Multi-spine structure segmentation method and system based on CT image

Abstract

The invention provides a multi-spine structure segmentation method and system based on a CT image, and belongs to the technical field of medical image segmentation. A to-be-segmented CT image and a plurality of historical CT images are acquired and preprocessed; inputting the historical CT image into the deep neural network for training; inputting a CT image to be segmented into the trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; extracting centrum features of the CT image by using a bone tissue segmentation network to obtain a centrum segmentation image; segmenting a soft tissue structure of the CT image by using a soft tissue enhanced network to obtain an intervertebral disc segmentation image and a nerve segmentation image; and integrating the vertebral body segmentation image, the intervertebral disc segmentation image and the nerve segmentation image to obtain a complete lumbar vertebra structure segmentation image. According to the method, the vertebral body, the intervertebral disc and the nerve in the spine CT image can be efficiently and automatically segmented at the same time.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to a multi-spinal structure segmentation method and system based on CT images. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the aggravation of the aging population and the changes in people's living and working styles, the prevalence of lumbar diseases is gradually increasing. At present, the clinical diagnosis and treatment of lumbar diseases still rely on the analysis and judgment of doctors, which is not only time-consuming but also easily affected by the doctor's subjective experience and state, and can no longer meet the current diagnosis and treatment needs. Therefore, it is very important to develop an automatic segmentation system for lumbar structures.

[0004] The segmentation method of lumbar anatomical structure is mainly based on MRI images, because MRI has strong imaging capabilities for soft tissues. However, the MRI examination time is much longer than CT examination, so patients who want to undergo MRI examination may need to wait for hours or even days, which undoubtedly causes trouble for outpatient patients; the price of MRI examination of the same part is generally higher than that of CT examination, which increases the financial burden of patients; and some special patients cannot undergo MRI examination (such as some patients who have undergone pacemaker implantation or patients who cannot lie flat for a long time due to pain, etc.). Therefore, after weighing the pros and cons, spinal surgeons give priority to fast and relatively cheap CT examinations for outpatient patients with lumbar diseases; if the CT examination shows that there are more obvious lesions in the lumbar spine, MRI examinations are performed. However, CT examinations are not as good as MRI in imaging soft tissues. Diagnosis of lumbar diseases by CT often requires doctors to have rich clinical experience, which places high demands on doctors and increases their burden.

[0005] Therefore, there is an urgent need to establish a new CT image segmentation method to assist radiologists and spinal surgeons in quickly and accurately completing the diagnosis of lumbar diseases only through CT images; and to intuitively understand the location of intervertebral disc herniation and the specific location and severity of compression of the spinal cord and nerve roots in 3D mode, so that spinal surgeons can understand the specific anatomical structure of the lumbar spine of each patient in more detail before and after lumbar surgery, which has important guiding significance for the preoperative and postoperative evaluation of lumbar surgery.

[0006] The inventors found that the existing CT image segmentation methods still have some technical problems, such as:

[0007] (1) For example, the patent application number is 202410283675.5, which is a two-stage AIS spine CT image multi-label segmentation method based on deep learning. This method creates new labels for vertebrae location in the image dataset to train a convolutional neural network for vertebrae location, and locates the vertebrae in the CT image based on the convolutional neural network. It can be seen that the existing CT image segmentation method can only achieve the segmentation of a single spinal structure (vertebrae), but cannot simultaneously segment other adjacent structures.

[0008] (2) Compared with MRI images, the soft tissue resolution of CT images is lower, resulting in a lower degree of differentiation between neural structures and surrounding tissues. This may lead to reduced visibility of details in CT images, increased misidentification rate, and ultimately make it difficult to identify soft tissue areas.

