Spine bone mineral density quantification method and device based on depth mark distribution learning

Abstract

The invention discloses a spine bone mineral density quantification method and device based on depth mark distribution learning, and the method comprises the steps: (1) obtaining a spine X-ray image data set of a patient, and carrying out the preprocessing; (2) positioning the preprocessed X-ray image to obtain a segmentation mask of a lumbar vertebra region; (3) inputting the preprocessed X-ray image into a data enhancement unit to obtain a pseudo-X-ray image, and taking the X-ray image and the pseudo-X-ray image as training samples; (4) constructing a depth mark distribution learning model, and training by adopting the training samples under different body basic conditions to obtain depth mark distribution learning models under different body basic conditions; (5) performing weighted mixing on the bone mineral density probability density functions of the bone mineral density prediction Gaussian distribution and the bone mineral density standard distribution under each body basic condition to generate a bone mineral density mixed prediction model; and (6) inputting an X-ray image to be predicted into the bone mineral density hybrid prediction model to obtain a bone mineral density prediction result. The method is low in error and high in sensitivity.

Description

Technical Field

[0001] This invention relates to computer vision technology, and more particularly to a method and device for quantifying spinal bone density based on depth-labeled distribution learning. Background Technology

[0002] Currently, dual-energy X-ray absorptiometry (DXA) is the gold standard for clinical bone mineral density testing. However, it has limitations such as the risk of ionizing radiation (approximately 5.4 μSv per scan), high equipment costs, and high per-test expenses. While deep learning models based on X-ray images can avoid radiation issues, traditional CNN models struggle to capture the anisotropic characteristics of trabecular microstructure (such as directional loss of trabecular bone in osteoporosis patients), and the sample size of publicly available datasets is insufficient (e.g., only 174 cases in TCB2014), limiting the model's generalization ability. Furthermore, existing models lack uncertainty quantification mechanisms (such as Monte Carlo Dropout or Bayesian neural networks), failing to output confidence intervals and reducing clinical reliability. Although traditional QCT technology can assess three-dimensional bone mineral density, it relies on in vitro mold calibration, limiting its widespread application.

[0003] The WHO recommends that osteoporosis screening should meet the following criteria: sensitivity ≥90% and MAE ≤0.05 g / cm³. 2 However, existing models (such as ResNet-based regression networks) have Dice coefficients of only 0.85-0.89 on the DXA validation set, far below the clinically required 0.912 (spinal ROI) and 0.893 (femoral neck ROI). Furthermore, existing models still have a localization error of over 15% in high-risk vertebral areas, indicating a high level of error. Summary of the Invention

[0004] To address the problems existing in the prior art, the purpose of this invention is to provide a method and device for quantifying spinal bone mineral density based on deep labeled distribution learning, which has relatively low error and high sensitivity.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] A method for quantifying spinal bone mineral density based on deep labeled distribution learning includes the following steps:

[0007] (1) Obtain an image dataset including spinal X-ray images of several patients and perform preprocessing;

[0008] (2) Perform coarse localization of the vertebral body on the preprocessed X-ray images, and perform fine localization based on the coarse localization results to obtain the segmentation mask of the lumbar region;

[0009] (3) Input the preprocessed X-ray image into the data augmentation unit to obtain the pseudo X-ray image. Use the X-ray image and the pseudo X-ray image together as training samples. The data augmentation unit is a generator of a trained conditional generative adversarial network. The conditional vector of the conditional generative adversarial network includes the segmentation mask and the patient's basic physical condition.

[0010] (4) Construct a deep label distribution learning model and train it using training samples under different physical conditions to obtain a deep label distribution learning model under different physical conditions. The deep label distribution learning model includes a localization branch, a prediction branch, a mean branch, a variance branch, and a bone density distribution output branch. The localization branch is used to locate the center of the lumbar vertebral body of the spine using a segmentation mask and training samples under different physical conditions. The prediction branch is used to predict the bone density of each vertebral body based on the located vertebral body center. The mean branch is used to calculate the mean bone density. The variance branch is used to calculate the variance bone density. The bone density distribution output branch is used to output the mean and variance of the predicted Gaussian distribution of bone density, which are respectively the mean and variance of the bone density.

[0011] (5) The probability density functions of the predicted Gaussian distribution of bone density under each physical baseline condition and the standard distribution of bone density under the same physical baseline condition are weighted and mixed to generate a mixed prediction model of bone density, and the weights are optimized using the EM method.

[0012] (6) Input the X-ray image to be predicted into the bone density mixed prediction model to obtain the bone density mixed distribution prediction result.

