Spine segmentation method and device, medium and program product

Through the coarse positioning, centroid positioning and Gaussian distribution sub-segmentation methods of spine CT images, the problems of difficult and low accuracy of spine segmentation are solved, and efficient and accurate spine segmentation is achieved, which is suitable for clinical applications.

CN119006496AActive Publication Date: 2024-11-22VISUAL3D MEDICAL TECH DEV CO LTD
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411146400.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-20
Publication Date
2024-11-22
Estimated Expiration
2044-08-20

AI Technical Summary

Technical Problem

The prior art is difficult to segment spinal CT images. Doctors need to spend more than half an hour to outline the spinal area. It is difficult to achieve precise positioning and segmentation based on neural networks, especially due to the complex structure of the spinal column and the large individual differences.

Method used

By first coarsely positioning the spinal area, and then subdividing it by using centroid positioning and Gaussian distribution, the difficulty of segmentation of the spinal bones is reduced and the segmentation accuracy is improved. Specific steps include obtaining three-dimensional CT images, cropping and corroding the vertebrae, positioning the centroid, segmenting the vertebrae using a Gaussian distribution, and identifying categories through a classifier.

Benefits of technology

It achieves the accuracy and efficiency of spinal segmentation, reduces the difficulty of segmentation, is suitable for large-scale clinical applications, and improves the segmentation accuracy of spinal images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006496A_ABST
    Figure CN119006496A_ABST
Patent Text Reader

Abstract

The invention provides a spine segmentation method and device, a medium and a program product, and relates to the field of intelligent medical treatment. The method comprises the following steps: acquiring a three-dimensional CT image of a spine, and segmenting the three-dimensional CT image to obtain an N-value mask image of the whole spine; cutting a single vertebral body in the N-value mask image of the whole vertebra to obtain an N-value mask image in which adjacent single vertebral bodies are not connected; positioning the center of mass of a single cone in the unconnected N-value mask image between the adjacent single cones to obtain an image with the positioned center of mass; and taking the single centroid as a center, and segmenting the image with the positioned centroid to obtain a single vertebra image corresponding to the single centroid. According to the method, the spine area is coarsely positioned firstly, and then different spine bone areas are finely positioned for segmentation, so that the segmentation difficulty of the whole spine bone is reduced, and the segmentation precision is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical treatment, and more specifically, to a spinal segmentation method, device, medium and program product. Background Art

[0002] In the medical field, medical image processing can usually be combined to achieve partial guided surgery, and image segmentation is the basis for subsequent medical processing, which includes: image segmentation, feature extraction, quantitative analysis and three-dimensional reconstruction in order according to the CT images of normal tissue or diseased tissue. Different from the direct segmentation of tumors in tumor diseases, for spinal surgery, due to the complex structure of the spine, which includes multiple vertebrae, pedicles, lamina, vertebral foramina, transverse processes, spinous processes, etc., it is quite difficult to segment the spinal CT images.

[0003] In the prior art, when a machine learning algorithm based on a neural network is implemented, a large amount of training set sample data is usually used to train the neural network, that is, the accuracy of the neural network model directly depends on the amount of training set sample data. In order to achieve the segmentation effect of each component unit of the spine, each component unit of the spine in the sample data is marked with a category label. However, in actual operation, it requires doctors to outline the spinal area layer by layer on hundreds of layers of two-dimensional images, which causes doctors to spend an average of more than half an hour to outline the spinal area of ​​a patient. Therefore, there is an urgent need for a spinal segmentation method that is less difficult and can ensure spinal positioning and segmentation accuracy. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention provides a spine segmentation method, device, medium and program product; the method of the present invention reduces the difficulty of segmenting the entire spine and improves the accuracy of segmentation by first roughly locating the spine area and then finely locating different spinal bone areas for segmentation.

