Osteoporosis risk assessment system and method for lumbar magnetic resonance image

Abstract

The invention relates to the technical field of osteoporosis detection, in particular to an osteoporosis risk assessment system and method for lumbar magnetic resonance images. The system comprises an image acquisition module, a structure processing module, a feature extraction module, a bone mineral density prediction module and a reasoning control module, through MRI image preprocessing and image feature vector extraction, in combination with non-image information such as age, gender, height, weight and the like, a backstepping mapping model is constructed in a multi-modal mode, and continuous estimation of bone mineral density and bone related characterization values is achieved. According to the method, quantitative extraction of the iconography parameters of the lumbar vertebra bone tissue can be achieved under the non-radiation condition, and a safe and repeatable opportunity analysis means is provided for iconography monitoring and follow-up visit of high-risk people; meanwhile, the adaptability of the model in multi-center data is remarkably enhanced, overfitting is avoided, and the consistency and reliability of calculation results of bone characterization parameter values between different devices and samples are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of osteoporosis detection, and in particular to an osteoporosis risk assessment system and method for lumbar MRI images. BACKGROUND

[0002] Osteoporosis is a metabolic bone disease characterized by bone mass loss and microstructure damage of bone tissue, and its serious consequence is the occurrence of brittle fracture, especially in the elderly population, which has a high rate of disability and death. Bone density, as an important indicator for the diagnosis and management of osteoporosis, is a core parameter for evaluating bone strength and fracture risk. Before fracture occurs, bone density measurement is the main means for diagnosing osteoporosis. Currently, the commonly used bone density detection methods in clinical practice include dual-energy X-ray absorptiometry (DEXA), quantitative CT (QCT), ultrasonic bone density detection, and indirect measurement techniques based on MRI images. Among them, DEXA is widely recognized as the gold standard, but it has the problem of radiation exposure and cannot be frequently applied to long-term follow-up populations; QCT has a higher radiation dose, limiting its popular application; ultrasonic detection is greatly affected by noise and operating conditions, and its accuracy is unstable. Therefore, developing a new bone density evaluation method that is radiation-free, highly accurate, and can achieve opportunistic screening has important clinical value and promotional significance.

[0003] Although some studies have attempted to use fat fraction or VBQ values extracted from MRI images to analyze the correlation of bone density, these methods have problems such as small sample size, lack of multi-center verification, long time interval between data and gold standard DEXA measurement, and relatively simple algorithm structure, which affect the stability and prediction accuracy of the model. In addition, most of the existing methods only stop at the level of classifying osteoporosis risk levels, and fail to directly regress the continuous bone density real value with high accuracy, and lack of in-depth modeling of feature extraction and mapping mechanism, which limits the generalization ability and device consistency of the model. SUMMARY

[0004] The present application provides a method for assessing the risk of osteoporosis in lumbar MRI images.

[0005] To achieve the above-mentioned purposes, the technical solutions adopted by the present application are as follows: A method for assessing the risk of osteoporosis in lumbar MRI images, comprising the following steps: S1: Collect T1-weighted and T2-weighted image data of the lumbar region by magnetic resonance imaging sequence, perform preliminary organizational structure recognition processing and region recognition on the image data, and extract the vertebral body image region as input; S2: Perform sliding window scanning on the vertebral body image region, extract local image feature vectors and encode them into multi-dimensional feature vectors; S3: input the multi-dimensional feature vector and patient non-image information (including age, gender, height and weight) into the model, construct a multi-modal feature-based back-propagation mapping model, and perform fitting learning on the potential bone density information of the MRI image according to the multi-modal model, so as to establish a mapping relationship between the MRI image and the bone density real value in the joint feature space; introduce individual basic information (such as age, gender, height and weight) as an additional input variable, and combine it with the image feature vector to form the final joint input feature, and construct an enhanced back-propagation mapping model. The model structure has the ability to accept joint input of images and texts, which is convenient for future access to more clinical information (such as medical history and lifestyle). It is helpful to realize fine individualized osteoporosis screening based on MRI images.

[0006] S4: input the non-image information of the patient as an additional modal, and input it into the model together with the image feature vector to form a transverse multi-modal fusion feature, which is used to enhance the bone density estimation performance; S5: input a new MRI image sample and extract an image feature vector, and at the same time, obtain the non-image information of the corresponding patient, input the image feature and the non-image information into the trained model, and output the estimated value of the bone quality related representation value.

[0007] Further, in the above step S1, the step of extracting the vertebral body image region includes performing gray scale normalization processing and spatial resolution resampling processing on the image data, and automatically identifying the vertebral body boundary region of the lumbar vertebra segment through an anatomical structure model-based region segmentation method. A set of fixed vertebral body slice selection rules are adopted to extract the specified region for subsequent feature modeling, forming a standardized input image data set for subsequent modeling calculation. The slice selection rule is based on the central slice of the coronal plane or the sagittal plane of each vertebral body, and preferentially selects the slice with the clearest structure and the least interference from artifacts as the analysis object, which is determined by the bone boundary maximization region or the central volume projection plane method.

