and dividing the image data into a training set and a test set; secondly, performing preprocessing, training and post-processing on the image data; thirdly, training a network model, training a frequency enhanced multi-granularity context segmentation model on the training set to obtain a segmentation model; and finally, testing the segmentation model effect. The application can avoid information loss by introducing a wavelet transform downsampling operation, and can enhance low-frequency and high-frequency features respectively, so that the accuracy and detail performance of a segmentation result can be improved, and important technical support can be provided for precision medicine. The application can more effectively fuse multi-granularity context information, retain image details and improve segmentation precision, and is suitable for various clinical application scenarios.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of deep learning, computer vision and image segmentation, and relates to a spine vertebra segmentation method based on frequency enhancement multi-granularity context. BACKGROUND

[0002] Medical image segmentation plays a crucial role in modern medicine, and with the advancement of imaging technology and deep learning, it has achieved efficient and accurate analysis of complex medical images. By accurately identifying and extracting lesion areas, it not only improves the accuracy of disease diagnosis and supports the development of personalized treatment plans, but also promotes medical research and teaching. High-quality segmentation results help monitor disease progression in real time and adjust treatment strategies, thereby improving patient prognosis and quality of life. Therefore, medical image segmentation is not only a key bridge connecting imaging data and clinical applications, but also an important driving force for the development of precision medicine.

[0003] In the field of orthopedics, the spine, as the second lifeline of the human body, supports the normal operation of various organs and is closely connected to the whole body through the central nervous system within it. The development of modern digital orthopedics relies on accurate segmentation technology of three-dimensional medical image data, which is crucial in clinical applications such as disease diagnosis, preoperative planning and surgical navigation. Existing vertebra segmentation methods usually use an encoder-decoder structure and have achieved positive results. However, traditional methods often overlook the key impact of downsampling techniques. Max pooling and bilinear interpolation are commonly used downsampling methods, but they often result in a large loss of fine-grained details during downsampling. This loss of fine-grained details is particularly harmful to blurred images, as preserving fine-grained details is crucial for accurate segmentation. In addition, vertebra segmentation requires sufficient context information to distinguish different vertebrae. Existing convolutional neural networks (CNNs) rely on local convolution kernels to extract features, which limits their ability to model global and geometric features. Although Transformer-based methods perform well in capturing long-range dependencies, their self-attention mechanism has high computational complexity. Recently, state-space models (SSMs) have captured long-range dependencies with linear complexity, providing a promising solution to the limitations of CNNs and Transformers. However, SSMs lack consideration of spatially varying context, which reduces their sensitivity to multi-scale context information of similar vertebrae, thereby weakening their ability to distinguish vertebrae.

[0004] Therefore, developing an efficient and accurate vertebra segmentation method that can comprehensively address these challenges remains a pressing problem, which is of great significance for improving medical quality and efficiency. SUMMARY

[0005] In view of the deficiencies of the above existing methods, the present application proposes a spine vertebra segmentation method based on frequency enhanced multi-granularity context, introduces wavelet transform downsampling to avoid information loss, and respectively enhances low-frequency and high-frequency features, thereby improving the performance of spine vertebra segmentation.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is:

[0007] A spine vertebra segmentation method based on frequency enhanced multi-granularity context, the spine vertebra segmentation method comprising the following steps:

[0008] Step 1: Collect data and divide, obtain the image data of spine vertebra CT and MRI in three-dimensional imaging mode, divide the image data into training set and test set;

[0009] Step 2: Standardized process construction of image data preprocessing, training and post-processing process:

[0010] The preprocessing, training and post-processing process of image data are fully integrated into the nnU-Net framework to establish an efficient and standardized training process, ensuring that each step from image data preparation to final model evaluation meets the optimal practice standard; specifically:

[0011] Step 2.1, in the data preprocessing stage, first standardize the image data obtained in step 1, including gray value normalization and resampling operation. Then, adjust the image size by cropping or padding to adapt to the network input size requirement.

[0012] Step 2.2, in the training process, use the automatic configuration function of nnU-Net framework to dynamically adjust hyperparameters such as learning rate, batch size and loss function type to adapt to the characteristics of different data sets. At the same time, introduce data augmentation operations such as random rotation, scaling, elastic deformation and intensity disturbance during training to increase data diversity and further improve the robustness and generalization ability of the model.