[0009] (3) Bones in CT images may contain artifacts, which may affect the display of surrounding soft tissue structures such as intervertebral discs and nerves. Therefore, the presence of artifacts may cause the neural network to ignore some structural features during learning, which are important for 3D visualization and clinical diagnosis and treatment. Summary of the invention

[0010] To overcome the above-mentioned deficiencies of the prior art, the present invention provides a multi-spinal structure segmentation method and system based on CT images, which can simultaneously and efficiently automatically segment the vertebral bodies, intervertebral discs and nerves in spinal CT images.

[0011] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0012] A first aspect of the present invention provides a multi-spinal structure segmentation method based on CT images.

[0013] A multi-spinal structure segmentation method based on CT images, comprising:

[0014] Acquire a CT image to be segmented and a plurality of historical CT images and form a data set to be segmented and a historical data set respectively; wherein the plurality of historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae;

[0015] Preprocessing the data set to be segmented and the historical data set respectively;

[0016] Inputting a plurality of historical CT images in the preprocessed historical data set into a deep neural network to train the deep neural network;

[0017] Inputting the CT image to be segmented in the data set to be segmented into a trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, using the bone tissue segmentation network to extract vertebral features of the CT image to obtain a vertebral segmentation image; using the soft tissue enhancement network to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image;

[0018] The obtained vertebral body segmentation image, intervertebral disc segmentation image and nerve segmentation image are integrated to obtain a complete lumbar vertebrae structure segmentation image.

[0019] Furthermore, the preprocessing includes: cropping, slice selection, volume resampling and normalization.

[0020] Furthermore, during the cropping process, all image slices are cropped to the same size; slice selection is then performed on all images in the CT image dataset to remove the areas of interest that are not included in the horizontal plane; and the volumes of all images in the image dataset are then resampled to the median voxel spacing of the corresponding dataset.

[0021] Furthermore, the soft tissue enhancement network includes a bone tissue removal module, a grayscale self-adjustment module and a soft tissue segmentation module.

[0022] Furthermore, the bone tissue clearing module performs a zero-minus-one operation on the vertebral features extracted by the bone tissue segmentation network, and then adds them to the initial image to eliminate the influence of the vertebral area in the CT image.

[0023] Furthermore, the grayscale self-adjustment module automatically adjusts and selects the best grayscale value in the CT image through a loss function to enhance the display of soft tissue in the image; and then sends the soft tissue enhanced data to the soft tissue segmentation module.

[0024] Furthermore, the soft tissue segmentation module includes: a coding layer, a bottleneck layer, a decoding layer and a skip connection.

[0025] A second aspect of the present invention provides a multi-spinal structure segmentation system based on CT images.

[0026] A multi-spinal structure segmentation system based on CT images, comprising:

[0027] An image acquisition module is configured to: acquire a CT image to be segmented and a plurality of historical CT images and form a data set to be segmented and a historical data set respectively; wherein the plurality of historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae;

[0028] A preprocessing module is configured to: preprocess the to-be-segmented data set and the historical data set respectively;

[0029] The model training module is configured to: input a plurality of historical CT images in the preprocessed historical data set into a deep neural network to train the deep neural network;

[0030] The CT image segmentation module is configured to: input the CT image to be segmented in the data set to be segmented into a trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, extract the vertebral features of the CT image using the bone tissue segmentation network to obtain a vertebral segmentation image; use the soft tissue enhancement network to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image; integrate the obtained vertebral segmentation image, intervertebral disc segmentation image and nerve segmentation image to obtain a complete lumbar structure segmentation image.

[0031] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method for segmenting multiple spinal structures based on CT images as described in the first aspect of the present invention.

[0032] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a method for segmenting multiple spinal structures based on CT images as described in the first aspect of the present invention are implemented.

[0033] One or more of the above technical solutions have the following beneficial effects:

[0034] (1) The present invention inputs the CT image to be segmented into a trained deep neural network, which is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, the bone tissue segmentation network is used to extract the vertebral features of the CT image to obtain a vertebral segmentation image; the soft tissue enhancement network is used to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image. Therefore, the present invention can simultaneously segment the vertebral body, intervertebral disc and nerve in the CT image, thereby realizing the simultaneous segmentation of multiple spinal structures.

