WO2024001140A1

WO2024001140A1 - Vertebral body sub-region segmentation method and apparatus, and storage medium

Info

Publication number: WO2024001140A1
Application number: PCT/CN2022/143887
Authority: WO
Inventors: 李危石; 邹达; 齐欢; 吕维加
Original assignee: 北京大学第三医院（北京大学第三临床医学院）
Priority date: 2022-06-28
Filing date: 2022-12-30
Publication date: 2024-01-04
Also published as: CN114998301A; CN114998301B

Abstract

Disclosed are a vertebral body sub-region segmentation method and apparatus, and a storage medium, relating to the technical field of medical image processing. The segmentation method comprises: preprocessing obtained spine vertebral body image data; inputting the preprocessed image data into a pre-trained neural network model to obtain a sub-region segmentation mask regression result corresponding to a spine vertebral body; and performing post-processing on the mask regression result by means of morphological operation and connectivity test to obtain three-dimensional segmentation masks for vertebral body superior and inferior bony endplate regions having a uniform and customizable thickness, and cortical bone side plate and cancellous bone regions, thereby completing vertebral body sub-region segmentation. The problems in the prior art that a vertebral body sub-region segmentation method cannot implement personalized endplate and inferior bony endplate individual sub-region segmentation matching the complex anatomical form of the vertebral body bony endplates, the segmentation performance for the vertebral body side plate and vertebral body cancellous bone regions is poor, and the generalization ability is low are solved.

Description

A vertebral body sub-region segmentation method, device and storage medium

Technical field

The present application relates to the field of medical image processing, and in particular to a vertebral body sub-region segmentation method, device and storage medium.

Background technique

The vertebral body subregions of the spine mainly include bony endplates, cortical bone side walls, and cancellous bone, all of which play an important supporting role in the human body. Among them, the bone quality of the bony endplate and the bone area adjacent to the endplate is one of the key factors affecting the efficacy of spinal interbody fusion. The worse the endplate bone quality, the higher the risk of endplate damage during interbody fusion, and postoperative endplate collapse and cage subsidence and displacement. Analyzing the vertebral bony endplates, cortical bone lateral plates, and cancellous bone areas (hereinafter referred to as vertebral subregions) from aspects such as bone density, mechanics, and morphological structure will help establish quantitative evaluation indicators for bone quality. Assist clinical diagnosis and surgical planning. Based on three-dimensional medical images (such as computed tomography, CT), accurate three-dimensional segmentation and three-dimensional reconstruction of vertebral body sub-regions can be performed. Existing technical solutions generally segment vertebral body images based on underlying image processing methods such as morphological algorithms and similarity indicators, and can automatically generate segmentation masks for vertebral body sub-regions.

The bony endplate segmented by existing methods mainly has the following problems. First, the endplate segmentation area cannot match the complex endplate anatomical shape well; secondly, in fact, the bone density of subregions with different thicknesses and shapes (hereinafter referred to as endplate subregions) within the same endplate range is also different. Different from each other, the implants implanted in the intervertebral spine during spinal surgery often only come into contact with a certain endplate sub-region and do not contact the entire endplate. Existing methods cannot achieve segmentation of customized endplate subregions; thirdly, existing methods cannot separate and exclude non-osseous endplate regions such as osteophytes.

On the other hand, existing methods have poor segmentation performance for the vertebral body lateral plates and vertebral body cancellous bone regions. There is a strong dependence on the quality of the vertebral body image, that is, the generalization ability is poor, and it is difficult to adapt to different imaging equipment (such as the brand and model of CT equipment), different scanning parameters (voltage, radiation dose, scanning layer thickness, reconstruction kernel). Function), different spinal morphology (structural abnormalities, such as deformities, fractures, etc.; abnormal vertebral body angles, such as scoliosis, anterior and posterior kyphosis, etc.) and other scenarios maintain high segmentation accuracy.

The above problems have restricted the widespread clinical application of existing methods.

