A method of bone density detection

Abstract

The present application relates to the technical field of bone density detection, and discloses a bone density detection method, comprising the following steps: step 1) image acquisition; step 2) image segmentation; step 3) three-dimensional surface reconstruction; step 4) three-dimensional body distribution reconstruction; and step 5) calibration calculation; the present application realizes rapid and accurate detection of lumbar vertebrae bone density by combining an automatic segmentation algorithm, a three-dimensional reconstruction algorithm and a gray density mapping technology based on Lambert-Beer's law, thereby providing a reliable basis for early diagnosis of osteoporosis; in addition, the present application is based on ordinary X-ray images, has low radiation dose and low equipment cost, is suitable for primary medical institutions and resource-limited areas, and has wide popularization value and application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bone density detection, in particular to a bone density detection method. BACKGROUND

[0002] Bone Mineral Density (BMD) is an important indicator for assessing bone health, widely used in the diagnosis of osteoporosis and the assessment of fracture risk. Osteoporosis is a common chronic disease, especially prevalent in the elderly, characterized by reduced bone mass and deterioration of bone microstructure, leading to increased bone fragility. By measuring bone density, osteoporosis and its progression can be detected early, helping doctors develop prevention and treatment plans.

[0003] With the intensification of global aging trend, the incidence of osteoporosis has also increased significantly. According to the World Health Organization (WHO), a bone fracture caused by osteoporosis occurs every three seconds worldwide. As a key means of assessing bone health, bone density measurement has become an important tool for preventing and managing osteoporosis.

[0004] Dual-energy X-ray Absorptiometry (DXA): This is the most commonly used bone density measurement technique in clinical practice, which measures the difference in absorbed X-ray intensity by passing X-rays of different energies through the bone. DXA has high precision and low radiation dose, and is the "gold standard" for diagnosing osteoporosis.

[0005] Quantitative Computed Tomography (QCT): Based on CT imaging technology, QCT can provide three-dimensional bone density information and measure the density of cancellous bone and cortical bone separately, suitable for early detection of osteoporosis. However, QCT has a high radiation dose, limiting its clinical application.

[0006] However, existing bone density measurement techniques still have some limitations. For example, DXA, although with high measurement accuracy, only provides two-dimensional projection information, cannot reflect the three-dimensional density distribution of the bone, and is easily affected by patient size, posture and other tissue structures, leading to measurement errors. In addition, DXA equipment is expensive and requires a high operating environment, limiting its application in primary medical institutions or resource-limited areas. While QCT can provide three-dimensional density information of the bone, its high radiation dose and high cost make it difficult to be used as a routine screening method. SUMMARY

[0007] In view of this, the purpose of the present application is to provide a new bone density detection method, aiming to simplify the detection process, reduce the cost of equipment and improve the detection accuracy, so as to effectively improve the early diagnosis popularization rate of osteoporosis.

[0008] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0009] A bone density detection method, comprising the following steps:

[0010] Step 1) image acquisition, the reference module with known equivalent bone density and the human body part to be measured are imaged under the same conditions, and the X-ray image containing the lumbar vertebrae and the reference module is obtained; the standard density value of the reference module is known, and the gray value distribution in the image is clear and distinguishable; (the reference module is a physical 3D model; the human lumbar vertebrae and the reference module are in the same X-ray image)

[0011] Step 2) image segmentation, the lumbar vertebrae and the reference module are separated from the background by an optimized segmentation algorithm, and the gray information of the segmented image is extracted;

[0012] Step 3) three-dimensional surface reconstruction, based on the segmented image, an optimized three-dimensional surface reconstruction algorithm is used to establish a basic three-dimensional surface model of the lumbar vertebrae and the reference module, and then the established basic three-dimensional surface model is continuously optimized according to the two-dimensional contour of the segmented image;

[0013] Step 4) three-dimensional body distribution reconstruction, based on an optimized three-dimensional body distribution reconstruction algorithm, the gray information of the segmented image is mapped to the three-dimensional space to obtain an initial three-dimensional body distribution model, and a series of optimizations are performed on the initial body distribution model to finally obtain three-dimensional space density distribution information matched with the gray information;

[0014] Step 5) calibration calculation, the mapping relationship between the three-dimensional space density information and the gray information is established by combining the optimized body distribution model, and then the gray information is converted into bone density value and output by the mapping relationship between the three-dimensional space density distribution information and the bone density.