[0035] (2) When the present invention segments the CT image, after extracting the vertebral features of the CT image using the bone tissue segmentation network, the vertebral features and the CT image to be segmented are simultaneously input into the soft tissue enhancement network, and the bone tissue removal module, grayscale self-adjustment module and soft tissue segmentation module in the soft tissue enhancement network are used to achieve feature enhancement and feature extraction in the soft tissue structure. Therefore, the present invention can reduce the self-recognition rate of the network model and enhance the recognition effect of the soft tissue area.

[0036] (3) The present invention uses a 3D U-Net network as the backbone network to effectively reduce artifacts. Therefore, the neural network can focus on more detailed structural features during learning and enhance the segmentation effect of multiple spinal structures in CT images.

[0037] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0039] Figure 1 This is a flow chart of a multi-spinal structure segmentation method based on CT images in Example 1 of the present invention.

[0040] Figure 2 This is an overall framework diagram of a multi-spinal structure segmentation method based on CT images in Example 1 of the present invention.

[0041] Figure 3 This is a framework diagram of the bone tissue segmentation network in Example 1 of the present invention.

[0042] Figure 4 This is a framework diagram of the soft tissue enhancement network in Example 1 of the present invention.

[0043] Figure 5 This is a framework diagram of the grayscale adaptive module in the first embodiment of the present invention.

[0044] Figure 6 It is a schematic diagram for qualitative comparison of segmentation results in the first embodiment of the present invention. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0046] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0047] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0048] Embodiment 1

[0049] This embodiment discloses a multi-spinal structure segmentation method based on CT images.

[0050] like Figure 1 As shown, a multi-spinal structure segmentation method based on CT images includes:

[0051] Step S1, obtaining a CT image to be segmented and a plurality of historical CT images and forming a data set to be segmented and a historical data set respectively; wherein the plurality of historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae;

[0052] Step S2, preprocessing the data set to be segmented and the historical data set respectively;

[0053] Step S3, inputting a plurality of historical CT images in the preprocessed historical data set into a deep neural network, and training the deep neural network;

[0054] Step S4, inputting the CT image to be segmented in the data set to be segmented into a trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, using the bone tissue segmentation network to extract vertebral features of the CT image to obtain a vertebral segmentation image; using the soft tissue enhancement network to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image; integrating the obtained vertebral segmentation image, intervertebral disc segmentation image and nerve segmentation image to obtain a complete lumbar structure segmentation image.

[0055] Based on the above steps, the present invention can simultaneously and efficiently automatically segment the vertebral body, intervertebral disc and nerves in the spinal CT image. In order to facilitate the understanding of the technical solution of the present invention, the above steps are further explained and illustrated, specifically:

[0056] Step S1, obtaining a CT image to be segmented and multiple historical CT images and forming a data set to be segmented and a historical data set respectively; wherein the multiple historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae.

[0057] The historical data set used in this example was collected from Qilu Hospital of Shandong University (Qingdao Campus), and the data in the data set (historical CT images) have been approved by the hospital's data ethics committee. The data in the historical data set includes 200 lumbar CT scan data (from 200 different subjects); and the voxel spacing of these CT images is 0.3467×0.3467×0.999mm 3 .

[0058] Step S2: pre-process the data set to be segmented and the historical data set respectively.

[0059] The dataset to be segmented and the historical dataset are preprocessed respectively, that is, each image in the dataset to be segmented and the historical dataset is sequentially subjected to preprocessing operations such as cropping, slice selection, volume resampling and normalization.

[0060] Step S2-1: During the cropping process, all image slices are cropped to the same size, that is, all image slices are cropped to 256*256, and each data (cropped to 256*256 image slices) is cropped to a non-zero value region.