Contents of the invention

In view of the above analysis, this application aims to provide a vertebral body sub-region segmentation method, device and storage medium, and obtain a three-dimensional segmentation mask of the vertebral body sub-region through the pre-trained vertebral body sub-region segmentation neural network model. This application solves problems not covered by vertebral body sub-region segmentation methods in the prior art, including: the inability to achieve customized endplate and sub-endplate bone sub-region segmentation that matches the complex anatomical shape of the vertebral bony endplate; The segmentation performance of the vertebral body lateral plate and vertebral body cancellous bone area is poor, and the generalization ability is not high. This application can be expanded and applied in the field of spinal local bone density calculation.

The purpose of this application is mainly achieved through the following technical solutions:

On the one hand, a vertebral body sub-region segmentation method is provided, including the following steps:

Preprocess the obtained spinal vertebral body image data;

Input the pre-processed image data into the pre-trained neural network model to obtain the sub-region segmentation mask regression results corresponding to the spinal vertebrae; among them, the neural network model is used to output the mask regression results of the upper and lower bony endplates of the vertebral body, cortex Bone lateral plate mask regression results, cancellous bone area mask regression results, and vertebral body mask regression results based on fusion;

The vertebral body mask regression results are post-processed through morphological operations and connectivity tests to obtain a three-dimensional segmentation mask of the upper and lower bony endplates, cortical bone lateral plates, and cancellous bone areas of the vertebral body to complete vertebral body sub-region segmentation;

The vertebral body subregion includes the upper and lower bony endplates of the vertebral body, cortical bone lateral plates, and cancellous bone areas.

Further, the pre-trained neural network model is a bifurcated multi-task convolutional neural network structure, including an encoder, three decoders and a MAX fusion unit: where,

The encoder is used to receive preprocessed image data to obtain feature maps;

The three decoders are respectively connected to the encoder and used to output the upper and lower bony endplate mask regression results, cortical bone lateral plate mask regression results and vertebral cancellous bone mask regression results based on the feature map;

The MAX fusion unit is connected to three encoders respectively and used to output the vertebral mask regression results.

Further, the pre-training process of the neural network includes:

For the preprocessed image data, upper and lower bony endplate mask labels, cortical bone lateral plate mask labels, vertebral body cancellous bone mask labels are constructed respectively, and the vertebral body mask label is obtained by fusion;

By simulating the image features obtained in different scenarios, the image data after constructing the label is augmented to obtain training samples to expand the training data set;

Calculate the loss functions of the four output results of the neural network model and the corresponding labels, and assign weights to the four output loss functions to obtain the total loss function of the neural network model; based on the total loss function of the neural network model, the gradient descent method is used to determine the model parameters. Iterative training.

Furthermore, the binary mask method was used to annotate the voxels in the training sample image, and the upper and lower bony endplate mask labels, cortical bone lateral plate mask labels, and vertebral cancellous bone mask were constructed for the training sample image. Label;

The vertebral body mask label is constructed by fusing the upper and lower bony endplate mask labels, cortical bone lateral plate mask labels, and vertebral body cancellous bone mask labels.

Among them, the endplate mask label is constructed, including customizing the thickness and shape of the endplate area to obtain the endplate mask label.

Further, the gradient descent method is used to iterate the model parameters, including:

The weight of cross entropy and Dice function is used as the loss function of each output result, where the cross entropy function is:

The Dice function is:

Among them, a is an output result in the neural network model, b is the labeling result, i is the voxel position index, and |n| is the total prime number.

Further, the MAX fusion unit is used to output the vertebral mask regression results, including: based on the position index of each voxel of the preprocessed spinal vertebral image, taking the maximum value of the corresponding index of the three decoders, and calculating the position index of each voxel. The three decoders perform MAX fusion on the maximum value of the corresponding index, and output the vertebral mask regression result.

Further, the preprocessing of the spinal vertebra image data includes resampling and pixel value normalization; among which,

The resampling process includes: dividing the spatial resolution of the obtained spinal vertebral body image by the preset spatial resolution to obtain the resampling ratio of the image data in three dimensions; based on the resampling ratio, using linear interpolation method to obtain the resampling The resulting image data with fixed spatial resolution.