[0013] Furthermore, step (1) specifically includes:

[0014] (1.1) Obtain an image dataset including spinal X-ray images of several patients;

[0015] (1.2) Perform grayscale enhancement processing on each X-ray image;

[0016] (1.3) The feature point matching technique was used to locate the center point of the lower edge of the L3 vertebral body in the X-ray image, and the X-ray image was aligned to the standard anatomical coordinate system by using the center point of the lower edge of the L3 vertebral body as the reference point through affine transformation.

[0017] (1.4) Perform grayscale normalization on the X-ray images.

[0018] Furthermore, step (2) specifically includes:

[0019] (2.1) For the preprocessed X-ray images, coarse localization is performed by a key point detection network based on sparse convolution to obtain the coarse localization region of the L1-L4 vertebrae.

[0020] (2.2) Input the preprocessed X-ray images and the coarse localization regions of the L1-L4 vertebrae into the U-Net++ network, which includes an encoder and a decoder, to obtain the segmentation mask of the lumbar region.

[0021] Furthermore, the encoder described in step (2.2) uses a ResNet-50 deep residual network as its backbone network. The ResNet-50 deep residual network embeds dynamic dilated convolution modules into the residual blocks of the ResNet-50 network and adds channel attention modules to the jump connections of the residual blocks.

[0022] The dynamic dilated convolution module comprises the following three parts:

[0023] The dynamic hole rate prediction submodule is used to generate a dynamic hole rate through a lightweight convolutional network, which is specifically a 1×1 convolutional layer, a ReLU layer and a Sigmoid layer connected in sequence.

[0024] The dilated convolutional layer group is used to perform dilated variable convolution operations on the dynamic dilation rate output by the dynamic dilation rate prediction submodule.

[0025] Residual connection is used to perform residual connection between the output of the dynamic hole rate prediction submodule and the output of the holed convolutional layer group;

[0026] The normalization module is used to normalize the output of the residual connection.

[0027] Furthermore, the training method for the conditional generative adversarial network described in step (3) includes:

[0028] (3.1) Construct a conditional generative adversarial network, wherein the conditional generative adversarial network includes a generator and a discriminator;

[0029] (3.2) Generate a condition vector for each X-ray image, wherein the condition vector c = (age, gender, BMI, Mask), where age represents the age of the patient to whom the X-ray image belongs, gender represents the sex, BMI is an indicator of body fatness, age, gender, and BMI form the basic physical condition, and Mask represents the segmentation mask of the lumbar spine region.

[0030] (3.3) The conditional vector of each X-ray image is concatenated with the random noise vector to obtain the input features of the generator.

[0031] (3.4) Input features Input generator to generate xfake image;

[0032] (3.5) Input the xfake image and the X-ray image into the discriminator at the same time to obtain the discrimination result;

[0033] (3.6) Calculate the joint loss based on the discrimination results, and continuously optimize the generator and discriminator based on the joint loss until training is completed.

[0034] Furthermore, the age, gender, and BMI of the condition vector c are normalized and encoded, and the random noise vector follows a Gaussian distribution N(0,1).

[0035] Furthermore, step (4) specifically includes:

[0036] (4.1) Construct a deep label distribution learning model, wherein the deep label distribution learning model includes:

[0037] The localization branch includes a four-level residual convolutional block and a fully connected regression layer. Each residual convolutional block includes two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The input is the segmentation mask and training samples under different body baseline conditions, and the output is the coordinates of the L1-L4 vertebral center points.

[0038] The prediction branch is specifically a pre-trained improved VGG19 network. This improved VGG19 network retains the first 16 convolutional layers of the original VGG19 network, removes the fully connected layers and their classification heads, and adds a mean branch and a variance branch after the conv5_4 convolutional layer of the VGG19 network. The input to the prediction branch is the output of the localization branch and the preprocessed X-ray image, and the output is the predicted Gaussian distribution of bone mineral density for each vertebra. The first 16 convolutional layers of the VGG19 network and the conv5_4 convolutional layer are used to predict the bone mineral density of each vertebra based on the preprocessed X-ray image. The mean branch is used to calculate the mean bone mineral density μ of the predicted bone mineral density based on the bone mineral density of each vertebra. pred The variance branch is used to calculate the bone mineral density variance σ of the predicted bone mineral density value output by the prediction branch. pred σ pred =log(1+exp(W) σ ν σ +b σ In the formula, W σ ν σ 、b σ These represent the weight matrix, bone density prediction vector, and bias term of the variance branch, respectively; the outputs then contain the mean and variance μ. pred σ pred The predicted Gaussian distribution of bone mineral density;

[0039] (4.2) Input the training samples and segmentation masks under different body baseline conditions into the depth label distribution learning model for training to obtain depth label distribution learning models under different body baseline conditions. The loss function used during training is:

[0040]

[0041] L loc Let p represent the loss function for the localization branch, N represent the number of training samples in the batch, and p represent the loss function for the localization branch. ij This represents the coordinates of the center point of the cone Lj in the i-th training sample output by the localization branch. L represents the actual center point coordinates. KL Let μ represent the loss function for the prediction branch using KL divergence. real σ real These represent the mean and variance of the true bone mineral density distribution, respectively.