[0015] Further,

[0016] S1.1 uses a conventional X-ray device to collect X-ray images I containing lumbar vertebrae and reference modules, and the images need to contain the target lumbar region and the reference module with known physical density;

[0017] S1.2: the standard density value of the reference module is known, and the gray value distribution in the image needs to be clear and distinguishable.

[0018] Further, the image segmentation includes a segmentation module and a gray extraction module;

[0019] The segmentation module is responsible for separating the lumbar vertebrae and the reference module from the background in the X-ray image containing the human body and the reference module, and completing this work through an optimal segmentation algorithm to ensure the accuracy and robustness of the segmentation and save the segmentation results.

[0020] The gray scale extraction module extracts the gray scale information of the lumbar vertebrae and the reference module from the segmented image and saves it, providing basic data for subsequent work.

[0021] Further,

[0022] S2.1: input the acquired lumbar X-ray image into the pre-trained automatic segmentation model (the segmentation model is U-net++), and separate the lumbar region of interest and the reference module therefrom;

[0023] S2.2: extract the gray scale information G(x, y) of the segmented region, wherein G(x, y) represents the gray scale distribution of the image on a two-dimensional plane;

[0024] S2.3: output the results as the lumbar segmentation image R(x, y) and the reference module segmentation image R ref (x, y).

[0025] Further, the segmentation module comprises:

[0026] The training data required by the segmentation model includes the original X-ray image and the L1-L4 lumbar vertebrae segmentation mask label;

[0027] The original X-ray image is used as the input during model training, the corresponding lumbar vertebrae segmentation mask label is used as the gold standard, the difference between the model segmentation result and the gold standard is used as the supervised learning target, the segmentation model is trained, and the segmentation model that minimizes the difference between the model segmentation result and the gold standard is obtained;

[0028] Further, the three-dimensional surface reconstruction comprises a three-dimensional surface model reconstruction module and a three-dimensional surface model optimization module.

[0029] The three-dimensional surface model reconstruction module uses the lumbar segmentation image to generate a basic three-dimensional model of the lumbar vertebrae and the reference module using an optimal three-dimensional surface reconstruction algorithm based on the segmentation image;

[0030] The three-dimensional surface model optimization module compares the projection image with the two-dimensional contour of the segmentation image by performing two-dimensional projection on the basic three-dimensional model, iteratively optimizes the basic three-dimensional surface model based on the contour difference, and makes the optimized three-dimensional surface model more accurately match the two-dimensional contour of the segmentation image.

[0031] Further,

[0032] S3.1: Based on the segmented lumbar region image R(x,y) and the reference module segmented image R ref (x,y), using a three-dimensional reconstruction algorithm, generate a lumbar three-dimensional surface model M(x,y,z) and a reference module three-dimensional surface model M ref (x,y,z);

[0033] S3.2: Project the generated lumbar three-dimensional surface model M(x,y,z) and the reference module three-dimensional surface model M ref (x,y,z) to a two-dimensional plane to generate a projection image I(x,y);

[0034] S3.3: Continuously optimize the three-dimensional surface model by reducing the difference between the projection I(x,y) and the real X-ray image I two-dimensional contour.

[0035] Further, the three-dimensional surface model reconstruction module and the three-dimensional surface model optimization module include:

[0036] Read the segmented image, set the threshold to binarize the image to obtain the two-dimensional contour of the lumbar vertebral body and the reference module;

[0037] Based on the two-dimensional contour of the vertebral body and the reference module, combined with prior knowledge, use the preferred three-dimensional surface reconstruction algorithm to establish a basic lumbar vertebral body and reference module three-dimensional model;

[0038] Project the established basic three-dimensional surface model to two dimensions, compare the difference between the generated projection image contour and the lumbar vertebral body and reference module two-dimensional contour, and continuously optimize the three-dimensional surface model according to the difference.

[0039] Further, the three-dimensional body distribution reconstruction includes: a body distribution reconstruction module and an optimization module;

[0040] The body distribution reconstruction uses a preferred three-dimensional body distribution reconstruction algorithm to reconstruct the two-dimensional gray scale information of the segmented image into three-dimensional space to obtain an initial three-dimensional body distribution model;

[0041] The body distribution reconstruction module includes:

[0042] Extract the gray scale information of the target region from the image segmentation result;

[0043] Based on prior knowledge, use the preferred three-dimensional body distribution reconstruction algorithm to reconstruct the two-dimensional gray scale information into three-dimensional space to obtain an initial three-dimensional body distribution model containing three-dimensional space density information.