[0008] Furthermore, in step S2, the local image feature vectors extracted by the sliding window scanning include tissue structure contrast features based on the grayscale co-occurrence matrix of the vertebral region, trabecular structural stability features based on the local binary pattern within the vertebral boundary, fractal dimension image feature vectors reflecting the internal complexity of the vertebral body, bone distribution energy features based on multi-scale wavelet decomposition of the vertebral region, and Gabor response features for the directionality of the vertebral image structure. The various types of image feature vectors extracted are dimensionally connected to form a unified feature vector, and the eigenvalues ​​of each type are normalized to ensure the consistency of different feature dimensions and equivalence in the neural network input. After extraction, the various image feature vectors are uniformly encoded to form a structurally consistent multidimensional feature vector, which serves as the basis for subsequent model input. The tissue structure contrast feature based on the grayscale co-occurrence matrix of the vertebral region is used to reflect the grayscale variation pattern and tissue detail complexity inside the vertebral body; the trabecular structure stability feature based on the local binary pattern within the vertebral boundary is used to describe the repetitive pattern and directional consistency of the microstructure of the vertebral region; the fractal dimension image feature vector reflecting the internal complexity of the vertebral body is used to characterize the spatial roughness and structural disorder of bone tissue at multiple scales; the bone distribution energy feature based on multi-scale wavelet decomposition of the vertebral region is used to express the spatial distribution pattern of bone tissue texture under different frequency components; the Gabor response feature for the directionality of the vertebral image structure is used to capture the arrangement characteristics and striped structural differences of bone tissue in multiple directions.

[0009] Furthermore, in the above step S3, the mapping function model from the feature space to the real value of bone density is , its structure adopts a neural network composed of multiple layers of linear transformation and nonlinear activation function. The neural network structure is a multi-layer perceptron (MLP) with 5 fully connected layers, each layer contains 128~512 neurons, the activation function uses ReLU, and the last layer is a linear output. The residual connection structure is introduced internally to stabilize the training process, and the feature channel weighting module is configured to enhance the discrimination of structural information. The feature channel weighting module is constructed based on the SE mechanism, and dynamically adjusts the importance weight of each feature channel through global average pooling and channel weight recalibration operations, thereby improving the influence of key features on bone density prediction. In the initial stage, the network parameters are adjusted. The parameters are set according to the Gaussian distribution or He initialization rule. When the neural network uses ReLU or its variants as the activation function, the He initialization strategy is used for parameter initialization. If a saturation activation function such as Sigmoid or Tanh is used, the standard Gaussian initialization is used. The formula is: ; in, represents all trainable parameters of the mapping network, Represented as the model pair The numerical value of the bone-related characterization value predicted by the sample.

[0010] Furthermore, in the above step S3, the inverse mapping model is a deep neural network structure, the input is the feature vector of the MRI image, and the output is the corresponding bone density value. The construction of the inverse mapping model includes the following sub-steps: S3.1: Create a training sample dataset containing MRI image feature vectors, patient non-image information, and paired bone-related characterization values. The image and non-image information are jointly encoded to form input samples. S3.2: Using supervised learning with measured bone density as the target, a loss function is constructed to fit the continuous mapping relationship between MRI image features and bone density. S3.3: Iteratively optimize the parameters of the inverse mapping model through error backpropagation.

[0011] Furthermore, in the above step S3.1, the image feature vector The corresponding bone-related characterization value Pairing to form a structured supervised training sample set , the formula is: ; in, For the Image feature vectors extracted from MRI images, is the reference value of bone-related characterization values ​​measured by DXA system, Indicates the total number of training samples.

[0012] Furthermore, in the above step S3.2, in the structured supervised training sample set Based on this, the supervised objective function of the training process is defined, and the mean square error is used as the main loss function. At the same time, the L2 regularization term is introduced to limit the complexity of the model to suppress the tendency of overfitting. The loss function formula is: ; in, represents the total number of training samples, Represents the neural network model for the The structural features extracted from MRI image samples, Indicates that from The structure and image feature vector extracted from the MRI image, Indicates the The true value of the bone-related representation value of the sample, The mean square error term represents the value of the predicted bone density and the actual bone quality correlation characterization value, denotes the L2 regularization term, is the square norm of the model parameter vector, is the regularization weight coefficient. The regularization coefficient is determined by cross-validation, and is usually set between 10^(-3) and 10^(-5) to balance the fitting ability of the model on the training set and the generalization performance on the test set.

[0013] Further, in the above step S3.3, based on the constructed loss function, an optimization algorithm based on gradient descent is used to iteratively update the network parameters , and through the error back propagation mechanism, the loss function is continuously minimized until the model converges on the training set, thereby forming a stable mapping relationship between the MRI image structure feature space and the bone density continuous numerical space.