[0042] Furthermore, step (5) specifically includes:

[0043] (5.1) The probability density function of the predicted Gaussian distribution of bone mineral density under each body baseline condition and the standard distribution of bone mineral density under the same body baseline condition are weighted and mixed to generate a mixed prediction model of bone mineral density:

[0044]

[0045] Among them, P mix (x) represents the bone mineral density probability density function of the density prediction model, ω represents the weights, x represents the location x, and μ WHO ,σ WHO Let μ represent the mean and variance of the standard distribution of bone mineral density, respectively. pred ,σ pred This represents the mean and variance of a Gaussian distribution used to predict bone mineral density. Indicates the standard distribution of bone mineral density at point x. This indicates the predicted Gaussian distribution of bone mineral density at point x;

[0046] (5.2) For the mixed bone density prediction model, the mixed distribution is first obtained through the EM algorithm. Then, the mixed distribution is used as the output, and the KL divergence is used as the loss function to optimize ω to obtain the optimized mixed bone density prediction model.

[0047] Furthermore, step (6) is followed by:

[0048] (7) Risk assessment of the predicted mixed distribution of bone mineral density:

[0049] High risk: P(x≤X)>0.8;

[0050] Medium risk: 0.5 <P(x≤X)≤0.8;

[0051] Low risk / normal: P(x>X)≥0.5;

[0052] Where X is the international standard reference value for bone mineral density, x represents the bone mineral density at position x, and P() represents the corresponding probability.

[0053] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0054] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention replaces the traditional regression model with deep labeled distribution learning, which realizes accurate quantitative assessment of measurement uncertainty, reduces positioning error, improves the accuracy of bone density quantification, thereby reducing quantification error and improving quantification sensitivity; combined with conditional generative adversarial networks to generate synthetic pseudo-X-ray images, it effectively solves the problem of insufficient image samples. When the labeled data is reduced by 50%, the model still maintains 98% prediction accuracy, with higher sensitivity and lower error. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an embodiment of the spinal bone mineral density quantification method based on deep labeled distribution learning provided by the present invention.

[0056] Figure 2 It is an X-ray image;

[0057] Figure 3 This invention provides an improved U-Net++ network structure.

[0058] Figure 4 This is a structural diagram of a residual block in the ResNet-50 deep residual network provided by this invention;

[0059] Figure 5 This is a structural diagram of the improved VGG19 network provided by the present invention. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0061] Example 1

[0062] This invention provides a method for quantifying spinal bone mineral density based on deep labeled distribution learning, such as... Figure 1 As shown, it includes the following steps:

[0063] (1) Obtain an image dataset including spinal X-ray images of several patients and perform preprocessing.

[0064] This step specifically includes:

[0065] (1.1) Obtain an image dataset including X-ray DICOM (Digital Imaging and Communications in Medicine) images of the spine (including lumbar spine (L1-L4)) of several patients; covering different ages, genders and pathological states, including osteoporosis, osteopenia and normal bone density.

[0066] DICOM format is an international standard for medical imaging. Its core value lies in the unified storage of high-precision images and clinical metadata (such as patient information and equipment parameters), ensuring data compatibility and integrity. DICOM supports 16-bit grayscale depth lossless rendering of trabecular bone and vertebral body details. Combined with the 3D reconstruction capabilities of multi-angle images, it provides a precise foundation for spinal curvature analysis, osteoporosis quantification, and fracture detection. Simultaneously, its structured metadata seamlessly integrates with hospital information systems (PACS / HIS), helping doctors quickly access and compare historical images and serving as a key input for medical AI development, driving the implementation of intelligent diagnostic technologies such as bone density prediction and vertebral segmentation. DICOM format permeates the entire process from image acquisition and clinical diagnosis to scientific research and innovation, serving as the cornerstone of medical digitalization and precision medicine.