[0044] The optimization module restricts the spatial density information mapped into the three-dimensional space to within the established three-dimensional model, and continuously iteratively optimizes the three-dimensional volume distribution model based on the difference in grayscale information between the simulated X-ray image and the segmented image obtained by performing simulated X-ray projection on the initial three-dimensional volume distribution model.

[0045] The optimization module is characterized by comprising: a projection module and a three-dimensional space density distribution information optimization module;

[0046] The projection module performs simulated X-ray projection on the three-dimensional spatial density distribution information to obtain a simulated X-ray image;

[0047] Projection module, including:

[0048] According to actual needs, set the position of the X-ray source and select the appropriate projection angle;

[0049] Along the propagation path of the ray, calculate the intersection of the ray and each spatial density information, and perform attenuation calculation;

[0050] The projection results are synthesized into a set of two-dimensional image data, and the projection images are post-processed as necessary to simulate the real X-ray imaging effect.

[0051] The three-dimensional spatial density distribution information optimization module compares the differences between the obtained simulated X-ray image and the corresponding areas of the original lumbar spine X-ray segmentation image, and continuously optimizes the distribution of the three-dimensional spatial density information based on the differences.

[0052] Going further,

[0053] S4.1: Import the 2D grayscale information G(x,y) of the lumbar vertebrae and reference module from the segmented image. Using the preferred 3D volume distribution reconstruction algorithm, map the 2D grayscale information to 3D space to obtain an initial 3D volume distribution model μ(x,y,z). (The segmented image refers to both the lumbar vertebrae and the reference module, not just the lumbar vertebrae.)

[0054] S4.2: Using the established 3D surface model of the lumbar spine and reference module, perform boundary constraints on the 3D information distribution and remove abnormal voxels. The boundary constraint is expressed as:

[0055]

[0056] Where M represents the three-dimensional model boundary of the lumbar spine and the reference module, and μ′(x, y, z) is the three-dimensional density information after constraint.

[0057] S4.3: Import the constrained three-dimensional density information μ′(x, y, z), determine the position of the ray source, and select an appropriate projection angle;

[0058] S4.4: Ray propagation simulation, simulate X-ray from the source point, through the three-dimensional space model, according to Lambert Beer's law, calculate the attenuation of each three-dimensional density information on the ray during the ray propagation;

[0059] S4.5: Project the results into a set of two-dimensional image data I combined And compare it with the two-dimensional gray information G(x,y) of the segmentation image, and continuously optimize the distribution of three-dimensional space density information according to the difference between the two.

[0060] Further, the calibration calculation includes:

[0061] Combined with the optimized three-dimensional body distribution model, the mapping relationship between the two-dimensional gray information and the three-dimensional space density distribution information is established;

[0062] And by the conversion relationship between the three-dimensional space density distribution information and the bone density, the three-dimensional space density distribution information in the three-dimensional body distribution model is converted into the corresponding bone density value.

[0063] Further,

[0064] S5.1: Combined with the optimized three-dimensional body distribution model, the mapping relationship between the three-dimensional space density distribution information μ'(x,y,z) and the two-dimensional gray information G(x,y) is established:

[0065] μ=a·G+b

[0066] Wherein, a is the proportion coefficient, b is the offset, and a and b are determined by least square fitting of the measurement data of the reference module;

[0067] S5.2: And by the conversion relationship between the three-dimensional space density distribution information and the bone density, the three-dimensional space density distribution information μ'(x,y,z) in the three-dimensional body distribution model is converted into the corresponding bone density value.

[0068] The beneficial effects of the present application are as follows:

[0069] The present application proposes a bone density detection scheme based on ordinary X-ray image, which realizes the rapid and accurate detection of lumbar bone density by combining automatic segmentation algorithm, three-dimensional reconstruction algorithm and gray density mapping technology based on Lambert Beer's law, and provides reliable basis for early diagnosis of osteoporosis.

[0070] The technical promotion of the present application is embodied in the following aspects: first, the automatic segmentation model of deep learning can efficiently and accurately separate the lumbar vertebrae and the reference module region from the X-ray image, reducing the interference of manual operation; second, the three-dimensional model generated based on the accurate segmentation of the X-ray image can accurately reconstruct the three-dimensional model of the lumbar vertebrae from the two-dimensional image, greatly improving the reconstruction accuracy; finally, by combining the known density information of the reference module, the mapping relationship between the gray value and the bone density value is established, effectively improving the accuracy of bone density calculation.

