Vertebral segmentation method and system based on edge-enhanced U-Net

Abstract

The present invention provides a vertebra segmentation method and system based on edge-enhanced U-Net, belonging to the field of image processing technology, obtaining a vertebral image to be segmented; using a pre-trained segmentation model to process the vertebral image to be segmented, and obtaining a final vertebral segmentation result; wherein the pre-trained segmentation model is trained using a training set with edge-enhanced U-Net as the basic network; the training set includes multiple vertebral images and labels of vertebral pixels in the annotated images. The present invention uses a U-Net network with an encoder and decoder structure to segment vertebral images, the U-Net encoder extracts deep features, and the skip connection technology retains shallow textures; the U-Net decoder restores the deep abstract information to the size of the original image, and finally calculates the cross entropy loss according to the true value of the label; different weights are given to each pixel, so that the loss of pixels near the boundary is increased, so that the network can learn the characteristics of the boundary information and obtain ideal segmentation results.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a vertebra segmentation method and system based on edge-enhanced U-Net. Background Art

[0002] The spine is the pillar of the body. It not only supports the trunk, but also protects the internal organs and assists body movement. With the accelerated pace of life, spinal diseases are not only more common among the elderly, but also tend to be younger. Spinal diseases such as lumbar disc herniation and scoliosis plague patients' lives, making the accurate diagnosis and prognosis of spinal diseases the focus of medical attention.

[0003] In recent years, with the development of medical imaging technologies such as X-ray, CT, and MRI, spinal imaging has greatly facilitated doctors and radiologists in diagnosing spinal diseases. Vertebral morphology estimation, especially intervertebral space measurement, is crucial for spinal surgery and preoperative planning: the accuracy of intervertebral space measurement will directly affect the shape parameters of the implanted prosthesis, and these shape parameters play a decisive role in the quality of patient rehabilitation in the later stage. However, the current vertebral morphology estimation based on imaging data is still highly dependent on the professional knowledge of doctors.

[0004] The introduction of emerging artificial intelligence technology is expected to train intelligent models based on a small amount of manually labeled data, achieve the purpose of intelligent identification of vertebrae, realize effective intelligent surgical planning, and ultimately promote the clinical treatment of spinal diseases. The key point of completing vertebral morphology estimation is to achieve effective segmentation of vertebrae based on spinal images. Once the vertebrae are segmented, the width, length, and bending angle (caused by spinal curvature) of the spine can be calculated according to the image scale.

[0005] Image segmentation usually uses morphological information such as object shape, boundary, texture, etc. to divide the image into different regions. Most of the early spinal image segmentation used traditional machine learning segmentation algorithms based on prior knowledge, such as threshold method, gradient method, region growing method, edge detection method, watershed method, morphological filter and Gaussian mixture model. Traditional machine learning methods have many processing steps and weak generalization ability. In addition, the spinal block morphology and structure are complex, resulting in segmentation effects that do not meet the needs of clinical auxiliary diagnosis.

[0006] After the birth of deep learning, computer vision technology has developed rapidly, and deep networks represented by convolutional neural networks have achieved excellent performance in image processing. U-Net is a typical image segmentation network, which learns deep features in medical images through multi-layer convolution and retains shallow texture features through skip-layer connections. A large number of research results show that compared with traditional machine learning methods, U-Net has superior image segmentation accuracy.

[0007] As for vertebrae segmentation based on X-ray images, due to spinal lesions, aging, etc., the vertebrae have irregular shapes and some vertebrae have blurred contours, which leads to poor U-Net segmentation results. Summary of the invention

[0008] The object of the present invention is to provide a vertebral segmentation method and system based on edge-enhanced U-Net, which combines the morphological characteristics of the vertebrae, considers more vertebral features, and improves the accuracy of vertebral segmentation, so as to solve at least one technical problem existing in the above-mentioned background technology.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] In one aspect, the present invention provides a vertebra segmentation method based on edge-enhanced U-Net, comprising:

[0011] Acquire a vertebrae image to be segmented;

[0012] The vertebrae image to be segmented is processed using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0013] Preferably, when preparing a training set, the image is preprocessed and a weight map is prepared, including: enhancing the training image to expand the training data set; filling the background of the training image and adjusting images of different sizes to a fixed size with uniform length and width; recalculating the weight of each pixel on the annotated image based on pre-manual annotation, assigning the highest weight to the boundary pixels, assigning a larger weight to the pixels closer to the boundary points, and assigning a smaller weight to the pixels farther from the boundary points.