[0014] Further, in the above step S4, the fusion of MRI image features and non-image information includes the following sub-steps: S4.1: The original non-image information including age, gender, height and weight is normalized and numerically processed, and mapped into a fixed-dimensional embedding vector so that its dimension can be spliced with the MRI image features; S4.2: Introduce a cross-modal alignment network module to perform affine transformation on the MRI image feature vector and the non-image feature vector respectively, so that they have comparability in the same semantic space; S4.3: Introduce an attention weight gating mechanism between image modalities and non-image modalities, automatically learn the modalities followed by the current sample according to the training data, and give different fusion weights according to their relevance; S4.4: Introduce a modal consistency constraint loss term to make the prediction results of image modalities and non-image modalities consistent in the embedding space, thereby improving the stability of feature collaboration.

[0015] Further, in the above step S5, when a new MRI image is input in the model inference stage, the corresponding patient non-image information (including age, gender, height, weight, etc.) is also obtained, which is respectively processed by the image preprocessing process consistent with the training stage, feature encoding and non-image data normalization, and spliced to form a joint feature vector as the model input. The joint feature input ensures that the input structure of the model in the inference stage is consistent with that in the training stage, so as to obtain the corresponding bone-related representation value output. The bone-related representation value has good cross-device comparability, and compared with the absolute bone density value, the bone-related representation value can be standardized and compared between different MRI systems and used as the final bone density output index.

[0016] A system for osteoporosis risk assessment based on lumbar magnetic resonance images, comprising an image acquisition module, a structure processing module, a feature extraction module, a bone density prediction module, and an inference control module; The image acquisition module is used to acquire T1-weighted and T2-weighted MRI image data of the lumbar spine area; The structure processing module is used to perform tissue structure recognition and region extraction on the MRI image data, automatically locate the vertebral region and generate standardized image input; The feature extraction module is used to extract local image feature vectors of the vertebral area based on a sliding window mechanism and encode the features into multi-dimensional feature vectors; The bone density prediction module is used to construct and call a reverse mapping model based on a deep neural network, input the feature vector and output the corresponding bone density estimation value; The inference control module is used to call the trained mapping model in the inference stage and perform consistent preprocessing, encoding and bone-related representation value output processes on newly input MRI image samples.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. By constructing a deep mapping model between MRI image structural features and bone-related characterization values, this invention can achieve quantitative extraction of lumbar vertebrae bone tissue imaging parameters under radiation-free conditions. This provides a safe and repeatable opportunistic analysis method for imaging monitoring and follow-up of high-risk populations, thereby enhancing the research and application value of MRI data. 2. The present invention significantly enhances the adaptability of the model in multi-center data through strategies such as multi-dimensional image feature vector extraction, residual connection neural network modeling, and regularized loss function training, avoids overfitting, and improves the consistency and reliability of the calculated results of bone characterization parameter values ​​between different devices and samples, thereby ensuring the robustness and scalability of the method in cross-center applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0020] Example 1, as Figure 1 As shown, the present invention discloses a method for osteoporosis risk assessment based on lumbar magnetic resonance images, comprising the following steps: S1: lumbar spine MRI image acquisition and structural region standardization processing, the original T1 weighted and T2 weighted image data of the lumbar spine region is collected through the standard magnetic resonance imaging sequence, the image data is preliminarily organized and structured, including removing the artifacts in the imaging data, unifying the image gray value, and resampling the size in the resolution space, then the vertebral body region corresponding to the lumbar segment is identified in the image data through a model based on spatial anatomical structure or an adaptive feature extraction algorithm, a set of fixed vertebral body slice selection rules is used to extract the cancellous bone area inside the vertebral body to remove the cortical bone for subsequent feature modeling, forming the vertebral body image data for subsequent modeling calculation, the fixed vertebral body slice selection rule is specifically to select the middle slice of each vertebral body height in the sagittal sequence, in addition, another embodiment is given: select the range of ±1 layer of the central layer of the vertebral body in the transverse sequence. The MRI image data is a multi-layer two-dimensional sequence image in standard DICOM format, constituting a three-dimensional volume data, in the image standardization processing, the gray scale normalization and size resampling operation are carried out based on the three-dimensional voxel space, to ensure the comparability of the input data of each sample in the spatial structure and intensity distribution.

[0021] Further, the T1 weighted image and the T2 weighted image in step S1 are processed in a separate manner, both of which first independently enter their respective feature extraction channels to maintain their differences in imaging physical characteristics and structural information, wherein the T1 weighted image focuses on extracting low-frequency gray statistical features and overall structure indicators that can reflect the fat content distribution and density changes inside the vertebral body, and the T2 weighted image focuses on extracting high-frequency image feature vectors and multi-scale fractal parameters related to water content differences and microstructure directionality. After extracting two types of feature vectors, the T1 features and the T2 features are encoded and merged in a unified feature space through a feature fusion operator to obtain a fused MRI image feature vector, which is input into a bone density back mapping model to calculate the bone density value of the target region, thereby realizing the inference calculation of bone density while preserving the key information of each imaging mode.