[0061] In the actual implementation process, on the one hand, the data set may contain some extreme values, which may not be the true features of the data, but noise or erroneous records. Cropping can limit these outliers to a certain range to avoid their influence on model training or data analysis; on the other hand, combined with the normalization method, it can keep the scales of various features consistent; therefore, in order to deal with outliers, outliers, and standardize the data, it is necessary to perform a cropping operation. At the same time, based on the cropping of step S2-1, pixel values ​​that are too bright or too dark can be removed, making the remaining data more representative and usable to ensure the quality of the image data; further, cropping the 512*512 image slices into 256*256 pixel image slices can greatly save GPU memory and thus reduce training time.

[0062] Step S2-2: slice all images in the dataset to be segmented and the historical dataset, and remove the horizontal plane that does not contain the area of ​​interest. First, find the slice range of the area of ​​interest and expand the slice range; then, select the slice after removing the slices that are too small.

[0063] Step S2-3: resample the volumes of all images in the dataset to be segmented and the historical dataset to the median voxel spacing of the corresponding datasets to reduce the inconsistency between different volumes.

[0064] The SimpleITK library (an open source library for medical image processing) is used to process all images in the dataset to be segmented and the historical dataset, and the voxel spacing of the dataset is unified by calculating the median voxel spacing. Each image is resampled according to the target voxel spacing to ensure that all images have consistent spatial resolution. Specifically: first, the voxel spacing of the image is obtained; then, the median voxel spacing of all images in the image dataset is calculated; then, the image is resampled to the target voxel spacing, and the SimpleITK ResampleImageFilter is used for interpolation resampling; then, each image in the dataset is resampled and the image file is read; finally, the resampled image is saved.

[0065] Step S2-4: perform volume intensity trimming by normalization. Specifically, the normalization formula is expressed as:

[0066]

[0067] Where I′ represents the normalized image intensity, I represents the cropped image intensity, and I 2％ Represents the 2% percentile of the cropped image intensity, I 98％ I represents the intensity of the cropped image 98％ Percentile. Thus, the intensity values ​​of the volume are clipped to the percentile of the entire volume intensity value range [2.0, 98.0].

[0068] Step S3: input multiple historical CT images in the preprocessed historical data set into the deep neural network to train the deep neural network.

[0069] The historical CT images are input as training data into the deep neural network to train the deep neural network model. The entire training process is implemented through iteration. Before iterative training of the model, the number of epochs, the initial learning rate, and the weight decay value are pre-set. Among them, epoch is the unit for measuring the progress of model training, which refers to the process in which the entire training data is traversed by the model once. In other words, when an epoch is completed, it means that the model has traversed each CT image data in the training set (historical data set) at least once. When starting to train a deep neural network, the historical data set can be divided into multiple small batches, each of which contains a certain number of CT images. In each iteration, the model processes the samples in a batch, calculates the loss and gradient based on these samples, and then updates the weights of the model. This process is repeated until all batches are processed, that is, an epoch is completed. In addition, since a single epoch may not be enough for the model to fully learn the characteristics of the data, this embodiment chooses to train the model for multiple epochs; at the same time, too many epochs may also lead to overfitting, that is, the model is too complex, so that the model begins to remember the noise and details of each sample in the historical data set, rather than learning the general rules of the data. Therefore, it is necessary to select an appropriate number of epochs for deep neural network training.

[0070] Furthermore, in this embodiment, the number of epochs is set to 30, and each epoch is iterated 200 times; at the same time, the initial learning rate is set to 0.0001 and the weight decay value is set to 0.0001. If the model performance has not improved after 30 epochs, the training is terminated.