[0013] Step 2.3, in the post-processing stage, optimize the prediction results to reduce false positive regions and boundary burr phenomenon. The specific operation includes isolated point removal based on connected domain analysis and morphological closing operation to smooth the boundary.

[0014] Step three: training the network model, training the Frequency-enhanced Multi-granularity Context Module (FMCM) on the training set, observing the change of the loss function value during training, and terminating the training when the loss value converges, to obtain the trained segmentation model. The segmentation model is constructed by a Wavelet Transform Downsampling (WTD) and a Wavelet Transform Upsampling (WTU) to form an encoder-decoder structure, integrating a High-Frequency Feature Refinement (HFR) and a Multi-Granularity State Space Model (MG-SSM) to enhance the ability to capture different granularity features; in particular:

[0015] Step 3.1, in the encoding stage, first decompose the features extracted by the encoder into high-frequency components and low-frequency components through the Wavelet Transform Downsampling (WTD), and apply the High-Frequency Feature Refinement (HFR) and the Multi-Granularity State Space Model (MG-SSM) to generate enhanced feature representations, and then fuse these enhanced features to generate the final features, effectively capturing semantic and detail information. In particular:

[0016] Step 3.1.1, the Wavelet Transform Downsampling (WTD) based on discrete wavelet transform (DWT) is used to process the features extracted by the encoder, which can effectively alleviate the irreversible distortion problem in the traditional downsampling process and reduce information loss.

[0017] The features extracted in the first stage of the encoder are defined as , which is represented as:

[0018] (1)

[0019] wherein, is the low-frequency component; is the high-frequency component. In the discrete wavelet transform (DWT), by applying a low-pass filter ( ) and a high-pass filter ( ) to each axis of the feature , the 8 sub-band components are obtained, including a low-frequency component ( and 7 high-frequency components ( ​The resolution of these components is half of the input feature resolution of the current encoding stage, which realizes lossless down-sampling.

[0020] Step 3.1.2, the low-frequency component contains most of the basic information of the image, while the high-frequency component can capture complex details, textures and edge information. Therefore, feature enhancement processing is performed on the low-frequency component and the high-frequency component at the same time:

[0021] (2)

[0022] (3)

[0023] wherein, and represent the enhanced low-frequency feature and high-frequency feature, respectively, is a multi-granularity state space model MG-SSM, is a high-frequency feature enhancement module HFR;

[0024] Step 3.1.2.1, in the enhancement process of the low-frequency component, the multi-granularity state space model MG-SSM first uses the residual block to process the feature extracted by the encoder to capture the local detail information contained therein. Then, three parallel 3D depth separable convolutions with different dilation rates are used to capture multi-scale depth features with different receptive fields These multi-scale depth features are input into three visual state space modules VSSM of different scales respectively, to learn multi-granularity context information. The whole process is as follows:

[0025] (4)

[0026] (5)

[0027] wherein, denotes a multi-scale depth feature; denotes a dilation rate, including three different dilation rates; denotes a depth separable convolution; denotes a residual block; denotes an input low-frequency component; denotes a feature of the first branch of the visual state space module VSSM; denotes a spatial state model, denotes a SILU activation function; 1 denotes a convolution layer with a convolution kernel of 1; denotes a linear layer; Representation layer normalization operation. This process addresses the limitations of individual VSSM in capturing spatially varying features and provides multi-granularity context information to distinguish similar vertebrae.

[0028] Subsequently, the extracted multi-scale deep features are fused and input into the second branch of the VSSM, obtaining the features of the second branch :

[0029] (6)

[0030] wherein, represents the features of the second branch, which are shared among the three VSSMs; represents the concatenation operation.

[0031] The features of the two branches of the VSSM are combined using Hadamard product to obtain the final output:

[0032] (7)

[0033] wherein, represents the output of the VSSM with an expansion rate of ; represents the Hadamard product.

[0034] The output of each VSSM is obtained by formula (7) , and finally fused to obtain the enhanced low-frequency feature .

[0035] Step 3.1.2.2. In the enhancement process of high-frequency components, the high-frequency feature enhancement module HFR contains two paths, namely the key feature amplification branch and the noise filtering branch. Both paths use spatial attention mechanism to dynamically modulate different high-frequency components during the training process.