Pixel value normalization processing includes: mapping the original pixel value range [M, N] to the preset value range [P, Q] through a linear function; where M is the minimum pixel value of the CT image, and N is the maximum pixel value of the CT image. , P is the lower bound of the preset value range, and Q is the upper bound of the preset value range.

Further, the mask regression results are post-processed through morphological operations and connectivity tests, including: using a three-dimensional morphological opening operation with a convolution kernel of 3*3*3, and using the sliding window method to calculate the neighbors around each voxel. Domains are morphologically eroded and dilated to remove fine-grained noise in the mask to remove areas of abnormal bone structure including vertebral osteophytes.

Use the skimage toolkit to calculate the connectivity of the upper and lower bony endplate masks, and retain the two largest connected areas in the results, corresponding to the upper and lower bony endplates respectively; the cortical bone lateral plate mask and the vertebral cancellous The bone mask retains only the largest connected area.

On the other hand, this application also provides a vertebral body sub-region segmentation device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the The computer program implements the aforementioned vertebral body sub-region segmentation method.

In a third aspect, a computer-readable storage medium is also provided. A computer program is stored on the storage medium, and the computer program can be executed by a processor to implement the aforementioned vertebral body sub-region segmentation method.

This application can also be extended to the field of local bone density calculation to calculate the average CT value within the segmented area. Based on the CT value, quantitative calculation and analysis of bone mineral content can be performed through automated positioning analysis of muscle fat.

Beneficial effects of this technical solution:

1. In view of the problem that existing methods cannot achieve personalized endplate sub-region segmentation that matches the complex endplate anatomy and have poor performance in segmenting vertebral body side walls and cancellous bone regions, this application divides the vertebral body into upper and lower bones. There are three regions: sexual endplate, cortical bone lateral plate and vertebral cancellous bone. It can achieve personalized endplate thickness segmentation, endplate sub-region segmentation matching the "implant-endplate" contact surface, and remove abnormal bone structures such as osteophytes. Segmenting the vertebral body more accurately further expands the application scope of the vertebral body segmentation method.

2. In view of the problem of poor generalization ability of the existing technology, this application proposes a machine learning technology method based on a pre-trained neural network model, and uses the data augmentation method to simulate the vertebral image features obtained in different scenarios. Increase the adaptability of the prediction model to image contrast, image noise, vertebral body posture, and image layer thickness, thereby obtaining stronger generalization ability. Compared with the existing technology, this method can learn image features independently without using a specific feature extraction method.

3. In view of the problem that the existing technology is not scalable, this application adopts a data-driven modeling method, which can improve the performance of the prediction model by adding training data, and can add expert opinions to the model training process through data annotation. , and continue to optimize.

Additional features and advantages of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the application may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

Description of drawings

The drawings are for the purpose of illustrating specific embodiments only and are not to be construed as limiting the application. Throughout the drawings, the same reference characters represent the same components.

Figure 1: Schematic flow chart of the vertebral body sub-region segmentation method according to the embodiment of the present application;

Figure 2: Schematic structural diagram of the vertebral body sub-region segmentation neural network model according to the embodiment of the present application;

Figure 3: Structural diagram of a neural network model for training vertebral body sub-region segmentation according to the embodiment of the present application;

Figure 4: Comparison of the original spine image and the corresponding data augmentation result according to the embodiment of the present application;

Figure 5: Structural diagram of the vertebral body sub-region segmentation method according to the embodiment of the present application;

Figure 6: A schematic diagram of the vertebral body osteophyte removal method according to the embodiment of the present application, which removes the outer ring through an erosion algorithm to remove peripheral osteophytes;

Detailed ways

Preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings. The drawings constitute a part of the present application and are used together with the embodiments of the present application to illustrate the principles of the present application, but are not intended to limit the scope of the present application.

This application obtains the vertebral body sub-region segmentation mask by preprocessing the spinal vertebral body image data, constructing mask labels, neural network model prediction and post-processing operations; the vertebral body sub-regions in this application include: vertebral body bony end plates, lateral cortical plates, and cancellous bone areas of the vertebral body. For example, the spinal vertebral body image data is CT image data.

