Spine image segmentation method based on frequency domain enhancement and feature fusion

Abstract

The invention discloses a spinal image segmentation method based on frequency domain enhancement and feature fusion, and relates to the technical field of spinal image segmentation, and the method comprises the following steps: obtaining a to-be-segmented spinal image; inputting the spine image to be segmented into an encoder based on a frequency domain enhancement module, a residual error convolution module and a Transform module to obtain an encoding feature sequence; and inputting the coding feature sequence into a decoder based on an up-sampling block, a segmentation head module and a post-processing refining module to obtain a spine segmentation mask. According to the method, the encoder combined with the frequency domain enhancement module, the residual error convolution module and the Transformer module is constructed, the decoder combined with the up-sampling block, the segmentation head module and the post-processing refining module is adopted, and the segmentation precision of the vertebral body and the intervertebral disc of the spine image is improved through the synergistic effect of all the modules.

Description

Technical Field

[0001] This application relates to the field of spinal image segmentation technology, and in particular to a spinal image segmentation method based on frequency domain enhancement and feature fusion. Background Technology

[0002] The spine, as the central skeletal system of the human body, is composed of vertebral bodies, intervertebral discs, and other structures, undertaking the functions of support and movement. However, it is susceptible to pathological effects such as degeneration and injury, which can cause damage to the body. In modern society, with the increasing prevalence of diseases such as intervertebral disc herniation and spinal stenosis due to population aging, spinal MRI images have become the primary means of diagnosing spinal diseases due to their radiation-free nature and high soft tissue resolution. Accurate lesion identification, however, heavily relies on image segmentation techniques. Semantic segmentation provides an automated solution for this purpose, effectively reducing the burden on doctors and improving diagnostic efficiency and accuracy. Therefore, research on algorithms for spinal MRI image segmentation is of significant research importance.

[0003] In the prior art, Chinese patent CN117392677A discloses a semantic segmentation method for spinal tumor MRI images based on a multi-attention mechanism. The method includes acquiring an MRI image set containing spinal tumors; obtaining feature mapping layers {C1, C2, C3, C4, C5} of different scales of the MRI images from bottom to top according to the convolutional layers of the backbone network; obtaining M2, M3, M4, and M5 respectively based on C5, C4, C3, and C2, and using 3×3 convolutional kernels to obtain new feature maps P2, P3, P4, and P5; inputting C1 into a self-attention mechanism module for processing to obtain a sub-feature map X1; inputting P2, P3, P4, and P5 into channel attention for processing to obtain corresponding feature maps; inputting the corresponding feature maps into a spatial attention module for processing to obtain sub-feature maps X2, X3, X4, and X5; fusing X1 with X2, X3, X4, and X5 to obtain a fused feature map; and performing semantic segmentation on the fused feature map to obtain the final feature map.

[0004] However, spinal MRI image segmentation still faces challenges such as low contrast, pathological interference, blurred boundaries, and artifacts. The aforementioned existing technologies have not considered or solved these problems, resulting in poor segmentation accuracy of vertebral bodies and intervertebral discs in spinal images. Summary of the Invention

[0005] This application provides a spinal image segmentation method based on frequency domain enhancement and feature fusion to solve the problem of poor segmentation accuracy of vertebral bodies and intervertebral discs in existing spinal image segmentation techniques.

[0006] On the one hand, this application provides a spine image segmentation method based on frequency domain enhancement and feature fusion, including the following steps: Step 1: Obtain the image of the spine to be segmented.

[0007] Step 2: Input the spinal image to be segmented into an encoder based on a frequency domain enhancement module, a residual convolution module, and a Transformer module to obtain an encoded feature sequence.

[0008] Step 3: Input the encoded feature sequence into a decoder based on an upsampling block, a segmentation head module, and a post-processing refinement module to obtain a spine segmentation mask.

[0009] The frequency domain enhancement module is used to perform frequency domain enhancement on the spinal image to be segmented, thereby obtaining frequency domain enhanced features.

[0010] The residual convolution module is used to perform residual downsampling on the frequency domain enhanced features to obtain a deep feature map.

[0011] The Transformer module is used to perform self-attention processing on the deep feature map to obtain the encoded feature sequence.

[0012] The upsampling block is used to upsample and splice the encoded feature sequence to obtain channel features.