[0005] The first aspect of the present application discloses a spine segmentation method, comprising:

[0006] 101, acquiring a three-dimensional CT image of a spine, and segmenting the three-dimensional CT image to obtain an N-value mask image of the entire spine;

[0007] 102, cropping the single vertebrae in the N-value mask image of the entire spine to obtain an N-value mask image in which adjacent single vertebrae are not connected;

[0008] 103, locating the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebrae, and obtaining an image of the located centroid;

[0009] 104 , taking the single centroid as the center, segmenting from the image of the located centroid to obtain a single vertebra image corresponding to the single centroid.

[0010] In some embodiments, the method of cutting out the single vertebral region in the N-value mask image of the entire spine in 102 includes: cutting out gradually from the edges of the single vertebral body to the center; preferably, eroding from the edges of the single vertebral body to the center in equal proportions;

[0011] Optionally, the volume of the single vertebral body after trimming in 102 is smaller than the volume of the single vertebral body before trimming;

[0012] Preferably, the ratio of the volume of the single vertebral body after cutting to the volume of the single vertebral body before cutting is: 1 / 10-9 / 10, preferably 1 / 5-4 / 5, preferably 1 / 2;

[0013] Optionally, the coordinates of the centroid are three-dimensional coordinates.

[0014] Optionally, the N-value mask is a binary mask.

[0015] In some embodiments, the method of locating the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebra in 103 comprises: using a connected domain algorithm to identify at least one separate region in the N-value mask image that is not connected between adjacent single vertebra; locating the centroid of the separate region is the centroid of the single vertebra; the separate region is the region where the single vertebra is located;

[0016] Optionally, when locating the centroid of a single vertebra in 103 , the vertebrae are sorted according to the z-axis coordinate of the centroid.

[0017] Optionally, the method for acquiring the single vertebral image in 104 includes:

[0018] Obtain a 3D Gaussian distribution with a fixed a×b×c value centered on the single centroid;

[0019] Get a 3D image with a fixed size of a×b×c centered at the center of mass;

[0020] According to the 3D Gaussian distribution centered at the centroid, a single vertebral image corresponding to the centroid around the Gaussian distribution is segmented from the 3D image.

[0021] In some embodiments, after 101 and before 102, the method further includes: cropping the N-value mask image of the vertebra to obtain an N-value mask image containing the target vertebra;

[0022] Optionally, the method for cropping the N-value mask image of the spine includes: finding the maximum and minimum values ​​of the N-value spine mask in three directions of the RAS three-dimensional space, and cropping all areas in the original CT that are less than the minimum value and greater than the maximum value in the three directions, thereby obtaining an N-value mask image containing the target spine.

[0023] In some embodiments, after 104, the method further includes 105, identifying a category of a single vertebra in the vertebral image based on the position of the centroid;

[0024] Optionally, when the category of a single vertebra cannot be determined according to the method in 105, the method further comprises: inputting the image of the single vertebra into a classifier to identify the category of the single vertebra;

[0025] Optionally, the training method of the classifier includes: inputting a single vertebral image of a training set sample into the classifier, obtaining a predicted classification result, comparing it with the classification label, and optimizing it according to the comparison result to obtain a constructed classifier;

[0026] Optionally, the classifier includes any one or more of the following: decision tree, logistic regression, naive Bayes, neural network; preferably DenseNet in the neural network.

[0027] In some embodiments, after 105, the method further includes 106, synthesizing at least two single vertebrae based on the categories of the single vertebrae to obtain a synthesized image including the target vertebrae.

[0028] In some embodiments, after acquiring the three-dimensional CT image and before segmenting the spinal image in 101, the method further includes: preprocessing the three-dimensional CT image to obtain a unified three-dimensional CT image; the preprocessing includes: rotating the three-dimensional CT image to the RAS direction, randomly flipping, normalizing, and randomly shearing the image to a size of 128×128×128.

[0029] A second aspect of the present application discloses a computer device, the device comprising: a memory and a processor;

[0030] The memory is used to store a computer program; the processor executes the computer program to implement the steps of the method described in the first aspect of the present application.

[0031] The third aspect of the present application discloses a computer-readable storage medium, characterized in that a computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method described in the first aspect of the present application are implemented.

[0032] A fourth aspect of the present application discloses a computer program product, including a computer program, characterized in that when the computer program is executed by a processor, the steps of the method described in the first aspect of the present application are implemented.