One embodiment of the present application provides a vertebral body sub-region segmentation method, as shown in Figure 1, including the following steps:

Step 1: Preprocess the obtained spinal vertebral body image data;

Specifically, preprocessing includes resampling and pixel value normalization. Divide the spatial resolution of the spine vertebral body image by the preset spatial resolution to obtain the resampling ratio of the image data in three dimensions; according to the resampling ratio, use a linear interpolation method to obtain the resampled image with a fixed Image data of spatial resolution; and map the original pixel value range to a preset value range through a linear function.

Specifically, after obtaining the spinal vertebral body image data, preprocessing it can eliminate the negative impact of different spatial resolutions and extreme pixel values of the original image data on subsequent steps.

First, resample the original image data: resampling requires obtaining the spatial resolution of the input spine vertebral body CT image, that is, the physical spatial size corresponding to each voxel. Image data based on the DICOM protocol will save this information as part of the metadata. As a specific embodiment, the spatial resolution of the CT image can be obtained through the image data based on the DICOM protocol. The target spatial resolution of resampling is a certain fixed spatial size, such as 150*90*90 mm. For this purpose, first preset a spatial resolution, such as 1*1*1 mm, divide the spatial resolution of the input CT image by the preset spatial resolution, and obtain the weight of the three dimensions of the original image data. Sampling ratio; as a specific embodiment, a linear interpolation method (trilinear interpolation) can be used to obtain CT image data with a fixed spatial size after resampling.

After resampling the input CT image data, the pixel values of the image need to be normalized, and the original pixel value range [M, N] is mapped to a preset value range [P, Q] through a linear function. , for example [-1,1], where M is the minimum pixel value of the CT image, N is the maximum pixel value of the CT image, P is the lower bound of the preset value range, Q is the upper bound of the preset value range, use (M,P ) and (N, Q) fit the linear function.

It should be noted that after resampling, the image data is represented with a unified spatial resolution, thereby eliminating the structural differences caused by different spatial resolutions (such as data layer thickness and reconstruction methods), making the sub-region segmentation model Focus the learning direction of feature representation on the semantics of the image itself. The purpose of pixel value normalization is to further eliminate the negative impact of extreme pixel values on subsequent steps; for example, some metal implants have abnormally high pixel values in CT images and need to be suppressed by pixel value normalization. The linear interpolation method (trilinear interpolation) is used for normalization processing, which can achieve faster processing speed while retaining image characteristics.

Step 2: Input the pre-processed image data into the pre-trained neural network model to obtain the sub-region segmentation mask regression results corresponding to the spinal vertebrae; wherein, the neural network model is used to output the upper and lower bony endplate masks of the vertebral body. Membrane regression results, cortical bone lateral plate mask regression results, cancellous bone area mask regression results and vertebral body mask regression results based on fusion; the four mask regression results are the prediction output of the neural network model; The four predicted outputs are subjected to regression iterative calculations to obtain the final vertebral upper and lower bony endplate mask regression results, cortical bone lateral plate mask regression results, cancellous bone region mask regression results, and vertebral fusion-based vertebral mask regression results. Volume mask regression results.

Specifically, this application adopts a bifurcated multi-task convolutional neural network structure, as shown in Figure 2. The multi-task learning strategy helps improve the generalization ability of convolutional neural networks and suppress over-fitting. The bifurcated structure consists of an encoder, three decoders and a MAX fusion unit; the encoder is used to receive the preprocessed image data to obtain feature maps; the outputs of the three decoders and a MAX fusion unit It is the four outputs of the model; the three decoders are connected to the encoder respectively, and are used to output the upper and lower bony endplate mask regression results, cortical bone lateral plate mask regression results and vertebral cancellous bone mask based on the feature map. Regression results; the MAX fusion unit is connected to three encoders respectively to output the vertebral mask regression results.