[0013] The segmentation head module is used to map the channel features to the number of segmentation categories to obtain a preliminary segmentation map.

[0014] The post-processing refinement module is used to perform boundary smoothing on the preliminary segmentation map to obtain a spine segmentation mask.

[0015] In one possible implementation, step two, the processing flow of the frequency domain enhancement module includes: The spinal image to be segmented is sequentially subjected to Gaussian blur convolution, Fourier transform, spectrum centering, high-pass filtering, inverse discrete Fourier transform, normalization, contrast enhancement, and weighted fusion to obtain frequency domain enhanced features.

[0016] In one possible implementation, in step two, the residual convolution module is constructed based on the ResNet50 backbone network.

[0017] In one possible implementation, in step two, the Transformer module performs self-attention processing on the deep feature map through a multi-head self-attention mechanism and a multilayer perceptron extension module to obtain an encoded feature sequence.

[0018] In one possible implementation, step three involves improving the upsampling block, and the processing flow of the improved upsampling block includes: The encoded feature sequence is sequentially subjected to feature preservation and upsampling, skip connection fusion, deep feature extraction, edge attention enhancement, and adaptive gated residual fusion to obtain channel features.

[0019] In one possible implementation, in step three, the segmentation head module employs a 3×3 convolution.

[0020] In one possible implementation, in step three, the post-processing refinement module employs a cascaded strategy of hierarchical morphological optimization and Gaussian smoothing.

[0021] The spine image segmentation method based on frequency domain enhancement and feature fusion in this application has the following advantages: An encoder combining a frequency domain enhancement module, a residual convolution module, and a Transformer module, and a decoder combining an upsampling block, a segmentation head module, and a post-processing refinement module are constructed. The frequency domain enhancement module enhances the spinal image to be segmented in the frequency domain, obtaining enhanced frequency domain features. The residual convolution module performs residual downsampling on the enhanced frequency domain features to obtain a deep feature map. The Transformer module performs self-attention processing on the deep feature map to obtain an encoded feature sequence. The upsampling block upsamples and concatenates the encoded feature sequence to obtain channel features. The segmentation head module maps the channel features to the number of segmentation categories to obtain a preliminary segmentation map. The post-processing refinement module smooths the boundaries of the preliminary segmentation map to obtain a spinal segmentation mask. The synergistic effect of these modules improves the segmentation accuracy of vertebral bodies and intervertebral discs in spinal images.

[0022] The proposed method involves sequentially performing Gaussian blur convolution, Fourier transform, spectral centering, high-pass filtering, inverse discrete Fourier transform, normalization, contrast enhancement, and weighted fusion on the spinal image to be segmented, thereby obtaining frequency domain enhancement features. The image is then converted to the frequency domain through Fourier transform, and combined with high- and low-frequency separation and dual noise reduction strategies, effectively strengthening edge features and enhancing semantic information expression.

[0023] The proposed method sequentially performs feature preservation and upsampling, skip connection fusion, deep feature extraction, edge attention enhancement, and adaptive gated residual fusion on the encoded feature sequence to obtain channel features. Dual-path feature fusion is achieved through dynamic edge attention maps, thereby better recovering detailed information.

[0024] The proposed post-processing refinement module employs a cascaded strategy of hierarchical morphological optimization and Gaussian smoothing to further optimize the smoothness and continuity of the segmentation boundaries. Attached Figure Description

[0025] 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating the spine image segmentation method based on frequency domain enhancement and feature fusion provided in this application embodiment; Figure 2 A schematic diagram of the spine image segmentation network architecture based on frequency domain enhancement and feature fusion constructed in steps two to three of the embodiments of this application; Figure 3 A schematic diagram of the processing flow of the frequency domain enhancement module provided in the embodiments of this application; Figure 4 A schematic diagram of the processing flow of the improved upsampling block provided in the embodiments of this application; Figure 5 A schematic diagram of the processing flow of the post-processing refinement module provided in the embodiments of this application; Figure 6 A schematic diagram showing a qualitative comparison between the spinal image segmentation method based on frequency domain enhancement and feature fusion provided in this application embodiment and five existing methods. Detailed Implementation

[0027] 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, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] like Figure 1 and Figure 2 As shown, this application provides a spine image segmentation method based on frequency domain enhancement and feature fusion, including the following steps: Step 1: Obtain the image of the spine to be segmented.