[0033] This application has the following beneficial effects:

[0034] 1. The present application innovatively discloses a spinal segmentation method. Due to the complex structure of the spinal column, which includes the spine and muscles, the spine is composed of arranged vertebrae, and each vertebra includes a vertebral body, a pedicle, a lamina, a vertebral foramen, a transverse process, a spinous process, a collateral process, a superior articular process, and a mastoid process. Therefore, it is quite difficult to segment a spinal CT image. The method disclosed in the present application can solve the above problem well. It first preliminarily locates the spinal region in the three-dimensional image, reduces the region, and reduces the influence of the region outside the spinal region on the subsequent segmentation. By corroding a single vertebra in the N-value mask image of the entire spine, adjacent vertebrae are not connected. Then, the connected domain algorithm is used to determine the different regions where the single vertebrae are located after corrosion, and the centroid of the single vertebrae in different regions is located as the centroid of the entire vertebra. A Gaussian distribution centered on the centroid is used to restrict the network to only segment the vertebrae of the centroid, thereby achieving fine segmentation of the spine, thereby further obtaining the entire spine. The above method performs bone segmentation through a three-step method of coarse positioning, centroid positioning, and fine segmentation, thereby reducing the difficulty of segmentation while ensuring segmentation accuracy.

[0035] 2. The present application overcomes the difficulties faced in segmenting spinal images and tumor images. Unlike the direct segmentation of tumor images, the spine is a special chain structure composed of a large number of similar units, and there are large individual differences in spinal images. Segmentation requires annotation, learning and training of multiple views of the spine. The process is cumbersome and difficult, which leads to the difficulty in implementing image segmentation based on neural network machine learning algorithms. The existing technology makes it difficult to achieve accurate positioning and segmentation of the various units that make up the spine.

[0036] 3. The spinal images acquired by the segmentation method in this application have high accuracy and low requirements for spinal sequence images, which is conducive to large-scale clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0038] Figure 1 is a schematic diagram of a method flow chart provided by the first aspect of an embodiment of the present invention;

[0039] Figure 2 is a schematic diagram of a spine segmentation system provided by an embodiment of the present invention;

[0040] Figure 3 is a schematic diagram of a computer device provided by an embodiment of the present invention;

[0041] Figure 4 is a schematic diagram of the architecture of an exemplary computing device provided by an embodiment of the present invention;

[0042] Figure 5 is a schematic diagram of a storage medium provided by an embodiment of the present invention;

[0043] Figure 6 It is an operation flow chart of the segmentation method (three-step method, first coarse positioning, then positioning the centroid, and finally fine segmentation) provided by an embodiment of the present invention; a is coarsely segmented and output as b, b is cropped and its corresponding three-dimensional CT image c is output, c is segmented and output as d, the centroid of a single vertebra in d is identified and output as e, the single vertebra is cropped with the centroid as the center to obtain a column of images f, the column of images f is ID-identified to identify the category class of the single vertebra (column of images g), and the column of images g is synthesized to obtain image h;

[0044] Figure 7 Schematic diagram of the rough segmentation process (segmenting the input three-dimensional CT image into a three-dimensional binary mask image) in the three-step method provided by an embodiment of the present invention; A is the input three-dimensional CT image, and B is the output three-dimensional binary mask image; A is outputted as result B after passing through the segrenet segmentation network;

[0045] Figure 8 The embodiment of the present invention provides a method for dividing an input edge-cropped 3D CT image into a 3D binary mask image containing only vertebrae and without connecting adjacent vertebrae; A is the input edge-cropped 3D CT image, and B is the output 3D binary mask image containing only vertebrae with half of the volume eliminated; A is subjected to a segrenet segmentation network to output the result B;

[0046] Fig. 9 1 is a flow chart of a method for acquiring a single vertebral image corresponding to a single centroid provided by an embodiment of the present invention; A is an input 3D Gaussian distribution with a fixed size of 224x224x128 with the vertebral body as the center, B is an input cropped 3D CT image with a fixed size of 224x224x128 with the vertebral body as the center, and C is an output 3D binary mask image of the vertebral body; A and B are segmented by a segrenet network to output the result C;