[0022] Further, for the artifacts in step S1, common artifacts in MRI imaging include motion artifacts, metal artifacts, chemical shift artifacts, etc., which will interfere with subsequent image analysis. Existing removal methods include frequency domain filtering-based processing (such as filtering out periodic artifact signals after fast Fourier transform), model-based artifact separation algorithms (using the difference between the spatial distribution pattern of artifacts and the real signal for decomposition), and deep learning-based artifact suppression networks (inputting original images and outputting artifact-eliminated images automatically restored by trained models). The specific method can be selected according to the characteristics of the MRI device and the type of artifact to minimize the interference signal while preserving structural details.

[0023] Exemplarily, a model building process based on spatial anatomy is given: A three-dimensional anatomical template containing the position, shape and adjacent structure relationship of the vertebral body is established, and the template is aligned with the actual MRI image through image registration, so as to mark the position and contour of each vertebral body.

[0024] Exemplarily, an application process of an adaptive feature extraction algorithm is given: Without relying on fixed templates, image segmentation and feature detection techniques are used to automatically detect the vertebral body region through gray scale distribution, edge morphology, texture pattern and other information. The rectangular structure with regular shape and separated by low signal intervertebral disc above and below is detected and identified as the vertebral body region.

[0025] S2: Feature representation and coding of image structure texture information. After obtaining the standardized vertebral body image data, a sliding window mechanism is used to scan and process the local pixel blocks in the vertebral body image region. The sliding window uses a fixed size (such as 32x32 pixels) to scan the vertebral body image in a tiled manner, and the window step is set to 1 / 2 of the window length to ensure complete coverage of local features with a certain overlap area. For each local image structure, a set of mixed features including tissue structure contrast features based on gray level co-occurrence matrix of vertebral body region, trabecular bone structure stability features based on local binary pattern within vertebral body boundary, fractal dimension image feature vector reflecting the complexity of vertebral body interior, bone quality distribution energy features based on multi-scale wavelet decomposition of vertebral body region, and Gabor response features for directional vertebral body image structure, etc. are extracted. Each type of image feature vector is uniformly coded after extraction to form a consistent multi-dimensional feature vector, which serves as the basis for subsequent model input.

[0026] Specifically, after obtaining the standardized vertebra image data, a sliding window mechanism is used to perform feature scanning processing on the local pixel blocks in the vertebra image region. The sliding window uses a fixed size (such as 32x32 pixels) to perform a tiled scan on the vertebra image, and the window step is set to 1 / 2 of the window side length to ensure that the local features cover the entire structure and have a certain overlap. For the local image structure at each position, the contrast features based on the organization structure of the vertebra region gray level co-occurrence matrix are formed by calculating the local pixel gray level co-occurrence matrix and extracting the contrast, correlation, energy and entropy statistics; the trabecular structure stability features based on the local binary pattern within the vertebra boundary are obtained by generating a local binary pattern by comparing the local pixel with the neighborhood pixel, and then obtaining the features by counting the pattern distribution and direction consistency; the fractal dimension image feature vector reflecting the complexity of the vertebra interior is obtained by calculating the fractal dimension through multi-scale box counting on the local pixel block; the bone mass distribution energy features based on the multi-scale wavelet decomposition of the vertebra region are obtained by calculating the energy of the different frequency sub-band coefficients and combining them to form the features through two-dimensional wavelet decomposition on the local pixel block; the Gabor response features for the directional structure of the vertebra image are obtained by applying multi-directional and multi-scale Gabor filters to the local pixel block, calculating the amplitude response mean and variance to form the features, and each type of image feature vector is uniformly encoded to form a multi-dimensional feature vector with consistent structure after extraction, which is used as the basis for subsequent model input.

[0027] The organization structure contrast features based on the vertebra region gray level co-occurrence matrix are used to reflect the gray level change rule and organization detail complexity of the vertebra interior; the trabecular structure stability features based on the local binary pattern within the vertebra boundary are used to describe the repetition pattern and direction consistency of the microstructure of the vertebra region; the fractal dimension image feature vector reflecting the complexity of the vertebra interior is used to depict the spatial roughness and structural disorder degree of the bone tissue at multiple scales; the bone mass distribution energy features based on the multi-scale wavelet decomposition of the vertebra region are used to express the spatial distribution rule of the bone tissue texture at different frequency components; the Gabor response features for the directional structure of the vertebra image are used to capture the arrangement features and striped structure differences of the bone tissue in multiple directions.