[0029] Step 2: Input the spinal image to be segmented into an encoder based on a frequency domain enhancement module, a residual convolution module, and a Transformer module to obtain an encoded feature sequence.

[0030] Step 3: Input the encoded feature sequence into a decoder based on an upsampling block, a segmentation head module, and a post-processing refinement module to obtain a spine segmentation mask.

[0031] The frequency domain enhancement module is used to perform frequency domain enhancement on the spinal image to be segmented, thereby obtaining frequency domain enhanced features.

[0032] The residual convolution module is used to perform residual downsampling on the frequency domain enhanced features to obtain a deep feature map.

[0033] The Transformer module is used to perform self-attention processing on the deep feature map to obtain the encoded feature sequence.

[0034] The upsampling block is used to upsample and splice the encoded feature sequence to obtain channel features.

[0035] The segmentation head module is used to map the channel features to the number of segmentation categories to obtain a preliminary segmentation map.

[0036] The post-processing refinement module is used to perform boundary smoothing on the preliminary segmentation map to obtain a spine segmentation mask.

[0037] Specifically, in this embodiment, the spinal image to be segmented is a spinal MRI image. In other possible embodiments, a spinal CT image or the like may also be used.

[0038] For example, in step two, the processing flow of the frequency domain enhancement module includes: The spinal image to be segmented is sequentially subjected to Gaussian blur convolution, Fourier transform, spectrum centering, high-pass filtering, inverse discrete Fourier transform, normalization, contrast enhancement, and weighted fusion to obtain frequency domain enhanced features.

[0039] Specifically, in this embodiment, the image is decomposed into two parts: a low-frequency smooth region and a high-frequency edge detail region. A Fast Fourier Transform (FFT), which has global analysis capabilities and can effectively capture edge information in any direction, is used to transform the image from the spatial domain to the frequency domain. High-pass filtering is performed in the frequency domain to preserve key boundary information, followed by inverse Fourier transform and dual noise reduction processing. Finally, the image is weighted and fused with the original image to achieve effective feature enhancement. The specific process is as follows: Figure 3 As shown.

[0040] Gaussian blurred convolution can be characterized by equation (1): (1) in, Indicates the position of the input grayscale image. pixel values, This indicates that the Gaussian convolution kernel is shifted. Weight value at the location, offset range , This indicates the blurred image after processing. The pixel value at that location needs to be normalized using the Gaussian kernel as follows: .

[0041] Subsequently, a Fourier transform is performed on the image to obtain its spectral representation, as shown in equation (2): (2) in, Represents spatial domain coordinates, Represents frequency domain coordinates, Represents frequency Complex values ​​at that location, and These represent the width and height of the image, respectively.

[0042] Next, a frequency domain centering operation is performed, shifting the zero-frequency component of the spectrum to the center position to obtain the center spectrum. As shown in equation (3): (3) Create a rectangular high-pass filter mask on the centralized spectrum. As shown in equation (4): (4) in, and These represent the horizontal and vertical coordinates of the center of the spectrum after centralization, respectively; and the vertical coordinate of the center of the spectrum is represented by . It is the cutoff radius. , This is the preset high-pass ratio; The value is 0 in the low-frequency region and 1 in the high-frequency region.

[0043] By applying this high-pass filter, the filtered, centered spectrum is obtained. As shown in equation (5): (5) Perform an inverse discrete Fourier transform (IDFT) on the filtered, centered spectrum to return it to the spatial domain, obtaining the edge response map. As shown in equation (6): (6) Finally, the edge response map is normalized and contrast-enhanced, and then weighted and fused with the original image to generate a feature-enhanced image, namely, frequency domain enhanced features.

[0044] For example, in step two, the residual convolution module is built based on the ResNet50 backbone network.

[0045] Specifically, in this embodiment, the residual convolution module is downsampled stepwise by the three-stage residual blocks of the ResNet50 backbone network to output a deep feature map.

[0046] For example, in step two, the Transformer module performs self-attention processing on the deep feature map through a multi-head self-attention mechanism and a multilayer perceptron extension module to obtain an encoded feature sequence.