[0047] Fig.10It is a flow chart of a method for identifying the category of a single vertebra in the vertebral image based on the position of the center of mass provided by an embodiment of the present invention; on the left is an input cropped 3D CT image of 224x224x128 fixed with the vertebra as the center, which is output as the position label class of the vertebra after passing through the DenseNet classification network. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0049] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0051] Figure 1 1 is a flow chart of a spine segmentation method provided by an embodiment of the present invention. Specifically, the method comprises the following steps:

[0052] 101: Acquire a three-dimensional CT image of the spine, and segment the three-dimensional CT image to obtain an N-value mask image of the entire spine;

[0053] In some embodiments, since the input image size is much larger than the spinal region during spinal segmentation, coupled with the high similarity between spinal vertebrae, it is difficult to directly segment individual vertebrae in each spine. In order to improve the segmentation accuracy, the present application adopts a method of first roughly locating the spinal region and then finely segmenting each unit of the spinal component to achieve accurate segmentation of the spine. Specifically, the spinal image in the three-dimensional CT image is binary, that is, the differences between the components of the spine are not distinguished in the coarse positioning step, and the entire spine is regarded as the same label. After segmentation, the "retain the maximum connected domain" post-processing operation is added; the "retain the maximum connected domain" post-processing can help reduce interference such as the skull. The purpose of coarse positioning: to find the maximum and minimum values ​​through the binary region. In the coarse positioning step, the image has impurities that have a great impact on the cropping of the fine segmentation step; however, the image lacks a certain part in the coarse positioning step, which has a relatively small impact on the cropping of the fine segmentation step. The entire spine in this application refers to the human spine including the arranged vertebrae.

[0054] In some embodiments, the N-value mask is a binary mask. When it is a binary mask, the three-dimensional CT image is segmented in 101 to obtain a binary mask image containing the entire spine, or any one of the cervical vertebrae, thoracic vertebrae, and lumbar vertebrae; when it is a four-value mask, the three-dimensional CT image is segmented in 101 to obtain a four-value mask image containing all the cervical vertebrae, thoracic vertebrae, and lumbar vertebrae.

[0055] In some embodiments, after 101 and before 102, the method further includes: cropping the N-value mask image of the spine to obtain an N-value mask image containing a target spine; optionally, the target spine is a spine that needs to be displayed according to demand, including any one or more of the following: cervical vertebrae, thoracic vertebrae, lumbar vertebrae, and the entire spine; specifically, the method of cropping the N-value mask image of the spine includes: finding the maximum and minimum values ​​of the N-value spine mask in three directions in the RAS three-dimensional space, and cropping all areas in the original CT that are less than the minimum value and greater than the maximum value in the three directions, thereby obtaining an N-value mask image containing the target spine.

[0056] 102: cropping a single vertebra in the N-value mask image of the entire spine to obtain an N-value mask image in which adjacent single vertebrae are not connected;

[0057] In some embodiments, the method of cutting out the single vertebral region in the N-value mask image of the entire spine in 102 includes: cutting out gradually from the edges of the single vertebral body to the center; preferably, eroding from the edges of the single vertebral body to the center in equal proportions;

[0058] Optionally, the volume of the single vertebral body after trimming in 102 is smaller than the volume of the single vertebral body before trimming;

[0059] Preferably, the ratio of the volume of a single vertebra after cropping to the volume of a single vertebra before cropping is: 1 / 10-9 / 10, preferably 1 / 5-4 / 5, preferably 1 / 2. The reason for cropping a single vertebra is: if the original size of the vertebra is predicted, there is a high probability that the vertebrae will be connected. If they are connected, the next step using the connected domain algorithm cannot distinguish them. The cropping corrosion in this application changes the size of the vertebra to prevent the connection between the vertebrae. Each vertebra is corroded by the corrosion algorithm to retain half of its volume (because it is impossible for two half vertebrae to be connected together, the half vertebrae are not corroded). By predicting a spine with only half the volume, it is ensured that the vertebrae are not connected. Then, each vertebra is obtained by the connected domain algorithm, and the center of mass of the spine is used as the center of mass of the entire spine. Although the preferred ratio is 1 / 2, in actual operation, the cropping ratio is adjusted according to the shape of the subject's spine, as long as it can ensure that adjacent vertebrae are not connected.