[0028] S3: The bone density inference model based on the inverse mapping structure is constructed. On the basis of completing the encoding of the image structure features, the text-based basic information of the subject (such as age, gender, height, weight) is introduced, and it is spliced ​​with the image feature vector to form a joint multimodal input feature to construct a multimodal feature input space. The potential bone density information of the MRI image is supervised and fitted by the inverse mapping model, so that the model establishes a continuous mapping relationship between the MRI image and the real value of bone density in the joint feature space. The inverse mapping model takes the multidimensional image feature vector and the digitized non-image information vector as input, and the real bone-related representation value of the MRI image pair as the supervision signal, and is trained using a deep neural network with residual connection and feature attention mechanism. A feature channel weighting module is introduced into the model to improve the recognition ability of the multimodal fusion feature, and the model parameters are iteratively updated through the mean square error loss function combined with the L2 regularization term, and finally a stable multimodal mapping function is formed for the bone density numerical output in the subsequent reasoning stage. Among them, the mapping function model from the feature space to the real value of bone density is Its structure adopts a neural network composed of multi-layer linear transformation and nonlinear activation function, and introduces a residual connection structure to stabilize the training process. At the same time, it configures a feature channel weighting module to enhance the discrimination of structural information. In the initial stage, the network parameters are adjusted. Set according to Gaussian distribution or He initialization rule, the formula is: ; in, represents all trainable parameters of the mapping network, Represented as the model pair The bone-related characterization values ​​predicted by the samples.

[0029] The inverse mapping model is a deep neural network structure, the input is the feature vector of the MRI image, and the output is the corresponding bone density value. The construction of the inverse mapping model includes the following sub-steps: S3.1: Create a training sample dataset containing MRI image feature vectors, patient non-image information, and paired bone-related characterization values. The image and non-image information are jointly encoded to form input samples. The image feature vector The corresponding bone-related characterization value Pairing to form a structured supervised training sample set , the formula is: ; in, For the Image feature vectors extracted from MRI images, reference values for bone quality related characterization values measured by DXA system, denotes the total number of training samples.

[0030] S3.2: Construct a loss function to fit the continuous mapping relationship between the MRI image features and bone density by supervised learning mode targeting the measured bone density values. On the basis of the structured supervised training sample set Define the supervised objective function of the training process, use mean square error as the main loss function, and introduce L2 regularization term to limit the model complexity to suppress the overfitting tendency. The loss function formula is: wherein, denotes the total number of training samples, denotes the structural features extracted by the neural network model for the th MRI image sample, denotes the structural and image feature vectors extracted from the th MRI image, denotes the true value of the bone quality related characterization value of the th sample, denotes the mean square error between the predicted bone density and the true bone quality related characterization value, denotes the L2 regularization term, is the square norm of the model parameter vector, is the regularization weight coefficient.

[0031] S3.3: Iteratively optimize the parameters of the back-propagation mapping model by error back-propagation.

[0032] On the basis of the constructed loss function, the optimization algorithm based on gradient descent is used to iteratively update the network parameters , through the error back-propagation mechanism, continuously minimize the loss function, until the model reaches the convergence state on the training set, thereby forming a stable mapping relationship between the MRI image structural feature space and the bone density continuous numerical space.

[0033] S4: Input the patient's non-image information as an additional modality, and input it into the model together with the image feature vector to form a transverse multi-modal fusion feature for enhancing the bone density estimation performance.

[0034] ​To improve the individual accuracy of bone density estimation, on the basis of image feature input, further introduce the non-image information of the patient for transverse multi-modal fusion modeling. The non-image information includes but is not limited to age, gender, height, weight, BMI, past medical history (such as fracture history, diabetes, etc.), laboratory indicators (such as blood calcium, vitamin D level), etc. In implementation, image modal feature channels and non-image modal feature channels are constructed respectively.

[0035] Specifically, the image modal feature channel is the image feature vector extracted based on the MRI image . The non-image modal feature channel is to construct a patient attribute feature vector , and all features are processed by numerical normalization or embedding representation. The final input model fusion feature vector is: ; The fused feature vector is input into the back-propagation mapping neural network model to establish the nonlinear mapping relationship between the MRI image structure and the non-image individual feature to the bone density estimation value. To ensure the stability of the training process, a weight regulation mechanism is designed for the two modal feature branches in the loss function optimization process, and Dropout is introduced for the non-image feature channel to reduce the risk of overfitting.

[0036] The fusion of MRI image features and non-image information includes the following sub-steps: S4.1: Non-image information coding representation, the original non-image information including age, gender, height and weight is normalized and numerical processed, and is mapped into a fixed dimension embedding vector , so that its dimension can be spliced with the MRI image feature. Part of the discrete variables (such as gender) uses one-hot encoding, and part of the continuous variables (such as BMI) uses z-score standardization.

[0037] , represents the non-image information vector corresponding to the th MRI image sample, including the structured data of the patient's age, gender, height, weight, BMI, etc. After normalization and numerical coding, it is represented as a fixed dimension feature vector.

[0038] S4.2: Cross-modal feature alignment mapping, a lightweight cross-modal alignment network module is introduced to perform affine transformation on the MRI image feature vector and the non-image feature vector , so that they have comparability in the same semantic space, and the formula is: ; ; Wherein, represents the non-image information vector corresponding to the The MRI image feature vector of samples (obtained by sliding window texture extraction), , represents the weight matrix, which is used to map image / non-image features to the shared semantic space. , represents the bias vector, , Represents the aligned image / non-image feature vector.