[0067] Furthermore, throughout the entire process of patient data acquisition and use, this study strictly adhered to the requirements of regulations such as the Personal Information Protection Law and the Guidelines for Desensitization Technology of Medical Imaging Data. For case data requiring informed consent, collaborating medical institutions clearly informed patients through written agreements that the data would be used for the development of bone density prediction algorithms and multi-center medical research, and patients could choose to withdraw from the research without affecting their medical rights. The scope of data use was limited to vertebral ROI feature extraction and mixed distribution modeling, and its use for unrelated research or commercial purposes was prohibited. Patient biometric data underwent double desensitization: 18 metadata fields, including PatientID and StudyDate, were removed from the DICOM file header information using automated tools (such as the DicomAnonymizer Toolkit); Gaussian blurring and edge pixel interference were applied to non-target areas (such as facial contours and pelvic anatomical features) at the image pixel level to ensure irreversible anonymization. The encrypted storage system used the AES-256 algorithm to encrypt the desensitized data in a hierarchical manner and set up a dynamic two-factor authentication mechanism (biometric identification + dynamic token), allowing only authorized researchers to access designated datasets under audit log monitoring. The retention period for raw data strictly follows the principle of minimization. Physical destruction is performed within 6 months after the model training is completed. Fragmented storage media are transferred to a third-party certification body for verification after demagnetization.

[0068] (1.2) Perform grayscale enhancement processing on each X-ray image.

[0069] Specifically, the CLAHE algorithm can be used for grayscale enhancement to highlight the texture of bone trabeculae and soft tissues. The formula is as follows:

[0070] I cnhanced =CLAHE(I raw (Tile = 8×8, ClipLimit = 2.0)

[0071] Where, I cnhanced This represents the enhanced image, where Tile indicates the image block size (the original X-ray image was divided into 8×8 local regions for adaptive equalization). raw This represents the input grayscale matrix data of the original X-ray image. ClipLimit represents the contrast limit threshold (controlling the histogram clipping range to prevent excessive noise amplification).

[0072] (1.3) The feature point matching technique was used to locate the center point of the lower edge of the L3 vertebral body in the X-ray image, and the X-ray image was aligned to the standard anatomical coordinate system by using the center point of the lower edge of the L3 vertebral body as the reference point through affine transformation.

[0073] Specifically, SIFT feature point matching technology can be used to locate the center point of the lower edge of the L3 vertebral body, and the image can be aligned to the standard anatomical coordinate system through affine transformation to eliminate geometric distortion caused by patient positional deviation.

[0074] (1.4) Perform grayscale normalization on the X-ray images.

[0075] Specifically, the Z-Score standardization method can be used for standardization to unify imaging differences between devices. The formula is as follows:

[0076]

[0077] Among them, I norn Represents the standardized image, μ global and σ global The grayscale mean and standard deviation of the entire dataset are provided. A standardized image dataset is generated after preprocessing. Furthermore, data augmentation can be performed through elastic deformation and random occlusion.

[0078] (2) Perform coarse localization of the vertebral body on the preprocessed X-ray images, and perform fine localization based on the coarse localization results to obtain the segmentation mask of the lumbar region.

[0079] This step specifically includes:

[0080] (2.1) For the preprocessed X-ray images, coarse localization is performed using a sparse convolution-based keypoint network (SCoNeT) to obtain the coarse localization region of the L1-L4 vertebrae.

[0081] (2.2) The preprocessed X-ray images and the coarse localization regions of the L1-L4 vertebrae are input into a U-Net++ network including an encoder and decoder to obtain a segmentation mask for the lumbar region, such as... Figure 2 As shown. The U-Net++ network is as follows. Figure 3 As shown.

[0082] The encoder uses a ResNet-50 deep residual network as its backbone network, which is an improvement on the ResNet-50 network.

[0083] ResNet-50 consists of multiple convolutional layers, pooling layers, fully connected layers, and residual blocks. Its overall structure is as follows:

[0084] Input: Dimensions are 3×224×224;

[0085] Convolutional layer: The first convolutional layer uses a 7×7 convolutional kernel, with 64 kernels, a stride of 2, padding of 3, and an output size of 64×112×112.

[0086] Pooling layers: Max pooling layers change the dimensions but do not affect the number of elements;

[0087] Residual block groups: Each group consists of multiple residual blocks, and each residual block is composed of multiple convolutional layers. These convolutional layers are processed using batch normalization and the ReLU activation function;

[0088] In addition, there are global average pooling layers and fully connected layers.

[0089] like Figure 4 As shown, the ResNet-50 deep residual network embeds a Dynamic Dilated Convolution module into the residual block of the ResNet-50 network. Its dilation rate is dynamically adjusted according to the resolution of the input feature map, and a Channel Attention module is added to the skip connections of the residual block.

[0090] The dynamic dilated convolution module is an improved method based on dilated convolution. Its core idea is to adaptively adjust the dilation rate of each convolutional layer according to the statistical or attention information of the input features to improve the flexibility of the receptive field and the context modeling capability. It includes the following three parts:

[0091] The dynamic hole rate prediction submodule is used to generate a dynamic hole rate through a lightweight convolutional network, which is specifically a 1×1 convolutional layer, a ReLU layer and a Sigmoid layer connected in sequence.