[0060] In some embodiments, a erosion algorithm is used in the training phase to train the clipping process to obtain a region of a single vertebra.

[0061] 103, locating the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebrae, and obtaining an image of the located centroid;

[0062] In some embodiments, the method for locating the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebrae in 103 includes: using a connected domain algorithm to identify at least one separate area in the N-value mask image that is not connected between adjacent single vertebrae; locating the centroid of the separate area is the centroid of the single vertebra; the separate area is the area where the single vertebra is located; first locating different areas, and then finding the centroid of different areas.

[0063] Optionally, when locating the centroid of a single vertebra in 103, the centroids are sorted according to the z-axis coordinates of the centroids. Because the sub-regions obtained by the connected domain algorithm are chaotic, the centroids are identified after being sorted according to the z-axis centroid coordinates.

[0064] In some embodiments, the coordinates of the center of mass are three-dimensional coordinates, and the role of the center of mass is to cut out a single vertebra and Gaussian distribution to segment the spine.

[0065] In some embodiments, the center of mass of a single vertebral body is used as the center of mass of a single vertebra, and the spine includes 24 vertebrae, ie, contains 24 centers of mass.

[0066] 104: Taking the single centroid as the center, segmenting from the image of the located centroid to obtain a single vertebra image corresponding to the single centroid;

[0067] In some embodiments, the method for acquiring the single vertebral image in 104 includes:

[0068] Obtain a 3D Gaussian distribution with a×b×c fixed around the single centroid; the values ​​of a / b / c need to meet the following conditions: Since a / b / c is a hyperparameter, it needs to be a multiple of 16;

[0069] Get a 3D image with a×b×c fixed at the center of mass. The values ​​of a / b / c need to meet the following conditions: Since a / b / c is a hyperparameter, it needs to be a multiple of 16. And it needs to be used to crop a single spine, so the cropping area must include the entire spine. The larger the area, the greater the interference. Therefore, it should be as small as possible while ensuring that the entire vertebra is covered. The preferred value is 224×224×128, but its specific value is variable.

[0070] According to the 3D Gaussian distribution centered at the centroid, a single vertebral image corresponding to the centroid around the Gaussian distribution is segmented from the 3D image. The 3D Gaussian and the 3D image are input together to use the Gaussian distribution for positioning and only segment the spine mask around the Gaussian distribution.

[0071] In some embodiments, after 104, the method further includes 105, identifying the category of a single vertebra in the vertebral image based on the position of the centroid, wherein the ID corresponds to the vertebral number one by one, such as 1 for thoracic 1, and 2 for thoracic 2, which is predetermined;

[0072] Optionally, when the category of a single vertebra cannot be determined according to the method in 105, the method further comprises: inputting the image of the single vertebra into a classifier to identify the category of the single vertebra;

[0073] Optionally, the training method of the classifier includes: inputting a single vertebral image of a training set sample into the classifier, obtaining a predicted classification result, comparing it with the classification label, and optimizing it according to the comparison result to obtain a constructed classifier;

[0074] Optionally, the classifier includes any one or more of the following: decision tree, logistic regression, naive Bayes, neural network; preferably DenseNet in the neural network.

[0075] Specifically, the minimum three-dimensional box where the binary spine region of the spine image in 101 is located is input into the step of finely segmenting each unit of the spine. At this time, each vertebra represents a label, and after the segmentation is complete, a spine that distinguishes each vertebra is formed;

[0076] The principle of the minimum three-dimensional box is as follows:

[0077] inds={(i,j,k)∣data[i][j][k]=1}

[0078] x1=min(inds[0])

[0079] x2=max(inds[0])

[0080] y1=min(inds[1])

[0081] y2=max(inds[1])

[0082] z1=min(inds[1])

[0083] z2=max(inds[1])

[0084] crop_data=data[x1:x2,y1:y2,z1:z2]

[0085] Here, x1, x2, y1, y2, z1, and z2 are integer indices that define the start and end positions of the cropping region. Specifically: x1 and x2 define the start and end indices for cropping along the first dimension (usually corresponding to the length). y1 and y2 define the start and end indices for cropping along the second dimension (usually corresponding to the width). z1 and z2 define the start and end indices for cropping along the third dimension (usually corresponding to the height).