[0039] S4.3: Modality Adaptive Weight Control Mechanism. To improve the interpretability and adaptability of the model, an attention weight gating mechanism is further introduced between the image modality and the non-image modality. The modality to which the current sample adheres is automatically learned based on the training data, and different fusion weights are assigned according to their relevance. The formula is: ; ; in ∈[0,1], represents the importance weight of the MRI modality, which is adaptively learned by neural network training and finally fused with features As a mapping model enter.

[0040] represents the concatenation of image and non-image feature vectors, 、 represents the weights and biases used for attention weight calculation, Represents the Sigmoid activation function, which limits the output to between 0 and 1. Indicates the The attention weight of a sample on the MRI modality reflects the degree of dependence of the model on image features. represents the weight corresponding to the non-image modality; Indicates the The fused feature vector of each sample combines image and non-image features, is interpretable, automatically adjusts the influence of different modalities, and improves the model's adaptability to individual differences.

[0041] S4.4: Joint loss optimization. In addition to the main loss function, a modality consistency constraint loss term is introduced into the overall loss function to encourage the consistency of the prediction results of the image modality and the non-image modality in the embedding space, thereby improving the feature collaboration stability. The expression formula of the joint loss function is: ; in, represents the final joint loss function, represents the main loss function (such as the MSE between the predicted value and the true BMD, respectively represent the prediction sub-models based on MRI features only and non-image features only, represent the square of the L2 norm, represents the weighting coefficient of the modality consistency loss term, and further controls its influence in the overall optimization target.

[0042] S5: Non-invasive bone density numerical output and image reverse mapping application, after the model is trained, input a new MRI image sample, and simultaneously collect the non-image information (such as age, gender, height, weight, etc.) of the patient corresponding to the sample, generate a structured image feature vector through the image preprocessing and image feature vector extraction process consistent with the training stage, and at the same time, standardize or embed the non-image information. Subsequently, the image and non-image features are spliced and fused to form a multi-modal joint feature vector, which is input into the trained multi-modal reverse mapping model to calculate the corresponding bone-related representation value, and finally a direct mapping relationship from the joint space of the MRI image and the non-image information to the continuous bone density value is formed, thereby realizing the continuous numerical output and subsequent quantitative evaluation of the bone density of the target region.

[0043] In the model inference stage, the multi-modal joint input structure is consistent with the training stage to ensure the stability and comparability of the prediction results. Since the bone-related representation value is a standardized bone density indicator, it has good device migration, so the bone-related representation value is used as the final output indicator for clinical evaluation of osteoporosis risk and efficacy tracking.

[0044] The bone representation parameter value is a continuous quantity calculated by the model in the multi-modal feature space, used to represent the fusion mapping result of the MRI image and non-image information in the bone-related dimension. This parameter value reflects the comprehensive representation of bone tissue in structure, texture, and potential density characteristics, and can be used as data output for subsequent research, data analysis, or clinical auxiliary reference, but does not directly constitute a medical diagnosis conclusion.

[0045] In embodiment two, based on embodiment one, a specific working principle of a method for evaluating the risk of osteoporosis in lumbar MRI images is proposed.

[0046] A system for a method of evaluating the risk of osteoporosis in lumbar MRI images, comprising an image acquisition module, a structure processing module, a feature extraction module, a bone density prediction module, and an inference control module. The image acquisition module is used to acquire T1-weighted and T2-weighted MRI image data of the lumbar region. The structure processing module is used to identify and extract the organizational structure of the MRI image data, automatically locate the vertebral body region, and generate a standardized image input. ​The feature extraction module is used for extracting local image feature vectors of the vertebral body region based on a sliding window mechanism, and encoding the features into a multi-dimensional image feature vector; at the same time, non-image information (including age, gender, height, weight, etc.) of the patient is obtained and normalized, and is spliced with the image feature vector to form a joint input feature vector; The bone density prediction module is used for constructing and calling a multi-modal back-propagation mapping model based on a deep neural network, inputting the joint feature vector into the trained model, and outputting the estimated value of the bone-related representation value; The inference control module is used for performing image preprocessing, feature encoding and multi-modal fusion processes consistent with the training stage on the newly input MRI image sample and corresponding non-image information in the model inference stage, ensuring the consistency of the prediction structure, and outputting the final bone-related representation value.

[0047] In the third embodiment, based on the first embodiment, an actual application case of the lumbar MRI osteoporosis risk assessment system and method is provided.