[0047] Specifically, the deep feature maps output by the residual convolution module are first flattened into a sequence through a patch embedding operation before entering the Transformer module. The Transformer module performs self-attention processing on the flattened deep feature maps into a sequence through a multi-head self-attention mechanism and a multilayer perceptron extension module, enhancing the global contextual association of the features to obtain the encoded feature sequence.

[0048] like Figure 4 As shown, exemplarily, in step three, the upsampling block is improved, and the processing flow of the improved upsampling block includes: The encoded feature sequence is sequentially subjected to feature preservation and upsampling, skip connection fusion, deep feature extraction, edge attention enhancement, and adaptive gated residual fusion to obtain channel features.

[0049] Specifically, in this embodiment, feature preservation and upsampling include: preserving the input feature map. Used for subsequent residual joins. Perform bilinear upsampling to increase the feature map size.

[0050] Skip connection fusion includes: if there are corresponding layer skip connection features from the encoder. Then, it is concatenated with the upsampled features along the channel dimension to fuse multi-scale information and obtain the fused features. .

[0051] Deep feature extraction includes: Deep feature extraction is performed sequentially through two convolutional layers (each containing 3×3 convolution, batch normalization, and ReLU activation), outputting the feature. .

[0052] Edge attention enhancement includes: constructing a lightweight edge attention subnetwork. Edge features are first extracted using a 1×1 convolutional layer with one channel compression. A spatial attention map is then generated using a sigmoid activation function. Highlighting the edge areas. and Element-wise multiplication is performed to obtain edge enhancement features. .

[0053] Adaptive gated residual fusion includes: The image is compressed into a single-channel gated map G using 1×1 convolution. The original features saved in step 1 are then processed. Perform bilinear upsampling (matching the current spatial size) to obtain Final output features (i.e., channel features) are obtained by weighted fusion of the gated graph G.

[0054] The improved upsampling block precisely enhances the quality of boundary segmentation through an edge attention mechanism, and effectively ensures the stability and integrity of feature transmission by utilizing adaptive gated residual fusion.

[0055] For example, in step three, the segmentation head module uses a 3×3 convolution.

[0056] For example, in step three, the post-processing refinement module adopts a cascaded strategy of hierarchical morphological optimization and Gaussian smoothing.

[0057] Specifically, such as Figure 5 As shown, firstly, the post-processing refinement module processes each non-background category individually: extracting the binary mask of the current category and verifying its validity. For valid masks, closing operations are performed sequentially to fill small holes, and opening operations are performed to eliminate isolated noise and small artifacts, forming a structurally coherent mask. Then, Gaussian blur is applied to the morphologically optimized mask to generate a smooth probability map, and the optimized mask is obtained through fixed-threshold binarization. Finally, the optimized category regions are marked in the final result array, while retaining other category regions in the original prediction map that are not covered by the currently processed category.

[0058] Through this cascaded strategy of hierarchical morphological optimization and Gaussian smoothing, the post-processing refinement module significantly improves boundary smoothness and regional continuity while preserving the original segmentation structure to the greatest extent, and at the same time ensures the rationality of key anatomical structures.

[0059] In one possible embodiment, the spinal image segmentation method based on frequency domain enhancement and feature fusion of this application is experimentally verified as follows: Dataset: The MRSpineSeg2021 dataset was released by the 2nd China Society of Image and Graphics Graphics Technology Challenge (https: / / wwwspinesegmentation-challenge.com). It contains 214 samples, each with expert-labeled T2-weighted images and labels, covering 20 label categories, including 10 vertebrae and 9 intervertebral discs, primarily concentrated in the thoracic and lumbar regions, with slices ranging from 12 to 18.

[0060] Experimental Setup: During preprocessing, the original data of varying sizes were uniformly scaled to 224×224 2D tiles. The dataset was then divided into a training set (172 cases / 2169 tiles), a validation set (22 cases / 276 tiles), and a test set (20 cases / 255 tiles). To address the class imbalance issue (classes 9 / 10 / 18 / 19 only accounted for 19.19% / 2.70% / 20.26% / 2.85% respectively), these four classes were removed, and the dataset was remapped to 16 classes. All experimental models used randomly initialized parameters, employed the SGD optimizer, weight decay of 0.0001, momentum of 0.9, and an initial learning rate of 0.01. Data augmentation included ±15° rotations and horizontal flips with a 50% probability. Early stopping and cosine annealing learning rate scheduling were used during training. The experiment was developed on Ubuntu 22.04, Python 3.9, using PyTorch 2.1.0 and CUDA 11.8, and an Nvidia GeForce RTX 3060 graphics card.