As a specific embodiment, this application draws on the UNet structure and uses skip connections to transfer the feature map of the encoder to three decoders, thereby ensuring that the local feature map can be effectively transferred to the decoder and making up for the information caused by downsampling. lost. In order to further suppress the over-fitting phenomenon of the model, this application uses the MAX fusion unit to perform voxel-level fusion of the outputs of the three decoders; that is, for each voxel position index, the maximum value of the corresponding index of the three decoders is obtained, which is equivalent to completing Cone subregion classification voting based on channel maxima. The three-channel MAX fusion output can restore the complete vertebral mask and filter out invalid feature areas.

In order to train the vertebral body sub-region segmentation neural network model described in this application, three types of training labels need to be constructed for the image data used for training: upper and lower bony endplate mask labels, cortical bone lateral plate mask labels, and vertebral body cancellous Bone mask label. The vertebral mask label can be obtained by fusing the aforementioned three labels in space.

Through the above method, the endplate segmentation area is basically consistent with the anatomical shape of the endplate of different spinal segments, and image segmentation can be performed based on the neural network model; and the endplate segmentation area can achieve a uniform thickness effect. Through morphological calculation, according to the pre-automatic Segmentation with a defined thickness can cover the bony endplate and the bone under the endplate in a custom thickness range; the endplate segmentation area can be eliminated by using a three-dimensional morphological erosion algorithm to eliminate non-interesting interference areas such as osteophytes; upper and lower bony endplates The mask label can be expanded to a customized endplate sub-region shape, and the shape and size of the "implant-endplate" contact surface under the geometric model can be obtained through the pre-designed geometric model of intervertebral implants such as cages. Matching mask labels for superior and inferior bony endplate subregions.

Specifically, the way to create mask labels can be manually annotated by experts, and a binary mask can be used to represent the area of interest, where the label 1 represents the voxels in the area of interest, and the label 0 represents the voxels in the non-interest area. The image data used for training are three-dimensional medical images, including computed tomography, CT, MRI and other imaging equipment to collect three-dimensional data.

Figure 3 is a schematic structural diagram of the neural network model for training vertebral body sub-region segmentation in this application;

First, perform data augmentation on the image data. Specifically, as shown in Figure 4, three-dimensional data augmentation methods include: random exponential transformation and logarithmic transformation of pixel values, random three-dimensional radial transformation (including translation, stretching, shrinkage, rotation, shearing, etc.), random salt and pepper noise Disturbance, random elastic deformation, etc.; through data augmentation, the training data set is greatly enriched and the generalization ability of the model is increased.

Expand the training data set through data augmentation, simulate CT image features acquired in different scenarios, and increase the model's adaptability to image contrast, image noise, vertebral body posture, and image layer thickness, thereby obtaining stronger generalization ability and being able to Applicable to a wider range of clinical scenarios.

Secondly, the expanded training data set is used to iteratively train the model parameters. Specifically, the gradient descent method is used to iterate the model parameters as a whole, by calculating the loss functions of the four outputs of the model and the corresponding labels, and assigning weights to the loss functions of the four outputs, for example, weighting according to the ratio of 1:1:1:1, The total loss function of the model is obtained; the gradient descent method is used for the total loss function of the model, and the model parameters are updated and iteratively optimized at a certain learning rate. This application uses cross entropy and Dice function weighted at 1:1 as the loss function of each output; where,

The cross entropy function is:

The Dice function is:

Among them, a is an output result in the neural network model, b is the labeling result, i is the pixel position index, and |n| is the total number of pixels.

It should be noted that this application greatly enhances the generalization ability of the vertebral body segmentation method through the expansion of the training sample set and iterative parameter training based on the mask label data annotated by experts, and solves the problems that existing technologies have on vertebral body image quality. It has strong dependence, poor generalization ability, and is difficult to adapt to problems such as different CT equipment, different scanning parameters, and different spine morphologies; and this application adds expert opinions to better improve segmentation performance and improve reliability. Expandability.

Those skilled in the art can understand that the loss function and its weight distribution ratio selected in this application are only examples related to the solution of this application and do not constitute a limitation on the application of the solution of this application. The types of loss functions actually used and each loss function The corresponding weights can be changed.