[0048] The application can extract standard vertebral body regions based on existing T2 weighted images during the lumbar MRI examination process without additional imaging equipment, and output continuous bone density estimation values through a deep neural network model in combination with image structure features and patient non-image information (such as age, gender, height, weight, etc.). Image preprocessing: grayscale normalization, spatial resampling and vertebral body segmentation are performed on the T2 weighted image; Feature extraction: local image feature vectors are extracted and encoded into multi-dimensional image feature vectors by sliding window scanning; Non-image information processing: normalize and numerize or embed and encode the patient's age, gender, height, weight, etc. Multi-modal fusion and prediction: image and non-image features are spliced and input into the trained multi-modal back-propagation mapping model to output the estimated value of the bone density bone-related representation value.

[0049] In this embodiment, a multi-center sample data set is constructed, and the inclusion criteria are that the subjects have paired lumbar MRI and DEXA detection records within 30 days, and the exclusion criteria include long interval between MRI and DEXA examination, vertebral body structure malformation, post-internal fixation, etc. Through the data of multiple hospitals, the MRI sequence all uses T2 weighted image. The system uses multi-slice images to complete vertebral body segmentation and spinal canal segmentation, and calculates the VBQ value (vertebral body average grayscale and cerebrospinal fluid grayscale ratio). After the VBQ, image feature vector and patient basic information are fused, they are input into the multi-modal regression model to predict the bone-related representation value and Z-score measured by DEXA.

[0050] The results show that on the internal validation set, the mean absolute error (MAE) is <0.35, and >0.78; The external validation set is slightly higher but still acceptable, indicating strong generalization ability.

[0051] The above results show that the system can be applied to large-scale non-invasive screening based on MRI through the multi-modal structure of image + non-image joint input, expand the means of osteoporosis identification, and realize opportunistic risk warning.

[0052] Bone density prediction and post-treatment follow-up evaluation scenarios: As a chronic disease, the treatment plan for osteoporosis usually includes long-term medication of calcium, vitamin D, bisphosphonate, etc. The efficacy evaluation needs to rely on continuous bone density detection. The traditional DEXA is subject to radiation and equipment frequency. The multi-modal model based on MRI image and patient information of the present application can provide stable and continuous bone density estimation results.

[0053] In this embodiment, subjects who have received anti-osteoporosis treatment are selected, and MRI acquisition and patient information input are performed before and after treatment, respectively. The system is applied to calculate the predicted values of bone density and bone quality related representation values twice, and compared with the paired DEXA results. The verification results show that the change trend of the predicted values is highly consistent with DEXA (R=0.85, P<0.001). At the same time, the system is also suitable for multi-center remote follow-up management because it does not depend on specific equipment and operators.

[0054] In summary, the present application can not only be used for opportunistic screening of asymptomatic people, but also be widely used for individualized follow-up evaluation of diagnosed patients. The first is the opportunistic screening scenario of osteoporosis, and the second is the bone density follow-up prediction scenario of the population after receiving treatment. The system of the present application has practicality and expansibility in clinical osteoporosis management, and has high practicality and popularization potential.

[0055] Of course, the present application can have other various embodiments. Those skilled in the art can make various corresponding changes and modifications to the present application without departing from the spirit and essence of the present application. However, these corresponding changes and modifications should all belong to the protection scope of the claims attached to the present application.

Claims

1. A method for osteoporosis risk assessment based on lumbar vertebrae magnetic resonance imaging, characterized in that: The following steps are involved: S1: T1-weighted and T2-weighted image data of the lumbar spine are collected through magnetic resonance imaging sequences, and preliminary tissue structure recognition and region recognition are performed on the image data to extract the vertebral image region as input; S2: performing a sliding window scan on the vertebral image area, extracting a local image feature vector and encoding it into a multi-dimensional feature vector; S3: The multidimensional feature vector and the patient's non-image information are jointly input to construct an inverse mapping model containing multimodal features, and the potential bone density information of the MRI image is fitted and learned based on the multimodal model, so that the model can establish a mapping relationship between the MRI image and the real value of the bone density in the joint feature space; S4: The patient's non-image information is used as an additional modality input and is jointly input into the model with the image feature vector to form a horizontal multimodal fusion feature; S5: Input a new MRI image sample and extract the image feature vector, while obtaining the non-image information of the corresponding patient. The image features and non-image information are jointly input into the trained model, and the corresponding bone-related representation value is output.

2. The method for osteoporosis risk assessment based on lumbar magnetic resonance imaging according to claim 1, wherein: In the above step S1, the step of extracting the vertebral image region includes performing grayscale normalization processing and spatial resolution resampling processing on the image data, and automatically identifying the vertebral boundary region of the lumbar segment by a region segmentation method based on an anatomical structure model.

3. The method for osteoporosis risk assessment based on lumbar MRI according to claim 1, wherein: In the above step S2, the local image feature vectors extracted by the sliding window scanning include tissue structure contrast features based on the grayscale co-occurrence matrix of the vertebral region, trabecular structure stability features based on the local binary pattern within the vertebral boundary, fractal dimension image feature vectors reflecting the internal complexity of the vertebral body, bone distribution energy features based on multi-scale wavelet decomposition of the vertebral region, and Gabor response features for the directionality of the vertebral image structure. After extraction, various image feature vectors are uniformly encoded to form a multidimensional feature vector with consistent structure, which serves as the basis for subsequent model input.