[0061] Evaluation metrics: This embodiment uses five metrics to comprehensively evaluate segmentation performance: Dice similarity coefficient (DSC) and intersection over union (IoU) measure the similarity between the predicted and the true mask; True positive rate (TPR) quantifies the proportion of correctly identified positive samples; Positive predictive value (PPV) assesses the reliability of true positives in the prediction results; 95% Hausdorff distance (95HD) characterizes the alignment accuracy of the segmentation boundary, with smaller values ​​being better. Higher values ​​for DSC, IoU, TPR, and PPV indicate higher segmentation quality. The relevant calculation methods are shown in equations (7), (8), (9), and (10). (7) (8) (9) (10) Where TP means the sample is actually true and the prediction is true; FP means the sample is actually false and the prediction is true; and FN means the sample is actually true and the prediction is false. and These represent the actual segmentation result and the predicted segmentation result, respectively.

[0062] Experimental Results: As shown in Table 1, this embodiment compares the performance of U-Net, TransU-Net, Gctx-UNet, CSwin-UNet, CS-UNet, and the spine image segmentation method based on frequency domain enhancement and feature fusion proposed in this application (denoted as FE-UNet). Experimental results show that FE-UNet performs excellently on core segmentation metrics: Dice coefficient (84.01%), IoU (74.69%), PPV (85.90%), and TPR (84.84%) are all significantly better than all the comparison models. Regarding boundary conditions, although FE-UNet's 95HD is slightly higher than CS-UNet, it is still significantly better than traditional U-Net and other networks.

[0063] Table 1 Comparative Experiments

[0064] As shown in Tables 2 and 3, FE-UNet demonstrates significant advantages in vertebral body and intervertebral disc segmentation: For vertebral body segmentation, it achieves optimal DSC accuracy in four classes: L2, L1, T11, and T12, while its segmentation performance for the sacral S and lumbar L3-L5 is close to the optimal model, with a maximum difference of only 1.47%; for intervertebral disc segmentation, this method achieves optimal performance for five consecutive intervertebral discs from L3 / L4 to T11 / T12, with only slight disadvantages in the lumbosacral junction regions of L5 / S and L4 / L5, which are 0.07% and 1.86% lower than the optimal model, respectively. The results indicate that FE-UNet comprehensively improves the boundary segmentation of key anatomical structures of the spine, especially demonstrating outstanding performance in the thoracolumbar vertebral body and intervertebral disc regions.

[0065] Table 2. DSC results of vertebral body segmentation

[0066] Table 3. DSC results of intervertebral disc segmentation

[0067] like Figure 6 The qualitative comparison results show that FE-UNet demonstrates a certain advantage in complex segmentation compared to the five comparison models. The other models generally suffer from blurred boundaries and semantic confusion at the white boxes. Especially in the last example, facing the weak boundary of the right vertebra due to partial volume effect, FE-UNet still maintains the complete segmentation of the anatomical structure, verifying its robustness to imaging artifacts.

[0068] To verify the effectiveness of each improved module, ablation experiments were conducted under a unified training setting, as shown in Table 4. Compared to the baseline model, FE-UNet achieved significant improvements in segmentation metrics: DSC (+1.58%), IoU (+1.82%), PPV (+1.12%), TPR (+1.91%), and a 95 HD reduction of 0.84 mm. These results confirm the synergistic optimization effect of frequency domain enhancement, attention-guided upsampling blocks, and morphological post-processing strategies, which can improve the segmentation accuracy of vertebral bodies and intervertebral discs.