[0036] In the key feature amplification branch, the spatial attention mechanism is defined as:

[0037] (8)

[0038] (9)

[0039] wherein, represents the sigmoid operation, is the feature of the amplified high-frequency component; calculates the contribution of each high-frequency component to the spatial attention map; denotes a group max-pooling operation, which extracts key features from each component; The operation aggregates and amplifies the key features of each high-frequency component; is The convolutional layer of The feature dimension after the operation; denotes an element-wise multiplication operation.

[0040] The structure of the denoising path of the noise filtering branch is the same as that of the key feature amplification path, except that average pooling is used instead of max pooling:

[0041] (10)

[0042] (11)

[0043] wherein, denotes the output of the spatial attention mechanism; denotes the feature of the high-frequency component after denoising; denotes a group average pooling operation.

[0044] Step 3.1.2.3, spatial attention map is multiplied by each high-frequency component to highlight key features, while performs similar operations to filter out noise from each high-frequency component. Finally, the enhanced and denoised high-frequency components are spliced to obtain enhanced high-frequency features .

[0045] (12)

[0046] Step 3.1.3, fuse the enhanced and of step 3.1.2 to obtain features rich in semantic information and detail information:

[0047] (13)

[0048] Step 3.2, in the decoding stage, use the wavelet transform-based upsampling module WTU for feature fusion and resolution enhancement, fully utilize the information extracted by the encoder, and thus realize high-quality vertebra segmentation.

[0049] Specifically, the wavelet transform-based upsampling module WTU performs upsampling in the decoding stage while effectively utilizing the features of different scales extracted by the encoder:

[0050] (14)

[0051] (15)

[0052] wherein, represents the feature from the encoder of the th stage, represents the feature of the decoder of the th stage after passing through the up-sampling module based on wavelet transform WTU; represents the Inverse Discrete Wavelet Transform (IDWT); , represents the feature after decomposition by the Discrete Wavelet Transform (DWT); represents the feature of the decoder of the th stage.

[0053] During the entire step three processing, firstly, the down-sampling module based on wavelet transform WTD is applied in the encoder stage to realize lossless down-sampling. Then, the feature enhancement module, including the high-frequency feature enhancement module HFR and the multi-granularity state space model MG-SSM, is used to further improve the feature expression ability. Then, in the decoder stage, the image resolution is restored through the up-sampling module based on wavelet transform WTU, so as to obtain the final segmentation model.

[0054] Step four: test the segmentation model effect, input the CT or MR image data in the test set into the trained segmentation model, and then generate the segmentation image and calculate the related engineering index;

[0055] The engineering index includes: the Dice similarity coefficient (DSC) is used to evaluate the regional level similarity between the segmentation result obtained by the segmentation model and the true label, and the 95% Hausdorff distance (HD95) is used to measure the boundary accuracy.

[0056] Compared with the prior art, the beneficial effects of the present application are:

[0057] (1) The high-frequency feature enhancement module HFR and the multi-granularity state space model MG-SSM can effectively utilize different granularity context information and retain details, so as to obtain a segmentation result rich in details and accurate, thereby significantly improving the accuracy and reliability of the vertebral body segmentation. This method not only helps to improve the accuracy of diagnosis, supports the development of individualized plans, but also optimizes the surgical planning and postoperative evaluation process, greatly improves the treatment effect and rehabilitation process of patients.

[0058] (2) The present application introduces wavelet transform down-sampling to avoid information loss, and respectively enhances low-frequency and high-frequency features, thereby effectively solving the problems brought by image blur and high similarity of vertebral body structure. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is the overall structure schematic diagram of the present application.

[0060] Figure 2 is the multi-granularity state space model schematic diagram of the present application.

[0061] Figure 3 is the high-frequency feature enhancement module schematic diagram of the present application.

[0062] Figure 4 is the wavelet transform-based up-sampling module schematic diagram of the present application.

[0063] Figure 5 is the visualization graph on the Verse 2019 dataset.

[0064] Figure 6 is the visualization graph on the LUMBAR dataset. DETAILED DESCRIPTION

[0065] The specific embodiments of the present application are further described below in combination with the drawings and technical solutions.