[0071] In addition, the present application is based on ordinary X-ray images, has low radiation dose and low equipment cost, is suitable for primary medical institutions and resource-limited areas, and has wide popularization value and application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a flowchart of a bone density detection method disclosed by an embodiment of the present application;

[0073] Figure 2 is a process diagram of a bone density detection method disclosed by an embodiment of the present application for lumbar vertebrae image segmentation and establishing a three-dimensional surface model of the lumbar vertebrae based on the segmentation result and optimization;

[0074] Figure 3 is a schematic diagram of a specific implementation process of a bone density detection method disclosed by an embodiment of the present application for mapping two-dimensional gray information to three-dimensional space and utilizing the established three-dimensional model for boundary constraint;

[0075] Figure 4 is a schematic diagram of a specific implementation process of a bone density detection method disclosed by an embodiment of the present application for generating a projection image according to the three-dimensional space density information distribution, comparing the projection image with the original X-ray segmentation image, and continuously iterating and optimizing the three-dimensional space density distribution information according to the gray information difference. DETAILED DESCRIPTION

[0076] The present application will be described in detail below with reference to specific embodiments. The following examples are helpful for those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that those skilled in the art can make several changes and improvements without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0077] The present application discloses a bone density detection method, which compared with the prior art, is beneficial to simplify the detection process, and thus can perform bone density detection under conventional medical conditions, thereby improving the early diagnosis popularization rate of osteoporosis. The following will be described in detail.

[0078] Embodiment 1:

[0079] Referring to Figure 1 , Figure 1 is the overall flowchart of the bone density detection method disclosed in the embodiments of the present application, and specifically comprises the following steps:

[0080] S1. Obtain an X-ray image containing the human lumbar vertebrae and the reference module;

[0081] S1.1. Collect the X-ray image of the lumbar vertebrae and the reference module of the individual to be detected;

[0082] S1.2. Standardize, normalize, denoise, and other pretreatments are performed on the image to ensure consistent image quality and remove unnecessary noise or artifacts.

[0083] S2. Separate the lumbar vertebrae and the reference module from the background by an automatic segmentation algorithm, and extract the gray scale information of the segmented image:

[0084] The construction of the automatic lumbar vertebrae segmentation model is based on a deep convolutional neural network framework. Commonly used networks include u-net, U-Net++, SegNet, DeepLab, etc. The input of the model is the preprocessed lumbar vertebrae X-ray picture and the lumbar vertebrae segmentation label annotated by experienced radiologists, and the output is a binary segmentation result picture of the lumbar vertebrae.

[0085] Optionally, the training process of the model is as follows:

[0086] S2.1. Collect the original X-ray image dataset required for model training and perform standardization, normalization, denoising, and other pretreatments on the dataset to ensure consistent image quality and remove unnecessary noise or artifacts;

[0087] S2.2. Under the guidance of professional technicians, manually or semi-automatically annotate the preprocessed X-ray images to generate label images (ground truth) required for segmentation. The accuracy of the annotation directly affects the subsequent segmentation effect;

[0088] S2.3. The preprocessed lumbar vertebrae X-ray picture is used as the input during model training, and the label image generated under the guidance of professional technicians is used as the gold standard. The difference (loss function Ls) between the model segmentation result and the gold standard is used as the supervised learning target to train the lumbar vertebrae segmentation model;

[0089] S2.4. Finally, a lumbar vertebrae automatic segmentation model is obtained, which minimizes the difference between the model segmentation result and the gold standard.

[0090] The deep learning segmentation model used in the embodiment is an improved U-net++. In the model training process, the input is the lumbar X-ray image after data augmentation (flipping, translation, cropping, mirroring, adding noise, etc.), and the output is the corresponding lumbar vertebra segmentation image. Adam is used as the optimizer for back propagation in network training, the iteration number (epoch) is 500 times, the batch-size is 8, and the input image size is 256*256.

[0091] S3. Based on the segmentation image, a three-dimensional model of the lumbar vertebra and the reference module is established by using a three-dimensional reconstruction algorithm, and the established three-dimensional model is continuously optimized according to the two-dimensional contour of the segmentation image:

[0092] The construction of the three-dimensional reconstruction model of the lumbar vertebra is based on the segmentation image, a binary method is selected, and a suitable threshold is set to obtain the two-dimensional contour of the lumbar vertebra.