[0086] In some embodiments, in some embodiments, after 105, the method further includes 106, synthesizing at least two single vertebrae based on the categories of the single vertebrae to obtain a synthesized image containing the target vertebra.

[0087] In some embodiments, after acquiring the three-dimensional CT image and before segmenting the spinal image in 101, the method further includes: preprocessing the three-dimensional CT image to obtain a unified three-dimensional CT image; the preprocessing includes: rotating the three-dimensional CT image to the RAS direction, randomly flipping, normalizing, and randomly shearing the image to a size of 128×128×128.

[0088] Figure 3 is a schematic diagram of a computer device provided by an embodiment of the present invention, such as Figure 3 As shown, the device may include: one or more processors, and one or more memories; wherein the memories store computer-readable codes, and when the computer-readable codes are run by the one or more processors, the method described above may be executed.

[0089] The processor in this embodiment can be an integrated circuit chip with signal processing capabilities. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, operations and logic block diagrams in the embodiments of the present disclosure can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc., which can be an X86 architecture or an ARM architecture.

[0090] In general, various example embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Certain aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that may be executed by a controller, microprocessor, or other computing device. When various aspects of the disclosed embodiments are illustrated or described as block diagrams, flow charts, or using some other graphical representation, it will be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuits or logic, general purpose hardware or controllers or other computing devices, or some combination thereof as non-limiting examples.

[0091] For example, the method or device according to the embodiment of the present disclosure may also be implemented by Figure 4 The architecture of the computing device 3000 shown in FIG. Figure 4 As shown, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. The storage device in the computing device 3000, such as ROM 3030 or hard disk 3070, may store various data or files used for processing and / or communication of the method provided by the present disclosure and program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, Figure 4 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 4 One or more components of a computing device are shown.

[0092] The embodiment of the present invention also provides a computer-readable storage medium, such as Figure 5As shown, it is a schematic diagram of a storage medium provided by an embodiment of the present invention, and a computer readable instruction 4010 is stored on the computer storage medium 4020. When the computer readable instruction 4010 is executed by the processor, the method according to the embodiment of the present disclosure described with reference to the above figures can be executed. The computer readable storage medium in the embodiment of the present disclosure can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM) or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory. It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0093] The embodiment of the present disclosure also provides a computer program product or system, which implements the steps of the above method when the computer program is executed by a processor;

[0094] A spine segmentation system disclosed in this embodiment, such as Figure 2 As shown, the system comprises:

[0095] An image rough segmentation module 201 is used to obtain a three-dimensional CT image of the spine, and segment the three-dimensional CT image to obtain an N-value mask image of the entire spine;

[0096] A cropping processing module 202 is used to crop the individual vertebrae in the N-value mask image of the entire spine to obtain an N-value mask image in which adjacent individual vertebrae are not connected;

[0097] A centroid positioning module 203 is used to locate the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebrae, and obtain an image of the located centroid;

[0098] A vertebra segmentation module 204, for segmenting a single vertebra image corresponding to the single centroid from the image of the located centroid with the single centroid as the center;

[0099] In some embodiments, after the vertebra segmentation module, the system further comprises: a vertebra identification module for identifying the category of a single vertebra in the vertebra image based on the position of the centroid; optionally, when the category of a single vertebra cannot be determined according to the method of the vertebra identification module, the method further comprises: inputting the single vertebra image into a classifier to identify the category of the single vertebra;

[0100] In some embodiments, after the vertebra identification module, the system further comprises: a vertebra synthesis module for synthesizing at least two single vertebrae based on the categories of the single vertebrae to obtain a synthesized image containing the target vertebrae.