[0066] The present application is implemented as follows:

[0067] Step one: collect data and divide it, obtain the image data of the CT and MRI of the vertebral body in the three-dimensional imaging mode, and divide the image data into a training set and a test set;

[0068] In this embodiment, first, medical image data for spinal vertebral body segmentation is needed. To ensure the diversity and comprehensiveness of the data, the present application mainly uses public authoritative data sets, including CT data set Verse 2019 and MRI data set. These data sets provide rich multi-modal data support for model training and verification. The Verse 2019 data set covers the segmentation of cervical, thoracic and lumbar vertebrae, providing comprehensive coverage of the spinal region; while the LUMBAR data set focuses on detailed segmentation of the lumbar region, enhancing the professionalism and meticulousness of the data set in specific areas. The combination of the two not only increases the completeness and diversity of the data, but also ensures that the model can effectively learn and verify under different anatomical structures and image conditions. The collected data needs to be divided into a training set and a test set for effective model training and evaluation.

[0069] Step two: standardization process construction of image data preprocessing, training and post-processing process:

[0070] The preprocessing, training and post-processing processes of image data are comprehensively integrated into the nnU-Net framework to establish an efficient and standardized training process, ensuring that each step from image data preparation to final model evaluation meets the optimal practice standard; in particular:

[0071] Step 2.1, in the data preprocessing stage, first standardize the image data obtained in step one, including gray value normalization and resampling operation. Then, adjust the image size by cropping to adapt to the network input size requirements.

[0072] Step 2.2, in the training process, use the automatic configuration function of nnU-Net framework to dynamically adjust the hyperparameters, including learning rate, batch size and loss function type, to adapt to the characteristics of different data sets. At the same time, introduce data enhancement operation during training, including random rotation, scaling, elastic deformation and intensity disturbance, increase data diversity, further improve the robustness and generalization ability of the model.

[0073] Step 2.3, in the post-processing stage, optimize the prediction results to reduce false positive regions and boundary burr phenomenon. Specific operations include isolated point removal based on connected component analysis, morphological closing operation to smooth the boundary.

[0074] Step three: train the network model.

[0075] After completing the operation of step two, the multi-granularity context efficient spine vertebra segmentation model FMCM based on frequency enhancement is trained using the training set. The model constructs an encoder-decoder structure through the wavelet transform-based downsampling module WTD and the wavelet transform-based upsampling module WTU, integrates the high-frequency feature enhancement module HFR and the multi-granularity state space model MG-SSM to enhance the ability to capture different granularity features; in particular:

[0076] Step 3.1, in the encoding stage, first decompose the features extracted by the encoder into high-frequency components and low-frequency components through the wavelet transform-based downsampling module WTD, and apply the high-frequency feature enhancement module HFR and the multi-granularity state space model MG-SSM to generate enhanced feature representations, and then fuse these enhanced features to generate the final features, effectively capturing semantic and detail information. In particular:

[0077] Step 3.1.1, the wavelet transform-based downsampling module WTD adopts discrete wavelet transform (Discrete Wavelet Transform, DWT) to process the features extracted by the encoder, which can effectively alleviate the irreversible distortion problem in the traditional downsampling process and reduce information loss.

[0078] The first encoder is defined as The features extracted in each stage are , which is expressed as:

[0079] (1)

[0080] in, is the low frequency component; is the high frequency component. In the discrete wavelet transform (DWT), by Apply a low-pass filter to each axis ( ) and high-pass filter ( ), the 8 sub-band components are obtained, including a low-frequency component ( and 7 high-frequency components ( , the resolution of these components is half of the input feature resolution in the current encoding stage, achieving lossless downsampling.

[0081] In step 3.1.2, low-frequency components contain most of the basic information of the image, while high-frequency components can capture complex details, textures, and edge information. Therefore, feature enhancement processing is performed on both low-frequency components and high-frequency components at the same time:

[0082] (2)

[0083] (3)

[0084] in, and Represent the enhanced low-frequency features and high-frequency features respectively, is the multi-granularity state space model MG-SSM, It is the high-frequency feature enhancement module HFR;

[0085] In step 3.1.2.1, during the enhancement of low-frequency components, the multi-granularity state space model MG-SSM first uses the residual block to process the features extracted by the encoder to capture the local detail information contained therein. Then, three parallel 3D depth-separable convolutions with different expansion rates are used to capture multi-scale depth features with different receptive fields. These multi-scale deep features are input into three visual state space modules (VSSMs) of different scales to learn multi-granularity context information. The whole process is as follows:

[0086] (4)

[0087] (5)

[0088] in, Represent multi-scale deep features; denotes the expansion rate, including three different expansion rates; denotes the depth separable convolution; denotes the residual block; denotes the input low-frequency component; denotes the feature of the first branch of the visual state space module VSSM; denotes the spatial state model, denotes the SILU activation function; 1 denotes the convolution layer with a kernel of 1; denotes the linear layer; denotes the layer normalization operation. This process solves the limitation of a single visual state space module VSSM in capturing spatially varying features, and provides multi-granularity context information to distinguish similar vertebrae.