Step 3: Post-process the mask regression results through morphological operations and connectivity tests to obtain a three-dimensional segmentation mask of the upper and lower bony endplates, cortical bone lateral plates, and cancellous bone areas of the vertebral body, and complete the vertebral body subdivision. Region segmentation.

The upper and lower bony endplate mask regression results, cortical bone lateral plate mask regression results, and vertebral body cancellous bone mask regression results output from the vertebral body subregion segmentation neural network model need to be post-processed. Specifically, the three-dimensional morphological opening operation is first used to process each mask regression result to remove the fine-grained noise in the mask; the convolution kernel of the opening operation is 3*3*3, and the essence of the three-dimensional morphological opening operation is The sliding window method is used to morphologically erode and expand the neighborhood around each voxel to remove abnormal bone structure areas including vertebral osteophytes. The schematic diagram of osteophyte removal is shown in Figure 6. Then, use the connectivity test to process the three mask regression results. Specifically, you can use the skimage toolkit to calculate the connectivity of the upper and lower bony endplate masks, and retain the two largest connected areas in the results, corresponding to There are two upper and lower bony endplates; the cortical bone lateral plate mask only retains the largest connected area; the vertebral cancellous bone mask retains only the largest connected area. After morphological operations and connectivity testing, the final vertebral sub-region segmentation mask area can be obtained.

Another embodiment of the present application provides a vertebral body sub-region segmentation device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the present invention is implemented when the processor executes the computer program. Apply the vertebral body sub-region segmentation method described in any of the above embodiments.

The third embodiment of the present application provides a computer-readable storage medium. A computer program is stored on the storage medium. The computer program can be executed by a processor to realize the vertebral subregion described in any of the above embodiments of the present application. Segmentation method.

To sum up, this application provides a vertebral body sub-region segmentation method: As shown in Figure 5, the pre-processed image data is input into the pre-trained vertebral body sub-region segmentation neural network model, and four prediction outputs are obtained. , including sub-region segmentation mask regression results and vertebral body segmentation mask regression results; the mask regression results are post-processed through morphological operations and connectivity tests to obtain the upper and lower vertebral body bony endplates, cortical bone lateral plates, loose The three-dimensional segmentation mask of vertebral body sub-regions such as the bone region is used to complete the segmentation of vertebral body sub-regions. This application uses artificial intelligence technology to promote the application of smart medical care in clinical settings; it solves the shortcomings of traditional methods such as poor generalization ability, poor scalability, and limited application scope. And this segmentation method can be extended to the field of local bone density calculation to calculate the average CT value in the segmented area. Based on the CT value, through the automatic positioning analysis of muscle fat, quantitative calculation and analysis of bone mineral content can be performed, thereby achieving Bone density calculation measurement.

Those skilled in the art can understand that all or part of the process of implementing the method in the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Wherein, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.

The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any person familiar with the technical field can easily think of changes or modifications within the technical scope disclosed in the present application. Replacements shall be covered by the protection scope of this application.

Claims

A vertebral body sub-region segmentation method, characterized by including the following steps:

Preprocess the obtained spinal vertebral body image data;

The pre-processed image data is input into the pre-trained neural network model to obtain the sub-region segmentation mask regression results corresponding to the spinal vertebral body; wherein the neural network model is used to output the upper and lower bony endplates of the vertebral body. Mask regression results, cortical bone lateral plate mask regression results, cancellous bone area mask regression results, and vertebral body mask regression results based on fusion;

The vertebral body mask regression results are post-processed through morphological operations and connectivity tests to obtain a three-dimensional segmentation mask of the upper and lower bony endplates, cortical bone lateral plates, and cancellous bone areas of the vertebral body to complete the vertebral body sub-regions. segmentation;

The vertebral body sub-region includes the upper and lower bony endplates of the vertebral body, cortical bone lateral plates and cancellous bone areas.
The vertebral body sub-region segmentation method according to claim 1, characterized in that the pre-trained neural network model is a bifurcated multi-task convolutional neural network structure, including an encoder, three decoders and a MAX Fusion unit: where,