4. The method for osteoporosis risk assessment based on lumbar MRI according to claim 1, wherein: In the above step S3, the mapping function model from feature space to real value of bone density is: Its structure adopts a neural network composed of multi-layer linear transformation and nonlinear activation function, and introduces a residual connection structure to stabilize the training process. At the same time, it configures a feature channel weighting module to enhance the discrimination of structural information. In the initial stage, the network parameters are adjusted. Set according to Gaussian distribution or He initialization rule, the formula is: ； in, represents all trainable parameters of the mapping network, Represented as the model pair The bone-related characterization values ​​predicted by the samples.

5. The method for osteoporosis risk assessment based on lumbar MRI according to claim 1, wherein: In step S3 above, the inverse mapping model is a deep neural network structure, the input is the feature vector of the MRI image, and the output is the corresponding bone density value. The construction of the inverse mapping model includes the following sub-steps: S3.1: Create a training sample dataset containing MRI image feature vectors, patient non-image information, and paired bone-related characterization values. The image and non-image information are jointly encoded to form input samples. S3.2: Using supervised learning with measured bone density as the target, a loss function is constructed to fit the continuous mapping relationship between MRI image features and bone density. S3.3: Iteratively optimize the parameters of the inverse mapping model through error backpropagation.

6. The method for osteoporosis risk assessment based on lumbar MRI according to claim 5, characterized in that: In the above step S3.1, the image feature vector The corresponding bone-related characterization value Pairing to form a structured supervised training sample set , the formula is: ； in, For the Image feature vectors extracted from MRI images, is the reference value of bone-related characterization values ​​measured by DXA system, Indicates the total number of training samples.

7. The method for osteoporosis risk assessment based on lumbar MRI according to claim 6, wherein: In the above step S3.2, in the structured supervised training sample set Based on this, the supervised objective function of the training process is defined, and the mean square error is used as the main loss function. At the same time, the L2 regularization term is introduced to limit the complexity of the model to suppress the tendency of overfitting. The loss function formula is: ； in, represents the total number of training samples, Represents the neural network model for the The structural features extracted from MRI image samples, Indicates that from The structure and image feature vector extracted from the MRI image, Indicates the The true value of the bone-related representation value of the sample, It represents the mean square error term between the predicted bone density and the actual bone quality characterization value, represents the L2 regularization term, is the square norm of the model parameter vector, is the regularization weight coefficient.

8. The method for osteoporosis risk assessment based on lumbar vertebrae magnetic resonance images according to claim 7, characterized in that: In the above step S3.3, based on the constructed loss function, the network parameters are optimized using the gradient descent optimization algorithm. Iterative updates are performed, and the loss function is continuously minimized through the error back propagation mechanism until the model reaches convergence on the training set, thereby forming a stable mapping relationship between the MRI image structural feature space and the continuous numerical space of bone density.

9. The method for osteoporosis risk assessment based on lumbar MRI according to claim 1, wherein: In the above step S4, the fusion of MRI image features and non-image information includes the following sub-steps: S4.1: Normalize the original non-image information including age, gender, height and weight and map them into fixed-dimensional embedding vectors And perform splicing operation with MRI image features; S4.2: Introduce a cross-modal alignment network module to MRI image feature vectors and non-image feature vectors Perform affine transformation respectively; S4.3: Introduce an attention weight gating mechanism between image modalities and non-image modalities, automatically learn the modality to which the current sample adheres based on the training data, and assign different fusion weights based on their relevance; S4.4: Introduce a modality consistency constraint loss term to make the prediction results of image modality and non-image modality consistent in the embedding space.

10. A system for osteoporosis risk assessment based on lumbar vertebrae magnetic resonance images, the system being implemented based on the method for osteoporosis risk assessment based on lumbar vertebrae magnetic resonance images according to any one of claims 1 to 9, characterized in that: It includes image acquisition module, structure processing module, feature extraction module, bone density prediction module and reasoning control module; The image acquisition module is used to acquire T1-weighted and T2-weighted MRI image data of the lumbar spine area; The structure processing module is used to perform tissue structure recognition and region extraction on the MRI image data, automatically locate the vertebral region and generate standardized image input; The feature extraction module is used to extract local image feature vectors from the vertebral area based on a sliding window mechanism, encode the features into multidimensional image feature vectors, and simultaneously digitize the patient's non-image information to form joint input features; The bone density prediction module is used to build and call a reverse mapping model based on a deep neural network, taking the multimodal joint features of image feature vectors and non-image information as input and outputting the estimated value of the corresponding bone-related representation value; The inference control module is used to call the trained mapping model during the inference phase, perform consistent preprocessing, joint encoding, and prediction processes on newly input MRI image samples and patient non-image information, and output bone-related representation values.

Images