[0089] Subsequently, the extracted multi-scale deep features are fused and input into the second branch of the VSSM to obtain the feature of the second branch:

[0090] (6)

[0091] wherein, denotes the feature of the second branch, which is shared among the three VSSMs; denotes the concatenation operation.

[0092] The features of the two branches of the VSSM are combined using the Hadamard product to obtain the final output:

[0093] (7)

[0094] wherein, denotes the output of the VSSM with an expansion rate of ; denotes the Hadamard product.

[0095] The output of each VSSM is obtained by formula (7), and finally fused to obtain the enhanced low-frequency feature .

[0096] Step 3.1.2.2. In the enhancement process of the high-frequency component, the high-frequency feature enhancement module HFR includes two paths, namely the key feature amplification branch and the noise filtering branch. Both paths use spatial attention mechanisms to dynamically modulate different high-frequency components during the training process.

[0097] In the key feature amplification branch, the spatial attention mechanism is defined as:

[0098] (8)

[0099] (9)

[0100] wherein, denotes a sigmoid operation, is the feature of the high frequency component after amplification; operation calculates the contribution of each high frequency component to the spatial attention map; denotes a group max pooling operation, which extracts key features from each component; operation aggregates and amplifies the key features of each high frequency component; is a convolutional layer for converting the feature dimension after the operation; denotes an element-wise multiplication operation.

[0101] The structure of the denoising path of the noise filtering branch is the same as that of the key feature amplification path, except that average pooling is used instead of max pooling:

[0102] (10)

[0103] (11)

[0104] wherein, denotes the output of the spatial attention mechanism; denotes the feature of the high frequency component after denoising; denotes a group average pooling operation.

[0105] Step 3.1.2.3, spatial attention map is multiplied by each high frequency component to highlight key features, while performs similar operations to filter out noise from each high frequency component. Finally, the enhanced and denoised high frequency components are spliced to obtain the enhanced high frequency feature .

[0106] (12)

[0107] Step 3.1.3, fuse the enhanced and of step 3.1.2 to obtain a feature rich in semantic information and detail information:

[0108] (13)

[0109] Step 3.2, in the decoding stage, feature fusion and resolution enhancement are performed using the wavelet transform-based upsampling module WTU, which makes full use of the information extracted by the encoder, thereby achieving high-quality vertebra segmentation.

[0110] Specifically, the wavelet transform-based upsampling module WTU effectively utilizes the features of different scales extracted by the encoder while upsampling in the decoding stage:

[0111] (14)

[0112] (15)

[0113] wherein, represents the feature from the encoder of the i-th stage, represents the feature of the decoder of the i-th stage after the wavelet transform-based upsampling module WTU. represents the inverse discrete wavelet transform (IDWT); , represents the feature after discrete wavelet transform (DWT) decomposition; represents the feature of the decoder of the i-th stage. During the entire step three processing, first, the wavelet transform-based downsampling module WTD is applied in the encoder stage to achieve lossless downsampling. Then, the feature enhancement module, including the high-frequency feature enhancement module HFR and the multi-granularity state space model MG-SSM, is used to further improve the feature expression capability. Then, in the decoder stage, the image resolution is restored through the wavelet transform-based upsampling module WTU, thereby obtaining the final segmentation model.

[0114] During this process, the network model is continuously optimized to minimize the loss function, thereby improving the quality of the segmentation result. As the training progresses, it is crucial to monitor the change of the loss function in real time. The loss value should gradually decrease, which indicates that the model is gradually learning and optimizing its parameters. By observing the convergence of the loss function, it is determined whether the training has reached the expected effect. When the loss value tends to be stable after multiple iterations, it means that the model has converged, and at this time it can be considered that the training process has reached the best state and the training can be terminated. After the training is completed, the best weights of the current model are saved for subsequent inference and testing in the step.