The encoder is used to receive the preprocessed image data to obtain a feature map;

The three decoders are respectively connected to the encoder, and are respectively used to output upper and lower bony endplate mask regression results, cortical bone lateral plate mask regression results and vertebral cancellous bone mask regression based on the feature map. result;

The MAX fusion unit is connected to the three encoders respectively, and is used to output the vertebral mask regression results.
The vertebral body sub-region segmentation method according to claim 2, characterized in that the pre-training process of the neural network includes:

For the preprocessed image data, upper and lower bony endplate mask labels, cortical bone lateral plate mask labels, vertebral body cancellous bone mask labels are constructed respectively, and the vertebral body mask label is obtained by fusion;

By simulating the image features obtained in different scenarios, the image data after constructing the label is augmented to obtain training samples to expand the training data set;

Calculate the four output results of the neural network model and the loss function of the corresponding label, assign weights to the four output loss functions to obtain the total loss function of the neural network model; use the gradient based on the total loss function of the neural network model The descent method performs iterative training of model parameters.
The vertebral body sub-region segmentation method according to claim 3, characterized in that the voxels in the training sample image are marked using a binary mask method, and the upper and lower bony endplates are constructed for the training sample image. Mask label, cortical bone lateral plate mask label, vertebral cancellous bone mask label;

The vertebral body mask label is constructed by fusing the upper and lower bony endplate mask labels, cortical bone lateral plate mask labels, and vertebral body cancellous bone mask labels.

Wherein, constructing the endplate mask label includes customizing the thickness and shape of the endplate area to obtain the endplate mask label.
The vertebral body sub-region segmentation method according to claim 3 or 4, characterized in that the gradient descent method is used to iterate the model parameters, including:

The weight of cross entropy and Dice function is used as the loss function of each output result, where the cross entropy function is:

The Dice function is:

Among them, a is an output result in the neural network model, b is the labeling result, i is the voxel position index, and |n| is the total prime number.
The vertebral body sub-region segmentation method according to any one of claims 2 to 4, characterized in that the MAX fusion unit is used to output a vertebral body mask regression result, including: based on the preprocessed spinal vertebral body For each voxel position index of the image, take the maximum value of the corresponding indexes of the three decoders, perform MAX fusion on the maximum value of the corresponding indexes of the three decoders for each voxel, and output the vertebral mask regression result.
The vertebral body sub-region segmentation method according to claim 1, wherein the preprocessing of the spinal vertebral body image data includes resampling processing and pixel value normalization processing; wherein,

The resampling process includes: dividing the obtained spatial resolution of the spine vertebral body image by a preset spatial resolution to obtain a resampling ratio of the image data in three dimensions; according to the resampling ratio, using The linear interpolation method obtains resampled image data with a fixed spatial resolution.

The pixel value normalization process includes: mapping the original pixel value range [M, N] to the preset value range [P, Q] through a linear function; where M is the minimum pixel value of the CT image, and N is the maximum CT image. Pixel value, P is the lower bound of the preset value range, Q is the upper bound of the preset value range.
The vertebral body sub-region segmentation method according to claim 1, wherein the post-processing of the mask regression results through morphological operations and connectivity testing includes: using a convolution kernel of 3*3* 3's three-dimensional morphological opening operation uses a sliding window method to morphologically erode and expand the neighborhood around each voxel to remove fine-grained noise in the mask to remove abnormal bone including vertebral osteophytes. Structural area.

Use the skimage toolkit to perform connectivity calculations on the upper and lower bony endplate masks, and retain the two largest connected areas in the results, corresponding to the upper and lower bony endplates respectively; the cortical bone lateral plate mask and the vertebral cancellous The bone mask retains only the largest connected area.
A vertebral body sub-region segmentation device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, which is implemented when the processor executes the computer program The vertebral body sub-region segmentation method according to any one of claims 1 to 8.
A computer-readable storage medium, characterized in that a computer program is stored on the storage medium, and the computer program can be executed by a processor to implement the vertebral body sub-region segmentation described in any one of claims 1-8. method.