[0101] Example: Segmentation based on the patient's preoperative CT images

[0102] The deep learning method swinunetr was used to segment the spine, and after training, good segmentation results were obtained; Swin UNetR combined the context modeling ability of Swin Transformer and the pixel-level prediction ability of U-Net to improve the performance of semantic segmentation tasks.

[0103] 1) Labeling Dataset

[0104] Mainstream deep learning-based segmentation methods require a large amount of data sets for training to generalize the model capabilities. First, ask doctors to mark the cervical, thoracic, and lumbar vertebrae.

[0105] 2) Deep learning algorithm training

[0106] At present, the U-shaped structure of medical image segmentation in the CNN field has achieved the most advanced performance benchmark, such as the U-Net model. Swinunetr is also based on the U-shaped structure design, including a decoder and an encoder. We know that the transformer model has shown excellent ability to capture random information in multiple fields, including natural language processing and computer vision. However, the kernel size of the convolution layer is limited, and its modeling performance for remote information is not ideal. Inspired by this, swinunetr's encoder uses transformers. This architectural design improves the algorithm's ability to capture information and reduces the algorithm's computational complexity.

[0107] 3) Implementation details

[0108] Rotate the data to the RAS direction;

[0109] Randomly flip the data;

[0110] Normalize the data;

[0111] Randomly cut the data into 128x128x128 size;

[0112] Preprocessing before segmentation, unifying the direction, constraining the data within a certain range (images have different coordinate systems), and unifying the images will result in better results during subsequent segmentation.

[0113] Specifically as attached Figure 6 FIG. 1 is a flow chart of the operation of the segmentation method provided by an embodiment of the present invention. Figure 6 After splitting into Figure 7 , 8 , 9, 10; when segmenting, you can first perform binary segmentation (roughly locate the spinal region first, and then reduce the region), and then perform multi-value segmentation, such as 25-value segmentation; you can also perform four-value segmentation first (roughly locate the cervical, thoracic, and vertebral regions first, and use the segmentation algorithm to preliminarily locate the thoracic and lumbar vertebrae to reduce the images outside the thoracic and lumbar vertebrae for the subsequent segmentation results), and then perform N-value segmentation for the three different regions; you can also add a centroid positioning step between coarse positioning and fine segmentation. The centroid can represent a single vertebra that makes up the spine. Because the vertebrae are not connected to each other, the pedicles are discarded during segmentation. The remaining vertebrae can be directly located. If adjacent vertebrae are connected, it is necessary to use corrosion algorithms or other algorithms to exploit the connected vertebrae so that the adjacent vertebrae are not connected. After locating multiple centroids (the centroid is a spatial point with three-dimensional coordinates), when locating a single vertebra according to the centroid, the size of the effect is unknown, so the image will be too large or too small when cropping. For large images, the cropped 3D CT is segmented into a single vertebra according to the 3D Gaussian distribution of a single centroid to obtain a single vertebra image and complete the cutting of the spine. If the spine is incomplete and the specific vertebra or vertebra cannot be determined, a classifier needs to be introduced to pre-learn the shape of each vertebra in the training set sample, and then training can be performed.

[0114] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0115] In general, various example embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, firmware, logic, or any combination thereof. Certain aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that may be executed by a controller, microprocessor, or other computing device. When various aspects of the disclosed embodiments are illustrated or described as block diagrams, flow charts, or using some other graphical representation, it will be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuits or logic, general purpose hardware or controllers or other computing devices, or some combination thereof as non-limiting examples.

[0116] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0117] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0118] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0119] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0120] The exemplary embodiments of the present disclosure described in detail above are merely illustrative and not restrictive. It should be understood by those skilled in the art that various modifications and combinations may be made to these embodiments or their features without departing from the principles and spirit of the present disclosure, and such modifications should fall within the scope of the present disclosure.