[0071] Step S4: Input the CT image to be segmented into the trained deep neural network. Figure 2 As shown, the deep neural network consists of a bone tissue segmentation network and a soft tissue enhancement network; specifically, the bone tissue segmentation network is used to extract the vertebral features of the CT image to obtain a vertebral segmentation image; the soft tissue enhancement network is used to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image; the obtained vertebral segmentation image, intervertebral disc segmentation image and nerve segmentation image are integrated to obtain a complete lumbar structure segmentation image. Specifically, step S4 can be implemented by the following process, namely:

[0072] Step S4-1: Input the original 3D CT volume to be segmented into the bone tissue segmentation network BTS-Net, and obtain more accurate vertebral features after simple bone tissue segmentation. The bone tissue segmentation network can divide the obtained vertebral features into two parts (the two cone features are exactly the same), one of which is output as the final result (i.e., vertebral image), and the other is input into the soft tissue enhancement network (STE-Net).

[0073] like Figure 3 As shown, due to the outstanding performance of U-Net in medical image segmentation, this embodiment uses 3D U-Net as the backbone network, the main purpose of which is to obtain more accurate vertebral features. Since 3D U-Net can operate in three dimensions, it can better capture and retain spatial information in the image, and at the same time can better handle the boundaries of the object and reduce artifacts and breaks in the segmentation results. This helps to improve the accuracy and stability of segmentation, especially in CT images with complex structures or noise. In addition, accurate segmentation results will help in the next step of soft tissue enhancement.

[0074] Step S4-2, the soft tissue enhancement network includes a bone tissue removal module, a grayscale self-adjustment module and a soft tissue segmentation module. The original 3D CT volume is input into the soft tissue enhancement network (STE-Net).

[0075] In order to more effectively segment low-resolution soft tissue in CT image data, this embodiment proposes a soft tissue enhancement network (STE-Net), which consists of three parts: a bone tissue removal module (Elimination of bonetissue), a gray self-adjusting module (Gray self-adjusting module) and a soft tissue segmentation network. In this embodiment, the vertebral features obtained in BTS-Net and the original CT image are first sent to the bone tissue removal module (Elimination of bonetissue) to make the gray value of the vertebral area negative. Then, the gray self-adjusting module (GSAM) automatically adjusts and selects the best gray value in the CT image through the loss function, thereby enhancing the display of soft tissue in the image, and then the soft tissue enhanced data is sent to the backbone network U-Net, and the segmentation results of nerves and intervertebral discs are output. Finally, the segmentation results of BTS-Net and STE-Net are fused to obtain a complete segmentation. It can be implemented specifically through the following process:

[0076] Step S4-2-1, input the original 3D CT volume into the soft tissue enhancement network (STE-Net), and clear the vertebral area through the bone tissue removal module (Elimination of bone tissue) together with the vertebral feature information that has been input. This can reduce the complexity and difficulty of the task, while saving computing resources and improving efficiency.

[0077] Specifically, in order to eliminate the influence of the vertebral area during the soft tissue segmentation process, this embodiment introduces a bone tissue removal module, which performs a zero-minus-one operation on the vertebral features extracted by the bone tissue segmentation network, and then adds them to the initial image to eliminate the influence of the vertebral area in the CT image, thereby improving the network's segmentation of nerves and intervertebral discs. Bone tissue removal can be expressed as:

[0078] x b =BTE-Net(x);

[0079] x e =x+x·arg(x b )-1;

[0080] Among them, x is the original input, x b is the output of the BTS-Net network, arg() and x e They are the outputs of the argmax function and the bone tissue removal module, respectively.

[0081] Step S4-2-2: Use the Gray self-adjusting module (GSAM) to automatically convert the bone window into a soft tissue window to discover more soft tissue structure details, thereby enhancing the network's learning ability for soft tissue and improving the accuracy of the lumbar structure segmentation task.