[0069] Table 4 Ablation Experiment

[0070] This application embodiment constructs an encoder combining a frequency domain enhancement module, a residual convolution module, and a Transformer module, and a decoder combining an upsampling block, a segmentation head module, and a post-processing refinement module. The frequency domain enhancement module performs frequency domain enhancement on the spinal image to be segmented, obtaining enhanced frequency domain features. The residual convolution module performs residual downsampling on the enhanced frequency domain features, obtaining a deep feature map. The Transformer module performs self-attention processing on the deep feature map, obtaining an encoded feature sequence. The upsampling block upsamples and concatenates the encoded feature sequence, obtaining channel features. The segmentation head module maps the channel features to the number of segmentation categories, obtaining a preliminary segmentation map. The post-processing refinement module smooths the boundaries of the preliminary segmentation map, obtaining a spinal segmentation mask. The synergistic effect of these modules improves the segmentation accuracy of vertebral bodies and intervertebral discs in the spinal image.

[0071] The proposed method involves sequentially performing Gaussian blur convolution, Fourier transform, spectral centering, high-pass filtering, inverse discrete Fourier transform, normalization, contrast enhancement, and weighted fusion on the spinal image to be segmented, thereby obtaining frequency domain enhancement features. The image is then converted to the frequency domain through Fourier transform, and combined with high- and low-frequency separation and dual noise reduction strategies, effectively strengthening edge features and enhancing semantic information expression.

[0072] The proposed method sequentially performs feature preservation and upsampling, skip connection fusion, deep feature extraction, edge attention enhancement, and adaptive gated residual fusion on the encoded feature sequence to obtain channel features. Dual-path feature fusion is achieved through dynamic edge attention maps, thereby better recovering detailed information.

[0073] The proposed post-processing refinement module employs a cascaded strategy of hierarchical morphological optimization and Gaussian smoothing to further optimize the smoothness and continuity of the segmentation boundaries.

[0074] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0075] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A spine image segmentation method based on frequency domain enhancement and feature fusion, characterized in that, Includes the following steps: Step 1: Obtain the image of the spine to be segmented; Step 2: Input the spinal image to be segmented into an encoder based on a frequency domain enhancement module, a residual convolution module, and a Transformer module to obtain an encoded feature sequence; Step 3: Input the encoded feature sequence into a decoder based on an upsampling block, a segmentation head module, and a post-processing refinement module to obtain a spine segmentation mask; The frequency domain enhancement module is used to perform frequency domain enhancement on the spinal image to be segmented to obtain frequency domain enhancement features; The residual convolution module is used to perform residual downsampling on the frequency domain enhanced features to obtain a deep feature map; The Transformer module is used to perform self-attention processing on the deep feature map to obtain the encoded feature sequence; The upsampling block is used to upsample and concatenate the encoded feature sequence to obtain channel features; The segmentation head module is used to map the channel features to the number of segmentation categories to obtain a preliminary segmentation map; The post-processing refinement module is used to perform boundary smoothing on the preliminary segmentation map to obtain a spine segmentation mask.

2. The spinal image segmentation method based on frequency domain enhancement and feature fusion according to claim 1, characterized in that, In step two, the processing flow of the frequency domain enhancement module includes: The spinal image to be segmented is sequentially subjected to Gaussian blur convolution, Fourier transform, spectrum centering, high-pass filtering, inverse discrete Fourier transform, normalization, contrast enhancement, and weighted fusion to obtain frequency domain enhanced features.

3. The spinal image segmentation method based on frequency domain enhancement and feature fusion according to claim 1, characterized in that, In step two, the residual convolution module is constructed based on the ResNet50 backbone network.

4. The spinal image segmentation method based on frequency domain enhancement and feature fusion according to claim 1, characterized in that, In step two, the Transformer module performs self-attention processing on the deep feature map through a multi-head self-attention mechanism and a multilayer perceptron extension module to obtain an encoded feature sequence.

5. The spinal image segmentation method based on frequency domain enhancement and feature fusion according to claim 1, characterized in that, In step three, the upsampling block is improved. The processing flow of the improved upsampling block includes: The encoded feature sequence is sequentially subjected to feature preservation and upsampling, skip connection fusion, deep feature extraction, edge attention enhancement, and adaptive gated residual fusion to obtain channel features.

6. The spinal image segmentation method based on frequency domain enhancement and feature fusion according to claim 1, characterized in that, In step three, the segmentation head module uses a 3×3 convolution.

7. The spinal image segmentation method based on frequency domain enhancement and feature fusion according to claim 1, characterized in that, In step three, the post-processing refinement module adopts a cascaded strategy of hierarchical morphological optimization and Gaussian smoothing.

Images