[0093] Please refer to Figure 2 , Figure 2 is a process diagram for the lumbar vertebra image segmentation and three-dimensional surface reconstruction of the segmentation result of the bone density detection method disclosed in the embodiment of the application, and specifically includes the following steps:

[0094] S3.1. Import the segmentation image, set a suitable threshold, and binarize the image to obtain the two-dimensional contour of the lumbar vertebra and the reference module;

[0095] S3.2. Based on the two-dimensional contour of the lumbar vertebra and the reference module, according to prior knowledge, a three-dimensional reconstruction algorithm is used (step 1. Contour curve input: input two-dimensional contour curve data, the curve is given in the form of a set of discrete points, and the coordinates of each point are represented as (x i, y i ); step 2. Setting of the rotation axis: 1) setting the rotation axis as the connecting line between the upper and lower middle points of the minimum circumscribed rectangle of the two-dimensional contour, and only retaining the curve on the right side of the rotation axis; 2) defining the rotation angle range as 0 to 2π; step 3. Three-dimensional point cloud generation, 1) for each point (x i, y i ) on the contour curve, calculating its three-dimensional coordinates (x i, y i ,z i ) at different angles; 2) repeating the above operation to form complete three-dimensional point cloud data; step 4. Constructing a triangular mesh according to the point cloud data to generate a three-dimensional surface), establishing a simple three-dimensional surface model of the lumbar vertebra and the reference module;

[0096] S3.3. Projecting the established three-dimensional surface model to a two-dimensional plane, in order to be as close to the real situation as possible, the perspective projection method is adopted;

[0097] Optionally, perspective projection is performed as follows:

[0098] Perspective projection maps the depth of a three-dimensional object onto a two-dimensional plane.

[0099]

[0100] Where (x, y, z) are the coordinates of the point in three-dimensional space, f is the focal length, and (x′, y′) are the projection coordinates on the plane.

[0101] S3.4. Compare the result of the two-dimensional projection with the binarized segmented image, and continuously optimize the simple benchmark three-dimensional surface model established in step S3.2 based on the difference between the two. Set a fixed error standard. When the optimized result meets the error requirement, output the result; if it does not meet the requirement, continue to optimize.

[0102] S4. Using the preferred three-dimensional volume distribution reconstruction algorithm (step 1. input data preparation: 1) to obtain a set of X-ray image data, represented as a two-dimensional grayscale image set {I 1, I 2, …,I n}, the resolution of each image is W×H; 2) based on prior knowledge, obtain the voxel data range of grayscale information in three-dimensional space; step 2. initialize the volume data matrix, establish a three-dimensional volume data matrix V(x, y, z) based on the surface three-dimensional model optimized by S3.4 and initialize it; step 3. fill the volume data matrix: fill the voxel data obtained in step 1 into the initialized volume data matrix to obtain an initial three-dimensional volume distribution model. ), map the grayscale information to three-dimensional space to obtain an initial three-dimensional volume distribution model, and perform a series of optimizations on it, including constraining the boundaries of the three-dimensional space density information distribution combined with the established three-dimensional volume distribution model and comparing the difference in grayscale information between the X-ray image simulated by the volume distribution model and the original X-ray segmentation image to continuously iteratively optimize the volume distribution model;

[0103] See also Figure 3 , Figure 3 This is a flowchart of a specific implementation process of a bone density detection method disclosed in an embodiment of the present invention, which maps two-dimensional grayscale information to a three-dimensional space and uses the established three-dimensional model to perform boundary constraints, specifically including the following steps:

[0104] S4.1. According to the Lambert-Beer law, when X-rays pass through an object, the attenuation of the rays is related to the density of the object and the path length of the rays. The formula is as follows:

[0105] I=I0·e -μx

[0106] where I0 is the intensity of the incident ray, I is the intensity of the ray after penetrating the object, μ is the linear attenuation coefficient of the object, and x is the path length of the ray through the object.