[0115] Step four: test the model effect.

[0116]

[0117] ​​​After model training is complete, the vertebral segmentation model is evaluated using the CT and MRI image data in the test set. The specific steps are as follows: (1) Loading saved weights: Load the best model weights saved during training. (2) Inputting test data: Input the CT and MRI images in the test set into the trained model to generate vertebral segmentation results. (3) Evaluating the segmentation effect: By calculating indicators such as the Dice similarity coefficient (DSC) and the 95% Hausdorff distance (HD95), the difference between the segmentation results output by the model and the true labels is compared to evaluate the model performance and compare with other methods for verification.

[0118] The current model achieved a DSC of 79.92% and an HD95 performance index of 3.04 on the Verse dataset, and a DSC of 79.13% and an HD95 of 17.13 on the LUMBAR dataset. Figure 5 and Figure 6 The visual segmentation results of the proposed model FMCM on the Verse2019 dataset and the LUMBAR dataset are shown respectively, and compared with other methods (such as nnU-Net, UMamba, MambaClnix, Segmamba and FMCM). Figure 5 The following figure shows the segmentation results of the thoracic, cervical, and lumbar regions in the Verse2019 dataset, from top to bottom. It can be seen that in areas with similar vertebrae, FMCM can more accurately distinguish adjacent vertebrae, demonstrating superior segmentation capabilities compared to other methods. Figure 6 The results show that FMCM has advantages in segmentation of the lumbar spine on the LUMBAR dataset. The results show that the model can achieve more accurate vertebral segmentation than existing methods.

[0119] The above-described embodiments merely express the implementation methods of the present invention, but should not be understood as limiting the scope of the present invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A method for vertebral segmentation based on frequency-enhanced multi-granularity context, characterized in that: The spinal vertebra segmentation method comprises the following steps: Step 1: Collect and divide data, obtain vertebral CT and MRI image data in 3D imaging mode, and divide the image data into training set and test set; Step 2: Preprocess, train and post-process the image data; Step 3: Train the network model. Train the frequency-enhanced multi-granularity contextual segmentation model (FMCM) on the training set, observe the changes in the loss function value during training, and terminate the training when the loss value converges to obtain a trained segmentation model. This segmentation model constructs an encoder-decoder structure by using a wavelet transform-based downsampling module (WTD) and a wavelet transform-based upsampling module (WTU). It integrates a high-frequency feature enhancement module (HFR) and a multi-granularity state space model (MG-SSM) to enhance the ability to capture features of different granularities. Step 3.1: In the encoding stage, the features extracted by the encoder are first decomposed into high-frequency components and low-frequency components through the wavelet transform-based downsampling module WTD. The high-frequency feature enhancement module HFR and the multi-granularity state space model MG-SSM are applied to generate enhanced feature representations. The enhanced features are fused to generate the final features, which effectively captures semantic and detail information. Step 3.2: In the decoding stage, the wavelet transform-based upsampling module (WTU) is used to perform feature fusion and resolution enhancement, making full use of the information extracted by the encoder to achieve high-quality vertebrae segmentation and obtain the final segmentation model. Step 4: Test the effectiveness of the segmentation model. Input the CT and MR image data in the test set into the trained segmentation model, generate segmented images, and calculate relevant engineering indicators.

2. The method for segmenting spinal vertebrae based on frequency-enhanced multi-granularity context according to claim 1, characterized in that: In step 2, the image data preprocessing, training, and post-processing processes are integrated into the nnU-Net framework. Specifically: Step 2.1: In the preprocessing stage, the image data obtained in step 1 is first standardized, including grayscale normalization and resampling operations; then, the image size is adjusted by cropping or padding to make it fit the network input size requirements; Step 2.2: During the training process, the automatic configuration function of the nnU-Net framework is used to dynamically adjust the hyperparameters to adapt to the characteristics of different datasets. At the same time, data augmentation operations are introduced during training to increase data diversity. Step 2.3: In the post-processing stage, the prediction results are optimized to reduce false positive areas and boundary burrs.