[0082] like Figure 5 As shown in FIG, the grayscale self-adjusting module automatically adjusts and selects the best grayscale value in the CT image through the loss function to enhance the display of soft tissue in the image; then the soft tissue enhanced data is sent to the soft tissue segmentation module. Specifically, the grayscale self-adjusting module adjusts the window width w in the CT image. w and window level w c To achieve the selection of appropriate grayscale values. Among them, the window width w w Determines the grayscale value range covered by the window, the window position w c Corresponding to the CT value at the center of the window. This embodiment sets the initial window width w based on experience. w and window level w c are (40, 350) respectively; in network training, the window width w in the CT image is adjusted by learning the two parameters α and β w and window level w c .like Figure 4 As shown in the figure, the original grayscale value of the 3D CT image (x) is used as input, and the data after adjusting the grayscale value distribution is the output data (x′) to improve the contrast of soft tissue structure and improve the network's learning ability of nerve and intervertebral disc structure. Specifically, first, a 1×1×1 convolution kernel with a parameter of θ is used to perform a convolution operation on the original input x to generate a c-channel feature map f, that is:

[0083] f = Conv(x,θ);

[0084] Among them, θ represents the convolution parameter and Conv() represents the convolution operation.

[0085] Then, the feature map f is sent to two parallel branches, each of which learns adaptive parameters α and β to adjust the grayscale values ​​w and c of the window width and window position; in each branch, a 1×1×1 convolution is performed on the feature map f to generate a 1-channel feature map, followed by a global average pooling (GAP) operation, and the value is compressed to the [0,1] range through the Sigmoid activation function. The adaptive parameters α and β are expressed as follows:

[0086] α=Sigmoid(GAP(Conv(f,θ w )))+t;

[0087] β=Sigmoid(GAP(Conv(f,θ c )))+t;

[0088] Among them, Sigmoid() represents the activation function, GAP() represents the global average pooling operation, θ w represents the convolution parameter of the adaptive parameter α, θ c represents the convolution parameter of the adaptive parameter β, and t represents the hyperparameter.

[0089] This embodiment adds a hyperparameter t (usually set to 0.5) to ensure that the values ​​of α and β are in the range of [0.5, 1.5] to avoid grayscale value ranges that are not suitable for segmentation tasks. Use the learned parameters α and β to adjust the original w w and w c Scaling is performed to obtain the adjusted w′ w and w′ c ,Right now:

[0090] w′ w =α·w w ;

[0091] w′ c =β·w c ;

[0092] Among them, w w The grayscale value representing the window width, w c Represents the gray value of the window level, w′ w Represents the grayscale value of the adjusted window width, w′ c Represents the grayscale value of the adjusted window level.

[0093] Finally, the deep neural network performs a windowing operation on the original input x, including normalization and thresholding. The normalization operation is based on the adjusted w′ w and w′ c The pixel values ​​are mapped to the range of [0,1], and then the values ​​are limited to [0,255] through the threshold operation to obtain the grayscale adjusted image x′, that is:

[0094] y = Normalization(x) = (xw′ c ) / (w′ w -w′ c );

[0095] x′=Threshold(y)=min(1,max(0,y))·255;

[0096] Among them, y represents the intermediate process quantity after normalization operation, Normalization() represents the normalization operation; x′ represents the image after grayscale adjustment, and Threshold() represents the threshold operation.

[0097] Step S4-2-3: Automatically segment the soft tissue structure in the soft tissue segmentation module to obtain segmented images of the intervertebral disc and nerves.

[0098] The soft tissue segmentation module includes: encoding layer, bottleneck layer, decoding layer and skip connection.

[0099] The soft tissue structure segmentation network is considered as a variant of 3D U-Net, which retains similar structures and functions, but differs in specific implementation and hierarchical details. Specifically, in the encoding stage, compared with the standard 3D U-Net, this embodiment introduces the normalization technology, the rectified linear unit, and the maximum pooling layer in sequence after the 3×3×3 convolution; among them, the normalization technology is BatchNorm3d, the rectified linear unit is ReLU, and each dimension of the maximum pooling layer uses 1×2×2 maximum pooling. This embodiment helps to improve the stability and training speed of the network by using normalization technology and a smaller pooling stride; in addition, the smaller pooling stride can enhance the stability and convergence speed of the network while retaining more spatial details. After the encoding stage, the next layer is the bottleneck layer, which is used to map high-dimensional features to low-dimensional space, thereby reducing the computational cost while maintaining the expressiveness of the features. In the last layer, a 1×1×1 convolution is used to reduce the number of output channels to the number of labels, which is 3 in this embodiment.