[0092] The dilated convolutional layer group is used to perform deformable convolution operations with holes on the dynamic dilation rate output by the dynamic dilation rate prediction submodule.

[0093] Residual connection is used to perform residual connection between the output of the dynamic hole rate prediction submodule and the output of the holed convolutional layer group;

[0094] The normalization module is used to normalize the output of the residual connection.

[0095] (3) Input the preprocessed X-ray image into the data enhancement unit to obtain the pseudo X-ray image, and use the X-ray image and the pseudo X-ray image together as training samples.

[0096] The data augmentation unit is a generator of a trained conditional generative adversarial network (cGAN), and the conditional vector of the cGAN includes a segmentation mask and the patient's basic physical information (including the patient's age, gender, and BMI index).

[0097] The training method for the conditional generative adversarial network includes:

[0098] (3.1) Construct a conditional generative adversarial network, which includes a generator (e.g., U-Net) and a discriminator (e.g., PatchGAN, consisting of 5 sets of convolutional layers (kernel size 4×4, stride 2));

[0099] (3.2) Generate a condition vector for each X-ray image, wherein the condition vector c = (age, gender, BMI, Mask), where age represents the age of the patient to which the X-ray image belongs, gender represents the sex, BMI is an indicator of body fatness, age, gender, and BMI form the basic physical condition and are normalized and encoded, and Mask represents the segmentation mask of the lumbar spine region.

[0100] (3.3) The conditional vector of each X-ray image is concatenated with the random noise vector to obtain the input features of the generator. The random noise vector z follows a Gaussian distribution N(0,1);

[0101] (3.4) Input features Input generator to generate xfake image;

[0102] (3.5) Input the xfake image and the X-ray image into the discriminator at the same time to obtain the discrimination result;

[0103] (3.6) Calculate the joint loss based on the discrimination results, and continuously optimize the generator and discriminator based on the joint loss until training is completed.

[0104] The joint loss function is:

[0105]

[0106] L adv X represents the combined loss. real Let D() represent an X-ray image, G() represent a discriminator function, and G() represent a generator function.

[0107] The age, gender, and BMI of the condition vector c are normalized and encoded.

[0108] (4) Construct a deep label distribution learning model and train it using training samples under different physical conditions to obtain a deep label distribution learning model under different physical conditions.

[0109] This step specifically includes:

[0110] (4.1) Construct a deep label distribution learning model, wherein the deep label distribution learning model includes:

[0111] The localization branch includes a four-level residual convolutional block and a fully connected regression layer. Each residual convolutional block includes two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The input is the segmentation mask and training samples under different body baseline conditions. The output is the coordinates of the center points of the L1-L4 vertebrae. The coordinate system takes the upper left corner of the image as the origin and the accuracy reaches the pixel level.

[0112] The prediction branch, specifically a pre-trained improved VGG19 network, such as... Figure 5As shown, the improved VGG19 network retains the first 16 convolutional layers of the VGG19 network, removes the fully connected layers and their classification heads, and adds a mean branch and a variance branch after the convolutional layer conv5_4 of the VGG19 network. The input of the prediction branch is the output of the localization branch and the preprocessed X-ray image, and the output is the predicted Gaussian distribution of bone mineral density for each vertebra. The first 16 convolutional layers of the VGG19 network and the convolutional layer conv5_4 are used to predict the bone mineral density of each vertebra based on the preprocessed X-ray image. The mean branch is specifically a fully connected layer used to calculate the mean bone mineral density μ of the predicted bone mineral density value based on the bone mineral density of each vertebra. pred The variance branch is specifically a fully connected layer used to calculate the bone mineral density variance σ of the predicted bone mineral density value output by the prediction branch. pred σ pred =log(1+exp(W) σ ν σ +b σ In the formula, W σ ν σ 、b σ These represent the weight matrix, bone density prediction vector, and bias term of the variance branch, respectively; the outputs then contain the mean and variance μ. pred σ pred The predicted Gaussian distribution of bone mineral density;

[0113] (4.2) Input the training samples and segmentation masks under different body baseline conditions into the depth label distribution learning model for training to obtain depth label distribution learning models under different body baseline conditions. The loss function used during training is:

[0114]

[0115] L loc Let p represent the loss function for the localization branch, N represent the number of training samples in the batch, and p represent the loss function for the localization branch. ij This represents the coordinates of the center point of the cone Lj in the i-th training sample output by the localization branch. L represents the actual center point coordinates. KL Let μ represent the loss function for the prediction branch using KL divergence. real σ real Let L represent the mean and variance of the true bone mineral density distribution (a Gaussian distribution centered on DXA measurements), respectively. KL The loss term prompts the network: when the predicted value μ pred When σ approaches the true value pred Approaching σ real (Reflecting inherent noise in the data), when μ pred When deviating from the true value, σ pred Automatically increases to compensate for forecast uncertainty.