[0107] From this equation, the relationship between the gray value and the density distribution of the object can be obtained. The gray value is usually related to the attenuation of the ray, so the gray information can reflect the physical density at different positions. To map the gray value to the three-dimensional space, a pre-trained mapping algorithm model can be used, which has learned the relationship between the gray information and the density value. The key steps of the mapping process include:

[0108] S4.1.1. Extract the gray value of each pixel from the X-ray image;

[0109] S4.1.2. Use the trained mapping algorithm model to convert the gray value of each pixel to the corresponding three-dimensional space density value;

[0110] S4.1.2. Position each converted density value to the corresponding position in the three-dimensional space through the established three-dimensional space coordinates;

[0111] S4.2. The three-dimensional surface model constrains the boundary of the three-dimensional space density information distribution:

[0112] The established three-dimensional surface model not only provides the geometric structure information of the space, but also serves as a spatial constraint for the density information. The shape and boundary of the three-dimensional model can help determine which areas belong to the valid physical space and which areas should not have density information or the density should be zero. The specific steps include:

[0113] S4.2.1. Define the spatial boundary of the three-dimensional surface model according to the reconstructed three-dimensional surface model;

[0114] S4.2.2. According to the boundary of the three-dimensional surface model, limit the density information to exist only within the model or the region of interest, and avoid generating unreasonable density values in the blank area;

[0115] S4.3. Simulate X-ray projection on the three-dimensional space density distribution information, and continuously iterate and optimize the three-dimensional space density distribution information according to the difference between the projection image and the segmented X image to obtain the three-dimensional space density distribution information that matches the segmented X image;

[0116] Please refer to Figure 4 , Figure 4 is a bone density detection method according to the three-dimensional space density information distribution to generate a projection image and an original X-ray segmented image, and continuously iterate and optimize the three-dimensional space density distribution information according to the gray information difference. The specific implementation process of the schematic diagram of the specific implementation process of the three-dimensional space density distribution information, which includes the following steps:

[0117] The generation process of the simulation X-ray image is as follows:

[0118] S4.3.1. Projection parameter setting, determining the projection direction, projection angle, focus position and other parameters of the X-ray. These parameters are consistent with the real X-ray imaging geometry, so that the generated simulation X-ray image is as similar as possible to the real X-ray image;

[0119] S4.3.2: Line integral calculation, calculate the cumulative attenuation value of each ray passing through the voxel when the X-ray passes through the three-dimensional model from a certain angle. The cumulative attenuation value reflects the absorption of the ray after passing through different density tissues in the body;

[0120] S4.3.3: Image generation, generate the corresponding gray-scale image through the calculated attenuation value;

[0121] S4.4. Compare the generated simulation X-ray image and the segmentation image, and continuously optimize the density information distribution in the three-dimensional space according to the difference between the two.

[0122] S5. Combined with the known physical density information of the reference module, the mapping relationship between the physical density information and the three-dimensional space density information is established, and the density value of the lumbar vertebrae is calculated and output;

[0123] S5.1. Combined with the optimized three-dimensional body distribution model, the mapping relationship between the three-dimensional space density distribution information and the two-dimensional gray-scale information is established; the known density reference is used for density distribution calibration and error correction, the extracted space density information is mapped with the gray value through the density mapping function, and the density distribution is processed and corrected in the three-dimensional model, to ensure the accuracy of the final result;

[0124] S5.2. The three-dimensional space density distribution information in the three-dimensional body distribution model is converted into the corresponding bone density value by the conversion relationship between the three-dimensional space density distribution information and the bone density, the bone density value of the lumbar vertebrae part is output, and a report is generated.

[0125] The specific embodiments of the application are described above. It should be understood that the application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other.