[0100] Step S4-2-4: Integrate the acquired vertebral body, intervertebral disc and nerve segmentation images to obtain a complete lumbar structure segmentation result.

[0101] In this embodiment, the integration operation is implemented based on the cat operation under the torch package in Python. Figure 6The schematic diagram for qualitative comparison of the segmentation results of the present invention is shown. It can be seen from the figure that the present embodiment realizes the automatic multi-class segmentation of vertebrae, intervertebral discs and nerves based on CT images by adopting a cascade structure of first segmenting bone tissue and then segmenting soft tissue. Specifically: in order to obtain more accurate vertebral features, a lightweight bone tissue segmentation network (BTS-Net) similar to 3D U-net is innovatively proposed to perform simple bone tissue segmentation on the original CT data; in order to realize automatic enhancement and extraction of soft tissue features, a soft tissue enhancement network (STE-Net) is innovatively proposed; STE-Net first removes the vertebral area in the vertebral features through the bone tissue removal module; then the grayscale self-adjustment module (GSAM) is used to enhance the network's learning ability for soft tissue, thereby improving the accuracy of the lumbar anatomical segmentation task; finally, the intervertebral disc and nerves are segmented through the soft tissue segmentation module. Since the grayscale adaptive module (GASM) can automatically adjust the window width and window level of the CT data to the appropriate soft tissue window, the contrast of the soft tissue structure is improved, and the recognition and segmentation ability of the new segmentation network for soft tissue is enhanced; in addition, the bone tissue removal module (EBT) and the grayscale adaptive module (GASM) can produce a complementary effect when used simultaneously, and the optimal segmentation results can be obtained under various segmentation conditions. In summary, the new segmentation network proposed in the present invention has significantly better segmentation performance for vertebrae, intervertebral discs and nerves than other classical methods, and the present invention can simultaneously and efficiently segment vertebrae, intervertebral discs and nerves in spinal CT images.

[0102] Based on the above process, the segmentation algorithm of the present invention can more quickly and effectively segment the vertebrae, nerves and intervertebral discs from spinal CT medical images; with the assistance of the new segmentation network, doctors can quickly and accurately complete the diagnosis of lumbar diseases only through CT images. And the accuracy of the segmented tissue is higher. This method of segmenting multiple spinal structures based on CT images can enable doctors to have a clear enough understanding of the relative position of the lumbar diseased area before surgery, which can assist in clinical disease diagnosis.

[0103] Embodiment 2

[0104] This embodiment discloses a multi-spinal structure segmentation system based on CT images.

[0105] A multi-spinal structure segmentation system based on CT images, comprising:

[0106] An image acquisition module is configured to: acquire a CT image to be segmented and a plurality of historical CT images and form a data set to be segmented and a historical data set respectively; wherein the plurality of historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae;

[0107] A preprocessing module is configured to: preprocess the to-be-segmented data set and the historical data set respectively;

[0108] The model training module is configured to: input a plurality of historical CT images in the preprocessed historical data set into a deep neural network to train the deep neural network;

[0109] The CT image segmentation module is configured to: input the CT image to be segmented in the data set to be segmented into a trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, extract the vertebral features of the CT image using the bone tissue segmentation network to obtain a vertebral segmentation image; use the soft tissue enhancement network to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image; integrate the obtained vertebral segmentation image, intervertebral disc segmentation image and nerve segmentation image to obtain a complete lumbar structure segmentation image.

[0110] Embodiment 3

[0111] The purpose of this embodiment is to provide a computer-readable storage medium.

[0112] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps in a method for segmenting multiple spinal structures based on CT images as described in the first embodiment of the present disclosure.