[0116] (5) The bone density prediction Gaussian distribution under each body baseline condition and the bone density standard distribution under the same body baseline condition are weighted and mixed to generate a bone density mixed prediction model, and the weight is optimized using the EM method.

[0117] (5.1) The predicted Gaussian distribution of bone mineral density under each physical baseline condition and the probability density function of the standard distribution of bone mineral density under the same physical baseline condition (same age group and same sex) are weighted and mixed to generate a mixed bone mineral density prediction model:

[0118]

[0119] Among them, P mix (x) represents the bone mineral density probability density function of the density prediction model, ω represents the weights, x represents the location x, and μ WHO ,σ WHO Table 1 shows the mean and variance of the standard distribution of bone mineral density (WHO distribution), respectively. pred ,σ pred This represents the mean and variance of a Gaussian distribution used to predict bone mineral density. Indicates the standard distribution of bone mineral density at point x. This indicates the predicted Gaussian distribution of bone mineral density at point x;

[0120] Table 1 Standard distribution of bone mineral density under different physical baseline conditions

[0121] gender age group Mean bone mineral density range Clinical significance female 20-29 years old 1.0-1.2 Peak bone mineral density 30-49 years old 0.95-1.15 slow descent 50-59 years old 0.85-1.05 The decline accelerates after menopause. ≥60 years old 0.7-0.9 High risk of osteoporosis male 20-29 years old 1.05-1.25 Peak bone mineral density 30-49 years old 1.0-1.2 slow descent 50-69 years old 0.95-1.15 Gradually decline ≥70 years old 0.8-1.0 Significant decline

[0122] (5.2) The weights ω are optimized by the EM method to obtain the optimized bone density hybrid prediction model.

[0123] The EM method includes two steps: the E step and the M step.

[0124] Step E: Calculate the posterior probability γ of training sample i belonging to the WHO distribution according to the following formula. i :

[0125]

[0126] M-step: Update weights:

[0127]

[0128] ω (t) This represents the weight ω value at the t-th iteration.

[0129] (6) Input the X-ray image to be predicted into the bone density mixed prediction model to obtain the bone density mixed distribution prediction result.

[0130] (7) Perform risk assessment on the predicted results of the mixed distribution of bone density and generate a report:

[0131] High risk: P(x ≤ X) > 0.8, that is, there is a probability of > 80% in the osteoporosis area;

[0132] Medium risk: 0.5 < P(x ≤ X) ≤ 0.8, that is, there is a probability of 50% - 80% in the osteoporosis area;

[0133] Low risk / normal: P(x > X) > 0.5, that is, there is a probability of more than 50% not in the osteoporosis area;

[0134] Among them, X is the internationally standard reference value of bone density, P() represents the corresponding probability, P(x ≤ X) > 0.8 represents the probability that the area where bone density is less than or equal to X is greater than 80%, 0.8 < P(x ≤ X) ≤ 0.5 represents the probability that the area where bone density is less than or equal to X is greater than 50% and less than or equal to 80%, and P(x > X) > 0.5 represents the probability that the area where bone density is greater than X is greater than 50%.

[0135] In the method of the present invention, the parameter design in the image preprocessing stage directly serves to improve the quality and standardization of medical images. The Tile parameter determines the range of local contrast enhancement. Smaller blocks can enhance the visibility of fine structures such as trabeculae, but may also amplify image noise; larger blocks can suppress noise but may blur the key anatomical features of the vertebral body edge. The Clip Limit parameter prevents artifacts from being generated due to over-enhancement in high-density areas (such as metal implants) by restricting the pixel proportion of a single gray level, ensuring that the true bone density information in low-contrast areas is retained. This design provides a reliable visual basis for subsequent analysis while improving the image quality.

[0136] The global mean (μ global ) and standard deviation (σ global ) parameters in the image standardization stage are used to eliminate the inherent differences between different imaging devices. The global mean reflects the average pixel gray level of all images in the dataset. By subtracting the mean, the brightness baselines of different devices can be aligned; the standard deviation is used to unify the dynamic range of image contrast, so that the model input is not affected by the fluctuations of device parameters. This processing significantly improves the compatibility of cross-institutional data. For example, after the images collected from different hospitals are standardized, the bone density prediction error can be reduced.

[0137] The generator (G) and discriminator (D) parameters in the generative adversarial network (cGAN) act synergistically to generate synthetic images that retain pathological features. The generator synthesizes spinal images with specific pathological features through the fusion of a noise vector and a conditional label. Its design needs to ensure the anatomical rationality of trabecular texture and vertebral body morphology; the discriminator improves the authenticity of the generated images through adversarial training.