Claims

1. A bone density detection method, characterized in that: The steps include: Step 1) image acquisition, obtaining X-ray images of the human lumbar spine and a reference module with known equivalent bone density; Step 2) Image segmentation: Separate the lumbar vertebrae and reference modules from the background using a segmentation algorithm and extract the grayscale information of the segmented image; Step 3) 3D surface reconstruction: Based on the segmented image, a 3D surface reconstruction algorithm is used to establish a basic 3D surface model of the lumbar vertebral body and the reference module, and then the basic 3D surface model is optimized based on the 2D contour of the segmented image; Step 4) 3D volume distribution reconstruction: Based on the 3D volume distribution reconstruction algorithm, the grayscale information of the segmented image is mapped to 3D space to obtain an initial 3D volume distribution model. The initial 3D volume distribution model is optimized to finally obtain 3D spatial density distribution information that matches the grayscale information. The three-dimensional volume distribution reconstruction in step 4) includes the following steps: 4.1) Import the 2D grayscale information G(x,y) of the lumbar vertebrae and reference module in the segmented image and use a 3D volume distribution reconstruction algorithm to map the 2D grayscale information into 3D space to obtain an initial 3D volume distribution model μ(x,y,z); 4.2) Using the established 3D surface models of the lumbar spine and reference module, we impose boundary constraints on the 3D information distribution and remove abnormal voxels. The boundary constraints are expressed as: Where M represents the three-dimensional model boundary of the lumbar spine and the reference module, and μ′(x, y, z) is the three-dimensional density information after constraint; 4.3) Import the constrained 3D density information μ′(x, y, z), determine the location of the ray source, and select the appropriate projection angle; 4.4) Ray propagation simulation: simulates X-rays traveling from a source point through a three-dimensional spatial model. According to the Lambert-Beer law, along the ray propagation path, the intersection of the ray and each spatial density information is calculated, and attenuation is calculated to obtain the simulated X-ray projection results; 4.5) Synthesize the simulated X-ray projection results obtained in 4.4) into a set of two-dimensional image data I combined , and compare it with the grayscale information G(x,y) of the segmented image, and iteratively optimize the distribution of three-dimensional spatial density information based on the difference between the two; Step 5) Calibration calculation, establishing a mapping relationship between the three-dimensional spatial density information and the grayscale information, and then converting the grayscale information into a bone density value based on the mapping relationship between the three-dimensional spatial density distribution information and bone density, and outputting the result; The calibration calculation in step 5) includes the following steps: 5.1) Based on the optimized 3D volume distribution model, a mapping relationship is established between the 3D spatial density distribution information μ′(x, y, z) and the 2D grayscale information G(x, y): μ=a·G+b Where a is the scale factor and b is the offset, which are determined by fitting the measurement data of the reference module using the least squares method. 5.2) Based on the conversion relationship between the three-dimensional spatial density distribution information and bone density, the three-dimensional spatial density distribution information μ′(x, y, z) in the three-dimensional volume distribution model is converted into the corresponding bone density value.

2. A bone density detection method according to claim 1, characterized in that: The image acquisition in step 1) includes the following steps: 1.1) Acquire an original X-ray image I containing a lumbar spine and a reference module, wherein the X-ray image I includes a target lumbar spine region and a reference module with known equivalent bone density; 1.2) The original X-ray image I is preprocessed by standardization, normalization, denoising, and artifact removal to obtain an X-ray image I.

3. A bone density detection method according to claim 1, characterized in that: Image segmentation in step 2) includes the following steps: 2.1) Input the acquired lumbar spine X-ray image into the pre-trained segmentation model to separate the lumbar spine region of interest and the reference module; The segmentation model training process is as follows: Obtaining the training data required for the segmentation model, including X-ray image I and lumbar vertebral segmentation mask labels; The X-ray image I is used as the input for model training, the corresponding lumbar vertebral segmentation mask label is used as GroundTruth, and the difference between the model segmentation result and GroundTruth is used as the supervised learning target. The segmentation model is trained and iteratively optimized to obtain a segmentation model that minimizes the difference between the model segmentation result and GroundTruth. 2.2) Extract the grayscale information G(x,y) of the segmented area, where G(x,y) represents the grayscale distribution of the image on a two-dimensional plane; 2.3) The output results are the lumbar segmentation image R(x,y) and the reference module segmentation image R ref (x,y).

4. A bone density detection method according to claim 1, characterized in that: The three-dimensional surface reconstruction in step 3) includes the following steps: 3.1) Read the lumbar vertebra segmentation image R(x,y), set a threshold to binarize the image to obtain the two-dimensional contours of the lumbar vertebral body and the reference module; 3.2) Based on the lumbar segmentation image R(x,y) and the reference module segmentation image R ref The 2D contour of (x, y) is used to generate the lumbar 3D surface model M(x, y, z) and the reference module 3D surface model M using the 3D surface reconstruction algorithm. ref (x,y,z); 3.3) Compare the generated lumbar 3D surface model M(x, y, z) with the reference module 3D surface model M ref (x, y, z) is projected onto a two-dimensional plane to generate a projection image I(x, y); 3.4) By reducing the difference between the two-dimensional contour of the projection image I(x,y) and the real X-ray image I, an error standard is set. When the optimized result meets the error requirement, the result is output; if it does not meet the error requirement, the optimization is continued.

Images