[0113] Embodiment 4

[0114] The purpose of this embodiment is to provide an electronic device.

[0115] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in a method for segmenting multiple spinal structures based on CT images as described in Embodiment 1 of the present disclosure are implemented.

[0116] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0117] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0118] 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 on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A multi-spinal structure segmentation method based on CT images, characterized in that: include: Acquire a CT image to be segmented and a plurality of historical CT images and form a data set to be segmented and a historical data set respectively; wherein the plurality of historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae; Preprocessing the data set to be segmented and the historical data set respectively; Inputting a plurality of historical CT images in the preprocessed historical data set into a deep neural network to train the deep neural network; Inputting the CT image to be segmented in the data set to be segmented into a trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, using the bone tissue segmentation network to extract vertebral features of the CT image to obtain a vertebral segmentation image; using the soft tissue enhancement network to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image; The obtained vertebral body segmentation image, intervertebral disc segmentation image and nerve segmentation image are integrated to obtain a complete lumbar vertebrae structure segmentation image.

2. The method for segmenting multiple spinal structures based on CT images according to claim 1, characterized in that: The preprocessing includes: cropping, slice selection, volume resampling and normalization.

3. The method for segmenting multiple spinal structures based on CT images as claimed in claim 2, characterized in that: During the cropping process, all image slices are cropped to the same size; slice selection is then performed on all images in the CT image dataset to remove the areas of interest that are not included in the horizontal plane; and the volumes of all images in the image dataset are then resampled to the median voxel spacing of the corresponding dataset.

4. The method for segmenting multiple spinal structures based on CT images according to claim 1, characterized in that: The soft tissue enhancement network includes a bone tissue removal module, a grayscale self-adjustment module and a soft tissue segmentation module.

5. The method for segmenting multiple spinal structures based on CT images as claimed in claim 4, characterized in that: The bone tissue removal module performs a zero-minus-one operation on the vertebral features extracted by the bone tissue segmentation network, and then adds them to the initial image to eliminate the influence of the vertebral area in the CT image.

6. The method for segmenting multiple spinal structures based on CT images as claimed in claim 4, characterized in that: The grayscale self-adjustment module automatically adjusts and selects the best grayscale value in the CT image through the loss function to enhance the display of soft tissue in the image; and then sends the soft tissue enhanced data to the soft tissue segmentation module.

7. The method for segmenting multiple spinal structures based on CT images as claimed in claim 4, characterized in that: The soft tissue segmentation module includes: a coding layer, a bottleneck layer, a decoding layer and a skip connection.

8. A multi-spinal structure segmentation system based on CT images, characterized in that: include: An image acquisition module is configured to: acquire a CT image to be segmented and a plurality of historical CT images and form a data set to be segmented and a historical data set respectively; wherein the plurality of historical CT images include spinal CT images of patients with lumbar spondylosis and spinal CT images of normal lumbar vertebrae; A preprocessing module is configured to: preprocess the to-be-segmented data set and the historical data set respectively; The model training module is configured to: input a plurality of historical CT images in the preprocessed historical data set into a deep neural network to train the deep neural network; The CT image segmentation module is configured to: input the CT image to be segmented in the data set to be segmented into a trained deep neural network, wherein the deep neural network is composed of a bone tissue segmentation network and a soft tissue enhancement network; specifically, extract the vertebral features of the CT image using the bone tissue segmentation network to obtain a vertebral segmentation image; use the soft tissue enhancement network to segment the soft tissue structure of the CT image based on the CT image and the obtained vertebral features to obtain an intervertebral disc segmentation image and a nerve segmentation image; integrate the obtained vertebral segmentation image, intervertebral disc segmentation image and nerve segmentation image to obtain a complete lumbar structure segmentation image.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the method for segmenting multiple spinal structures based on CT images as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the multi-spinal structure segmentation method based on CT images as described in any one of claims 1 to 7 are implemented.

Images