[0014] Preferably, constructing an edge-enhanced U-Net network model includes constructing an encoder: for an input image, a final encoding result is obtained through stacked convolution-downsampling operations; wherein, in each layer of convolution-downsampling operations, the image undergoes two multi-channel convolutions to extract features; after the convolution operation, ReLu is used as an activation function; maximum pooling acts as a downsampling layer, and maximum pooling is used to reduce the size of the feature map.

[0015] Preferably, constructing an edge-enhanced U-Net network model also includes constructing a decoder: based on the encoding result of the encoder, a final pixel vertebral block membership probability map is obtained through stacked upsampling-convolution operations; wherein, in each layer of upsampling-convolution operations, the feature map is upsampled by bilinear interpolation to obtain a feature map with doubled dimension, and the feature map is skipped with the feature map of the corresponding level of the encoder according to the channel to obtain a feature map with doubled number of channels; convolution is performed on the feature map with doubled number of channels; at the highest layer of the decoder, the final pixel vertebral block membership probability map is output after convolution.

[0016] Preferably, when training the segmentation model, the weighted cross entropy between the segmentation result and the manual annotation is used as the loss function, and the formula is:

[0017]

[0018] Among them, Ω represents the feature map, w:Ω→R represents the pre-set pixel weight map; l:Ω→{1,2,...,K} represents the category of the pixel point, K is the total number of categories; p k (x) is the softmax function:

[0019]

[0020] Among them, a k (x) represents the activation value of pixel x in the kth category;

[0021] The Adam optimizer is used, the learning rate multiplication factor is 0.96, and it is updated once every epoch;

[0022] The training adopts frozen training. First, the encoder part is frozen and the decoder part is trained; then the encoder part is unfrozen and the encoder and decoder parts are trained.

[0023] Preferably, a custom vertebral boundary weight map is used, and the weight calculation formula of each pixel point in the vertebral boundary weight map is as follows:

[0024]

[0025] Among them, w c :Ω→R is the basic weight, which is always 1; d:Ω→R represents the distance from the pixel point x to its nearest vertebral boundary pixel; w 0 is a hyperparameter that controls the weight of the boundary point; σ is a hyperparameter that controls the radiation range of the boundary point.

[0026] In a second aspect, the present invention provides a vertebra segmentation system based on edge-enhanced U-Net, comprising:

[0027] An acquisition module, used for acquiring a vertebrae image to be segmented;

[0028] The segmentation module is used to process the vertebrae image to be segmented using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0029] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the vertebral segmentation method based on edge-enhanced U-Net as described above is implemented.

[0030] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed on one or more processors, is used to implement the vertebra segmentation method based on edge-enhanced U-Net as described above.

[0031] In a fifth aspect, the present invention provides an electronic device, comprising: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes instructions for implementing the vertebrae segmentation method based on edge-enhanced U-Net as described above.

[0032] The invention has the following beneficial effects: a U-Net network with an encoder and decoder structure is used to segment vertebral images, wherein the encoder uses a VGG16 feature extraction backbone, which has a strong feature extraction capability; image inversion, translation, rotation and other enhancement techniques are used to prevent overfitting; the U-Net encoder extracts deep features, and the skip-layer connection technology retains shallow textures; the U-Net decoder restores deep abstract information to the size of the original image, and finally calculates the cross entropy loss according to the true value of the label; in the inference process, different weights are given to each pixel point, so that the loss of pixels near the boundary is increased, so that the network can learn the characteristics of the boundary information and obtain an ideal segmentation result.

[0033] Additional advantages of the present invention will be more clearly given in the following description or learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0035] Figure 1 Schematic diagram of the structure of the U-Net network model described in an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of a real label according to an embodiment of the present invention.

[0037] Figure 3 This is the weight graph described in the embodiment of the present invention.

[0038] Figure 4 It is a schematic diagram of the prediction result without edge enhancement according to an embodiment of the present invention.

[0039] Figure 5 It is a schematic diagram of the prediction result of edge enhancement according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below by the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0041] It should be understood by those skilled in the art that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.

[0042] It should also be understood that terms, such as those defined in commonly used dictionaries, should be understood to have a meaning consistent with that in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless as defined herein.

[0043] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements and / or groups thereof.

[0044] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0045] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings, and the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0046] Those skilled in the art should understand that the drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily necessary for implementing the present invention.