3. The method for segmenting spinal vertebrae based on frequency-enhanced multi-granularity context according to claim 1, characterized in that: The step 3.1 is specifically as follows: Step 3.1.1, the wavelet-based downsampling module WTD uses discrete wavelet transform DWT to process the features extracted by the encoder; define the encoder first The features extracted in each stage are , which is expressed as: （1）； in, is the low frequency component; is the high frequency component; In discrete wavelet transform DWT, by Apply low-pass filter and high-pass filter to each axis of the image to obtain 8 sub-band components, including one low-frequency component and 7 high-frequency components, to achieve lossless downsampling; Step 3.1.2: Perform feature enhancement on both low-frequency and high-frequency components to obtain enhanced and : （2）； （3）； in, and Represent the enhanced low-frequency features and high-frequency features respectively, is the multi-granularity state space model MG-SSM, It is the high-frequency feature enhancement module HFR; Step 3.1.3, the enhanced version of step 3.1.2 and Fusion is performed to obtain features rich in semantic information and detail information : F={ （13）。 4. The method for segmenting spinal vertebrae based on frequency-enhanced multi-granularity context according to claim 3, characterized in that: The step 3.1.2 is specifically as follows: Step 3.1.2.1, in the process of enhancing low-frequency components, the multi-granularity state space model MG-SSM first uses the residual block to process the features extracted by the encoder; then, three parallel 3D depth-separable convolutions with different expansion rates are used to capture multi-scale depth features with different receptive fields. ; The multi-scale deep features are input into three visual state space modules (VSSM) of different scales to learn multi-granularity context information; the whole process is as follows: （4）； （5）； in, Represent multi-scale deep features; Represents the expansion rate, including three different expansion rates; represents depth-wise separable convolution; represents the residual block; Represents the low-frequency component of the input; Represents the features of the first branch of the visual state space module VSSM; represents the spatial state model, Represents SILU activation function; 1 Represents a convolution layer with a convolution kernel of 1; represents a linear layer; Representation layer normalization operation; The extracted multi-scale deep features Fusion is performed and input into the second branch of VSSM to obtain the features of the second branch : （6）； in, represents the features of the second branch, which are shared among the three VSSMs; Indicates a connection operation; The features of the two branches of VSSM are combined using the Hadamard product to obtain the final output: （7）； in, The expansion rate is The output of VSSM; represents the Hadamard product; The output of each VSSM is obtained by formula (7): , and finally fused to obtain enhanced low-frequency features ; In step 3.1.2.2, during the enhancement of high-frequency components, the high-frequency feature enhancement module (HFR) consists of two paths: a key feature amplification branch and a noise filtering branch. Both paths use a spatial attention mechanism to dynamically modulate different high-frequency components during training. In the key feature amplification branch, the spatial attention mechanism Defined as: （8）； （9）； in, Represents the sigmoid operation, is the characteristic of the high-frequency component after amplification; The operation calculates the contribution of each high-frequency component to the spatial attention map; Represents the grouped maximum pooling operation, which is used to extract key features from each component; The operation aggregates and amplifies the key features of each high-frequency component; yes The convolutional layer is used to convert The feature dimension after the operation; Represents element-wise multiplication operation; The structure of the denoising path of the noise filtering branch is the same as the key feature amplification path, except that average pooling is used instead of maximum pooling: （10）； （11）； in, Represents the output of the spatial attention mechanism; Represented as the characteristics of high-frequency components after denoising; Represents the group average pooling operation; Step 3.1.2.3, spatial attention mechanism is multiplied by each high frequency component to enhance it, and Filter out the noise from each high-frequency component to perform denoising; finally, concatenate the enhanced and denoised high-frequency components to obtain the enhanced high-frequency features ; （12）。 5. The method for segmenting spinal vertebrae based on frequency-enhanced multi-granularity context according to claim 4, characterized in that: The step 3.2 is specifically as follows: The wavelet transform-based upsampling module WTU performs upsampling in the decoding stage and effectively utilizes features of different scales extracted by the encoder: （14）； （15）； in, Indicates that it comes from The characteristics of the encoder in the stage, It represents the first The characteristics of the decoder at each stage; represents the inverse discrete wavelet transform; , Represents the features after discrete wavelet transform DWT decomposition; Indicates the The characteristics of the decoder at this stage.

6. The method for segmenting spinal vertebrae based on frequency-enhanced multi-granularity context according to claim 1, characterized in that: The engineering metrics include: evaluating the region-level similarity between the segmentation results obtained by the segmentation model and the true label through the Dice similarity coefficient, and measuring the boundary accuracy using the 95% Hausdorff distance.

Images