Claims

1. A spine segmentation method, characterized in that: The method comprises: 101, acquiring a three-dimensional CT image of a spine, and segmenting the three-dimensional CT image to obtain an N-value mask image of the entire spine; 102, cropping the single vertebrae in the N-value mask image of the entire spine to obtain an N-value mask image in which adjacent single vertebrae are not connected; 103, locating the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebrae, and obtaining an image of the located centroid; 104 , taking the single centroid as the center, segmenting from the image of the located centroid to obtain a single vertebra image corresponding to the single centroid.

2. The spine segmentation method according to claim 1, characterized in that: The method of cutting out the single vertebral region in the N-value mask image of the whole spine in 102 includes: cutting out gradually from the edges of the single vertebral body to the center; preferably, cutting out from the edges of the single vertebral body to the center in equal proportion; Optionally, the volume of the single vertebral body after trimming in 102 is smaller than the volume of the single vertebral body before trimming; Preferably, the ratio of the volume of the single vertebral body after cutting to the volume of the single vertebral body before cutting is: 1 / 10-9 / 10, preferably 1 / 5-4 / 5, preferably 1 / 2; Optionally, the coordinates of the centroid are three-dimensional coordinates; Optionally, the N-value mask is a binary mask.

3. The spine segmentation method according to claim 1, characterized in that: The method for locating the centroid of a single vertebra in the N-value mask image that is not connected between adjacent single vertebra in 103 includes: using a connected domain algorithm to identify at least one separate region in the N-value mask image that is not connected between adjacent single vertebra; locating the centroid of the separate region is the centroid of the single vertebra; the separate region is the region where the single vertebra is located; Optionally, when locating the centroid of a single vertebra in 103 , the vertebrae are sorted according to the z-axis coordinate of the centroid; Optionally, the method for acquiring the single vertebral image in 104 includes: Obtain a 3D Gaussian distribution with a fixed a×b×c value centered on the single centroid; Get a 3D image with a fixed size of a×b×c centered at the center of mass; According to the 3D Gaussian distribution centered at the centroid, a single vertebral image corresponding to the centroid around the Gaussian distribution is segmented from the 3D image.

4. The spine segmentation method according to claim 3, characterized in that: After 101 and before 102, the method further includes: cropping the N-value mask image of the vertebra to obtain an N-value mask image containing the target vertebra; Optionally, the method of cropping the N-value mask image of the spine includes: finding the maximum and minimum values ​​of the N-value spine mask in three directions of the RAS three-dimensional space, and cropping all regions in the original CT that are smaller than the minimum value and larger than the maximum value in the three directions, thereby obtaining an N-value mask image containing the target spine; Optionally, the target vertebra is a vertebra that needs to be displayed according to demand.

5. The spine segmentation method according to any one of claims 1 to 4, characterized in that: After 104, the method further includes 105, identifying the category of a single vertebra in the vertebral image based on the position of the centroid; optionally, when the category of a single vertebra cannot be determined according to the method in 105, the method further includes: inputting the single vertebral image into a classifier to identify the category of the single vertebra; Optionally, the classifier includes any one or more of the following: decision tree, logistic regression, naive Bayes, neural network; preferably DenseNet in the neural network.

6. The spine segmentation method according to claim 5, characterized in that: After 105 , the method further includes 106 , synthesizing at least two single vertebrae based on the categories of the single vertebrae to obtain a synthesized image containing the target vertebrae.

7. The spine segmentation method according to claim 1, characterized in that: In the 101, after acquiring the three-dimensional CT image and before segmenting the spinal image, the method further includes: preprocessing the three-dimensional CT image to obtain a unified three-dimensional CT image; the preprocessing includes: rotating the three-dimensional CT image to the RAS direction, randomly flipping, normalizing, and randomly shearing the image to a size of 128×128×128.

8. A computer device, characterized in that: The device comprises: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Citation Information

Patent Citations

  • 3D multi-vertebra segmentation method based on cascade convolution neural network

    CN109493317A

  • Spine image processing method based on artificial intelligence and related device

    CN110599508A

  • Digital instrument character recognition method

    CN110659645A

  • Vertebra detection method and computer equipment

    CN112184664A

  • Automatic segmentation and identification method for spine of X-ray film

    CN113205535A