[0047] Example 1

[0048] This embodiment 1 provides a vertebra segmentation system based on edge-enhanced U-Net, the system comprising:

[0049] An acquisition module, used for acquiring a vertebrae image to be segmented;

[0050] The segmentation module is used to process the vertebrae image to be segmented using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0051] In this embodiment 1, the vertebra segmentation system based on edge enhancement U-Net is used to implement a vertebra segmentation method based on edge enhancement U-Net, including:

[0052] Acquire the vertebrae image to be segmented by using an acquisition module;

[0053] The vertebrae image to be segmented is processed by a segmentation module using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is trained using an edge-enhanced U-Net as a basic network using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0054] When making a training set, the image is preprocessed and a weight map is made, including: enhancing the training image to expand the training data set; filling the background of the training image and adjusting the images of different sizes to a fixed size with uniform length and width; recalculating the weight of each pixel on the annotated image according to the pre-manual annotation, assigning the highest weight to the boundary pixels, assigning a larger weight to the pixels closer to the boundary points, and assigning a smaller weight to the pixels farther from the boundary points.

[0055] In this embodiment 1, constructing an edge-enhanced U-Net network model includes constructing an encoder: for an input image, a final encoding result is obtained through stacked convolution-downsampling operations; wherein, in each layer of convolution-downsampling operations, the image is subjected to two multi-channel convolutions to extract features; ReLu is used as an activation function after the convolution operation; maximum pooling acts as a downsampling layer, and the size of the feature map is reduced by maximum pooling. Constructing an edge-enhanced U-Net network model also includes constructing a decoder: based on the encoder encoding result, a final pixel vertebral block membership probability map is obtained through stacked upsampling-convolution operations; wherein, in each layer of upsampling-convolution operations, the feature map is upsampled by bilinear interpolation to obtain a feature map with doubled dimensions, and the feature map is skipped and connected with the feature map of the corresponding level of the encoder according to the channel to obtain a feature map with doubled channel numbers; convolution is performed on the feature map with doubled channel numbers; at the highest level of the decoder, the final pixel vertebral block membership probability map is output through convolution.

[0056] When training the segmentation model, the weighted cross entropy between the segmentation result and the manual annotation is used as the loss function, and the formula is:

[0057]

[0058] Among them, Ω represents the feature map, w:Ω→R represents the pre-set pixel weight map; l:Ω→{1,2,...,K} represents the category of the pixel point, K is the total number of categories; p k (x) is the softmax function:

[0059]

[0060] Among them, a k (x) represents the activation value of pixel x in the kth category;

[0061] The Adam optimizer is used, the learning rate multiplication factor is 0.96, and it is updated once every epoch;

[0062] The training adopts frozen training. First, the encoder part is frozen and the decoder part is trained; then the encoder part is unfrozen and the encoder and decoder parts are trained.

[0063] A custom vertebral boundary weight map is used. The weight calculation formula for each pixel point of the vertebral boundary weight map is as follows:

[0064]

[0065] Among them, w c :Ω→R is the basic weight, which is always 1; d:Ω→R represents the distance from the pixel point x to its nearest vertebral boundary pixel; w 0 is a hyperparameter that controls the weight of the boundary point; σ is a hyperparameter that controls the radiation range of the boundary point.

[0066] Example 2

[0067] like Figures 1 to 5 As shown, in this embodiment 2, an algorithm for improving the accuracy of vertebral segmentation by using spine boundary information is provided. The specific implementation steps of the algorithm are as follows:

[0068] Step 1: Preprocess the image and make a weight map;

[0069] Step 1.1: Enhance the training images, including inversion, rotation, mirror symmetry and other operations to expand the training data set.

[0070] Step 1.2: Perform background filling on the training images and resize the images of different sizes to a fixed size [3×512×512] with uniform length and width.

[0071] Step 1.3: Recalculate the weight of each pixel on the annotated image based on the pre-annotated manual annotations, assign the highest weight to the boundary pixels, and assign larger (smaller) weights to pixels that are closer (farther) from the boundary points.

[0072] Step 2: Build the U-Net network model;

[0073] Step 2.1: Construct the encoder part. For an input image of size [3×512×512], the final encoding result is obtained through stacked convolution-downsampling operations. The convolution-downsampling operation of each layer adopts a similar model structure: (1) The image is subjected to two [3×3] multi-channel convolutions to extract features; the feature extraction backbone uses VGG16, and the convolution layer uses padding, the feature map size remains unchanged but the number of channels increases; (2) ReLu is used as the activation function after the convolution operation; (3) The maximum pooling acts as a downsampling layer, and the maximum pooling of [2×2] is used to reduce the feature map size. After the [2×2] maximum pooling, the feature map dimension is halved. The above is the encoder part of U-Net.