[0138] The hybrid distribution weights ω in labeled distribution learning dynamically balance model prediction with clinical standards. When the model's prediction uncertainty is high (e.g., bone mineral density values ​​are close to the diagnostic threshold), ω is automatically reduced to prioritize the adoption of WHO standards, reducing the risk of misdiagnosis. This design enables the algorithm to proactively prompt doctors to review complex cases and transform uncertainty into auxiliary information for clinical decision-making.

[0139] Example 2

[0140] This invention provides a computer device that provides services for implementing the method described in Embodiment 1. The device may include: a memory storing a computer-executable program; a processor coupled to the memory; and the processor calling the computer-executable program stored in the memory to execute the steps of the method described in Embodiment 1.

[0141] The memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the memory may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard disk drive"). A program / utility having a set (at least one) of program modules may be stored in, for example, memory. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The computer-executable program of the program modules typically performs the functions and / or methods described in the embodiments of the present invention.

[0142] The processor executes various functional applications and data processing by running programs stored in memory, such as the method provided in Embodiment 1 of the present invention.

[0143] The code of a computer executable program can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages.

[0144] It should be understood that the embodiments and descriptions above are only the principles, main features and advantages of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the invention, and all such changes and modifications fall within the protection scope of the present invention.

Claims

1. A method for quantifying spinal bone mineral density based on deep labeled distribution learning, characterized in that, The method includes the following steps: (1) Obtain an image dataset including spinal X-ray images of several patients and perform preprocessing; (2) Perform coarse localization of the vertebral body on the preprocessed X-ray images, and perform fine localization based on the coarse localization results to obtain the segmentation mask of the lumbar region; (3) Input the preprocessed X-ray image into the data augmentation unit to obtain the pseudo X-ray image. Use the X-ray image and the pseudo X-ray image together as training samples. The data augmentation unit is a generator of a trained conditional generative adversarial network. The conditional vector of the conditional generative adversarial network includes the segmentation mask and the patient's basic physical condition. (4) Construct a deep label distribution learning model and train it using training samples under different physical conditions to obtain a deep label distribution learning model under different physical conditions. The deep label distribution learning model includes a localization branch, a prediction branch, a mean branch, a variance branch, and a bone density distribution output branch. The localization branch is used to locate the center of the lumbar vertebral body of the spine using a segmentation mask and training samples under different physical conditions. The prediction branch is used to predict the bone density of each vertebral body based on the located vertebral body center. The mean branch is used to calculate the mean bone density. The variance branch is used to calculate the variance bone density. The bone density distribution output branch is used to output the mean and variance of the predicted Gaussian distribution of bone density, which are respectively the mean and variance of the bone density. (5) The probability density functions of the predicted Gaussian distribution of bone density under each physical baseline condition and the standard distribution of bone density under the same physical baseline condition are weighted and mixed to generate a mixed prediction model of bone density, and the weights are optimized using the EM method. (6) Input the X-ray image to be predicted into the bone density mixed prediction model to obtain the bone density mixed distribution prediction result.

2. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 1, characterized in that, Step (1) specifically includes: (1.1) Obtain an image dataset including spinal X-ray images of several patients; (1.2) Perform grayscale enhancement processing on each X-ray image; (1.3) The feature point matching technique was used to locate the center point of the lower edge of the L3 vertebral body in the X-ray image, and the X-ray image was aligned to the standard anatomical coordinate system by using the center point of the lower edge of the L3 vertebral body as the reference point through affine transformation. (1.4) Perform grayscale normalization on the X-ray images.

3. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 1, characterized in that, Step (2) specifically includes: (2.1) For the preprocessed X-ray images, coarse localization is performed by a key point detection network based on sparse convolution to obtain the coarse localization region of the L1-L4 vertebrae. (2.2) Input the preprocessed X-ray images and the coarse localization regions of the L1-L4 vertebrae into the U-Net++ network, which includes an encoder and a decoder, to obtain the segmentation mask of the lumbar region.

4. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 3, characterized in that, The encoder described in step (2.2) uses ResNet-50 deep residual network as the backbone network. The ResNet-50 deep residual network embeds dynamic dilated convolution modules into the residual blocks of the ResNet-50 network and adds channel attention modules to the jump connections of the residual blocks. The dynamic dilated convolution module comprises the following three parts: The dynamic hole rate prediction submodule is used to generate a dynamic hole rate through a lightweight convolutional network, which is specifically a 1×1 convolutional layer, a ReLU layer and a Sigmoid layer connected in sequence. The dilated convolutional layer group is used to perform dilated variable convolution operations on the dynamic dilation rate output by the dynamic dilation rate prediction submodule. Residual connection is used to perform residual connection between the output of the dynamic hole rate prediction submodule and the output of the holed convolutional layer group; The normalization module is used to normalize the output of the residual connection.

5. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 1, characterized in that, The training method for the conditional generative adversarial network described in step (3) includes: (3.1) Construct a conditional generative adversarial network, wherein the conditional generative adversarial network includes a generator and a discriminator; (3.2) Generate a condition vector for each X-ray image, wherein the condition vector c = (age, gender, BMI, Mask), where age represents the age of the patient to whom the X-ray image belongs, gender represents the sex, BMI is an indicator of body fatness, age, gender, and BMI form the basic physical condition, and Mask represents the segmentation mask of the lumbar spine region. (3.3) The conditional vector of each X-ray image is concatenated with the random noise vector to obtain the input features of the generator. (3.4) Input features Input generator to generate xfake image; (3.5) Input the xfake image and the X-ray image into the discriminator at the same time to obtain the discrimination result; (3.6) Calculate the joint loss based on the discrimination results, and continuously optimize the generator and discriminator based on the joint loss until training is completed.

6. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 5, characterized in that, The age, gender, and BMI of the condition vector c are normalized and encoded, and the random noise vector follows a Gaussian distribution N(0,1).

7. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 1, characterized in that, Step (4) specifically includes: (4.1) Construct a deep label distribution learning model, wherein the deep label distribution learning model includes: The localization branch includes a four-level residual convolutional block and a fully connected regression layer. Each residual convolutional block includes two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The input is the segmentation mask and training samples under different body baseline conditions, and the output is the coordinates of the L1-L4 vertebral center points. The prediction branch is specifically a pre-trained improved VGG19 network. This improved VGG19 network retains the first 16 convolutional layers of the original VGG19 network, removes the fully connected layers and their classification heads, and adds a mean branch and a variance branch after the conv5_4 convolutional layer of the VGG19 network. The input to the prediction branch is the output of the localization branch and the preprocessed X-ray image, and the output is the predicted Gaussian distribution of bone mineral density for each vertebra. The first 16 convolutional layers of the VGG19 network and the conv5_4 convolutional layer are used to predict the bone mineral density of each vertebra based on the preprocessed X-ray image. The mean branch is used to calculate the mean bone mineral density μ of the predicted bone mineral density based on the bone mineral density of each vertebra. pred The variance branch is used to calculate the bone mineral density variance σ of the predicted bone mineral density value output by the prediction branch. pred σ pred =log(1+exp(W) σ ν σ +b σ In the formula, W σ ν σ b σ These represent the weight matrix, bone density prediction vector, and bias term of the variance branch, respectively; the outputs then contain the mean and variance μ. pred σ pred The predicted Gaussian distribution of bone mineral density; (4.2) Input the training samples and segmentation masks under different body baseline conditions into the depth label distribution learning model for training, and obtain the depth label distribution learning model under different body baseline conditions. The loss function used during training is: L loc Let p represent the loss function for the localization branch, N represent the number of training samples in the batch, and p represent the loss function for the localization branch. ij This represents the coordinates of the center point of the cone Lj in the i-th training sample output by the localization branch. L represents the actual center point coordinates. KL Let μ represent the loss function for the prediction branch using KL divergence. real σ real These represent the mean and variance of the true bone mineral density distribution, respectively.

8. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 1, characterized in that, Step (5) specifically includes: (5.1) The probability density function of the predicted Gaussian distribution of bone mineral density under each body baseline condition and the standard distribution of bone mineral density under the same body baseline condition are weighted and mixed to generate a mixed prediction model of bone mineral density: Among them, P mix (x) represents the bone mineral density probability density function of the density prediction model, ω represents the weights, x represents the location x, and μ WHO ,σ WHO Let μ represent the mean and variance of the standard distribution of bone mineral density, respectively. pred ,σ pred This represents the mean and variance of a Gaussian distribution used to predict bone mineral density. Indicates the standard distribution of bone mineral density at point x. This indicates the predicted Gaussian distribution of bone mineral density at point x; (5.2) For the mixed bone density prediction model, the mixed distribution is first obtained through the EM algorithm. Then, the mixed distribution is used as the output, and the KL divergence is used as the loss function to optimize ω, so as to obtain the optimized mixed bone density prediction model.

9. The spinal bone mineral density quantification method based on deep labeled distribution learning according to claim 1, characterized in that, Step (6) is followed by: (7) Risk assessment of the predicted mixed distribution of bone mineral density: High risk: P(x≤X)>0.8; Medium risk: 0.5 <P(x≤X)≤0.8; Low risk / normal: P(x>X)≥0.5; Where X is the international standard reference value for bone mineral density, x represents the bone mineral density at position x, and P() represents the corresponding probability.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes the computer program to implement the method as described in any one of claims 1-9.

Images