[0074] Step 2.2: Construct the decoder. Based on the encoder encoding result, the final pixel vertebral block membership probability map is obtained through stacked upsampling-convolution operations. The upsampling-convolution operation of each layer adopts a similar model structure. Take the upsampling-convolution operation of the bottom layer (the layer with the lowest feature dimension) as an example: (1) The feature map of the bottom layer is upsampled by bilinear interpolation to obtain a feature map with doubled dimension; (2) The feature map is skipped with the feature map of the corresponding layer of the encoder according to the channel to obtain a feature map with doubled number of channels; (3) For the feature map with doubled number of channels, a [3×3] convolution is performed. The convolution layer is padded with padding. The feature map size remains unchanged but the number of channels is reduced. At the highest layer of the decoder (the layer with feature dimension close to the input image), the final pixel vertebral block membership probability map is output after [1×1] convolution, with a size of [2×512×512]. The above is the decoder part of U-Net.

[0075] Step 2.3: Use the weighted cross entropy between the segmentation result and the manual annotation as the loss function. The formula is as follows:

[0076]

[0077] Where Ω represents the feature map, w:Ω→R represents the preset pixel weight map; l:Ω→{1,2,...,K} represents the category of the pixel point, K is the total number of categories, here K=2; p k (x) is the softmax function:

[0078]

[0079] Among them, a k (x) represents the activation value of pixel x in the kth category.

[0080] Step 2.4: The Adam optimizer is used for training, and the learning rate multiplication factor is 0.96, which is updated once every epoch. The training adopts frozen training. First, the encoder part is frozen and the decoder part is trained. Then the encoder part is unfrozen and the encoder and decoder parts are trained.

[0081] Among them, a custom spine edge weight map is used in step 1.3, and the weight map calculation formula is as follows:

[0082]

[0083] Among them, w c :Ω→R is the basic weight, which is always 1; d:Ω→R represents the distance from the pixel point x to its nearest vertebral boundary pixel; w 0 is a hyperparameter that controls the weight of the boundary point, and its empirical value is 10; σ is a hyperparameter that controls the radiation range of the boundary point, and its empirical value is 5.

[0084] Example 3

[0085] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, a vertebra segmentation method based on edge-enhanced U-Net is implemented. The method includes:

[0086] Acquire a vertebrae image to be segmented;

[0087] The vertebrae image to be segmented is processed using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0088] Example 4

[0089] Embodiment 3 of the present invention provides a computer program (product), including a computer program, wherein when the computer program is executed on one or more processors, the computer program is used to implement a vertebra segmentation method based on edge-enhanced U-Net, the method comprising:

[0090] Acquire a vertebrae image to be segmented;

[0091] The vertebrae image to be segmented is processed using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0092] Example 5

[0093] Embodiment 4 of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing a vertebra segmentation method based on edge-enhanced U-Net, the method comprising:

[0094] Acquire a vertebrae image to be segmented;

[0095] The vertebrae image to be segmented is processed using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images.

[0096] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0099] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0100] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative work on the basis of the technical solution disclosed in the present invention should be included in the scope of protection of the present invention.

Claims

1. A vertebral segmentation method based on edge-enhanced U-Net, It is characterized in that include: Acquire a vertebrae image to be segmented; Processing the vertebrae image to be segmented using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is based on an edge-enhanced U-Net network and is trained using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images; When making a training set, the image is preprocessed and a weight map is made, including: enhancing the training image to expand the training data set; filling the background of the training image and adjusting the images of different sizes to a fixed size with uniform length and width; recalculating the weight of each pixel on the annotated image based on the pre-manual annotation, assigning the highest weight to the boundary pixels, assigning a larger weight to the pixels closer to the boundary points, and assigning a smaller weight to the pixels farther from the boundary points; Building the edge-enhanced U-Net network model includes building an encoder: for the input image, the final encoding result is obtained through stacked convolution-downsampling operations; in each layer of convolution-downsampling operations, the image undergoes two multi-channel convolutions to extract features; after the convolution operation, ReLu is used as the activation function; the maximum pooling acts as a downsampling layer, and the maximum pooling is used to reduce the size of the feature map; Constructing the edge-enhanced U-Net network model also includes constructing a decoder: based on the encoding result of the encoder, a final pixel vertebral block membership probability map is obtained through stacked upsampling-convolution operations; wherein, in each layer of upsampling-convolution operations, the feature map is upsampled by bilinear interpolation to obtain a feature map with doubled dimensions, and the feature map is skipped with the feature map of the corresponding level of the encoder according to the channel to obtain a feature map with doubled channels; convolution is performed on the feature map with doubled channels; at the highest layer of the decoder, the final pixel vertebral block membership probability map is output through convolution; When training the segmentation model, the weighted cross entropy between the segmentation result and the manual annotation is used as the loss function, and the formula is: Among them, Ω represents the feature map, w:Ω→R represents the pre-set pixel weight map; l:Ω→{1,2,...,K} represents the category of the pixel point, K is the total number of categories; p k (x) is the softmax function: Among them, a k (x) represents the activation value of pixel x in the kth category; The Adam optimizer is used, the learning rate multiplication factor is 0.96, and it is updated once every epoch; The training adopts frozen training. First, the encoder part is frozen and the decoder part is trained; then the encoder part is unfrozen and the encoder and decoder parts are trained.

2. The vertebral segmentation method based on edge-enhanced U-Net according to claim 1, It is characterized in that A custom vertebral boundary weight map is used. The weight calculation formula for each pixel point of the vertebral boundary weight map is as follows: Among them, w c :Ω→R is the basic weight, which is always 1; d:Ω→R represents the distance from the pixel point x to its nearest vertebral boundary pixel; w 0 is a hyperparameter that controls the weight of the boundary point; σ is a hyperparameter that controls the radiation range of the boundary point.

3. A vertebral segmentation system based on edge-enhanced U-Net, It is characterized in that include: An acquisition module, used for acquiring a vertebrae image to be segmented; A segmentation module, used to process the vertebrae image to be segmented using a pre-trained segmentation model to obtain a final vertebrae segmentation result; wherein the pre-trained segmentation model is trained using an edge-enhanced U-Net as a basic network using a training set; the training set includes multiple vertebrae images and labels of vertebrae pixels in the annotated images; When making a training set, the image is preprocessed and a weight map is made, including: enhancing the training image to expand the training data set; filling the background of the training image and adjusting the images of different sizes to a fixed size with uniform length and width; recalculating the weight of each pixel on the annotated image based on the pre-manual annotation, assigning the highest weight to the boundary pixels, assigning a larger weight to the pixels closer to the boundary points, and assigning a smaller weight to the pixels farther from the boundary points; Building the edge-enhanced U-Net network model includes building an encoder: for the input image, the final encoding result is obtained through stacked convolution-downsampling operations; in each layer of convolution-downsampling operations, the image undergoes two multi-channel convolutions to extract features; after the convolution operation, ReLu is used as the activation function; the maximum pooling acts as a downsampling layer, and the maximum pooling is used to reduce the size of the feature map; Constructing the edge-enhanced U-Net network model also includes constructing a decoder: based on the encoding result of the encoder, a final pixel vertebral block membership probability map is obtained through stacked upsampling-convolution operations; wherein, in each layer of upsampling-convolution operations, the feature map is upsampled by bilinear interpolation to obtain a feature map with doubled dimensions, and the feature map is skipped with the feature map of the corresponding level of the encoder according to the channel to obtain a feature map with doubled channels; convolution is performed on the feature map with doubled channels; at the highest layer of the decoder, the final pixel vertebral block membership probability map is output through convolution; When training the segmentation model, the weighted cross entropy between the segmentation result and the manual annotation is used as the loss function, and the formula is: Among them, Ω represents the feature map, w:Ω→R represents the pre-set pixel weight map; l:Ω→{1,2,...,K} represents the category of the pixel point, K is the total number of categories; p k (x) is the softmax function: Among them, a k (x) represents the activation value of pixel x in the kth category; The Adam optimizer is used, the learning rate multiplication factor is 0.96, and it is updated once every epoch; The training adopts frozen training. First, the encoder part is frozen and the decoder part is trained; then the encoder part is unfrozen and the encoder and decoder parts are trained.

4. A non-transitory computer-readable storage medium, It is characterized in that The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by the processor, the vertebrae segmentation method based on edge-enhanced U-Net as claimed in claim 1 or 2 is implemented.

5. A computer program product, It is characterized in that The invention comprises a computer program, which, when running on one or more processors, is used to implement the vertebra segmentation method based on edge-enhanced U-Net as claimed in claim 1 or 2.

6. An electronic device, It is characterized in that include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes instructions for implementing the vertebrae segmentation method based on edge-enhanced U-Net as described in claim 1 or 2.

Images