Artificial intelligence-based sarcopenia-osteoporosis diagnosis and fracture risk prediction method

Abstract

The application is suitable for the technical field of medical artificial intelligence and medical image analysis, and provides a sarcopenia-osteoporosis diagnosis and fracture risk prediction method based on artificial intelligence, which comprises the following steps: preprocessing the clinical structured data to obtain clinical feature data; performing feature analysis on the quantitative CT image data to obtain image feature data; performing feature fusion on the clinical feature data and the image feature data to obtain fusion feature data; and finally obtaining a fracture risk grade and a sarcopenia-osteoporosis auxiliary diagnosis result. The application has the advantages that the combined auxiliary diagnosis capability for the bone metabolism state and the sarcopenia state can be improved, and the information loss caused by a single data source can be reduced; the risk probability and the risk grade of a fracture event within a preset time period can be output, and a more forward-looking risk stratification basis can be provided for the clinic; and the diagnosis objectivity, consistency and risk prediction capability can be improved.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence and medical image analysis technology, and in particular relates to an artificial intelligence-based method for the diagnosis of sarcopenia-osteoporosis and the prediction of fracture risk. Background Technology

[0002] As the population ages, the incidence of sarcopenia and osteoporosis is gradually increasing among middle-aged and elderly individuals. Clinical research and practice have shown a strong correlation between decreased muscle mass and function and reduced bone mass and strength. When these two conditions occur together, they not only exacerbate the decline in mobility but also significantly increase the risk of falls and fractures. Therefore, how to comprehensively assess muscle and bone metabolic status and, based on this assessment, predict fracture risk has become a crucial issue in the management of degenerative musculoskeletal diseases in the elderly.

[0003] In current clinical practice, sarcopenia and osteoporosis are typically assessed separately. The diagnosis of sarcopenia relies heavily on physical measurements, functional assessments, and some imaging parameters, while the diagnosis of osteoporosis relies more on bone mineral density tests, bone metabolism-related biochemical indicators, and imaging results. Because the diagnostic processes for these two diseases are relatively independent, there is a lack of unified integration and analysis between clinical structured data and medical imaging data. This often fails to comprehensively reflect the overall state of coordinated muscle and bone degeneration, thus limiting the accuracy and consistency of combined diagnostic results.

[0004] Furthermore, quantitative CT imaging can provide a variety of objective information related to bone and muscle status, such as vertebral bone mineral density, paraspinal muscle cross-sectional area, fat infiltration rate, and muscle grayscale. Clinical structured data also includes important information such as age, sex, weight, symptoms and signs, bone mineral density test results, and bone metabolism biochemical indicators. However, current technologies primarily utilize these multimodal data through individual analysis or simple overlay, resulting in a low degree of fusion between clinical structured data and Q-CT image features. This leads to the data's value not being fully realized, and diagnostic and assessment results still largely rely on human experience.

[0005] In fracture risk assessment, existing methods focus more on limited indicators such as single bone mineral density levels or past fracture history, and lack sufficient joint modeling of muscle status, bone metabolism status, and quantitative imaging characteristics. Therefore, they suffer from insufficient accuracy and stability in early risk identification and stratified prediction. Particularly in scenarios where sarcopenia and osteoporosis co-occur, current technologies lack a unified approach that can simultaneously utilize structured clinical data and quantitative CT imaging data to assist in the diagnosis of sarcopenia-osteoporosis and further output fracture risk probability and risk level. Summary of the Invention

[0006] The purpose of this application is to provide an artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk, aiming to solve the problem that the existing technology lacks a unified technical solution that can simultaneously utilize clinical structured data and quantitative CT imaging data to assist in the diagnosis of sarcopenia and further output the probability and level of fracture risk.

[0007] This application provides an artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk, comprising:

[0008] Acquire the clinical structured data and quantitative CT image data of the subjects to be analyzed;

[0009] The clinical structured data is preprocessed to obtain clinical feature data; the quantitative CT image data is subjected to feature analysis to obtain image feature data.

[0010] The clinical feature data and the imaging feature data are fused to obtain fused feature data;

[0011] The fused feature data is input into the joint auxiliary diagnostic model, which outputs bone metabolism status results and sarcopenia status results. Based on the bone metabolism status results and the sarcopenia status results, an auxiliary diagnostic result for sarcopenia-osteoporosis is obtained.

[0012] The fused feature data is input into the fracture risk prediction model, which outputs the risk probability of a fracture event occurring within a preset time period. The fracture risk level is obtained based on the risk probability.

[0013] Preferably, the method for preprocessing the clinical structured data to obtain clinical feature data includes:

[0014] The original case information is structured and organized according to a preset field dictionary;

[0015] Continuous indicators are retained as numerical variables, binary symptoms and signs are converted into 0 / 1 coded variables, and graded conclusions are converted into ordinal or unique heat coded variables; text-based examination conclusions are converted into predefined category codes.

[0016] After the data has been encoded, data cleaning, outlier identification, missing value handling, and numerical normalization are performed to obtain clinical feature data.

[0017] Preferably, the method for performing feature analysis on the quantitative CT image data to obtain image feature data includes:

[0018] The quantitative CT image data is converted into a model input tensor.

[0019] The target layer is located by performing target layer localization on the input tensor of the model to obtain the anatomical layer to be analyzed;

[0020] The vertebral body region and paravertebral muscle region in the anatomical plane are located and segmented based on an artificial intelligence segmentation model.

[0021] Based on the localization and segmentation results, image quantitative features are extracted. The image quantitative features include at least one or more of the following: bone density features of the vertebral body region, cross-sectional area of ​​the paraspinal muscles, fat infiltration rate, and muscle grayscale statistical features.

[0022] The image feature data is generated based on the image quantization features.

[0023] Preferably, the method for fusing the clinical feature data and the imaging feature data to obtain fused feature data includes:

[0024] The clinical feature data is encoded into a clinical feature vector;

[0025] The image feature data is encoded into an image feature vector;

[0026] The clinical feature vector and the image feature vector are respectively mapped to the feature space to obtain a unified dimension of clinical potential representation and image potential representation;

[0027] Dimensional alignment is performed between the clinical latent representation and the image latent representation;

[0028] The dimensionally aligned clinical latent representation and the image latent representation are concatenated or weighted and summed to obtain the fused feature data.

[0029] Preferably, the method for inputting the fused feature data into a joint auxiliary diagnostic model and outputting bone metabolism status results and sarcopenia status results, and obtaining sarcopenia-osteoporosis auxiliary diagnostic results based on the bone metabolism status results and the sarcopenia status results includes:

[0030] The fused feature data is input into the bone metabolism status determination branch, and the bone metabolism status result is output as normal, reduced bone mass, or osteoporosis.

[0031] The fused feature data is input into the sarcopenia status determination branch, and the sarcopenia status result is output as either sarcopenia present or sarcopenia absent.

[0032] When the bone metabolism status result is osteoporosis and the sarcopenia status result is sarcopenia, the auxiliary diagnostic result is determined to be sarcopenia-osteoporosis.

[0033] When the bone metabolism status result is osteopenia and the sarcopenia status result is sarcopenia, the auxiliary diagnostic result is determined to be sarcopenia combined with low bone mass.

[0034] Preferably, the method for inputting the fused feature data into the fracture risk prediction model, outputting the risk probability of a fracture event occurring within a preset time period, and obtaining the fracture risk level based on the risk probability is as follows:

[0035] The fused feature data is then normalized and subjected to consistency processing.

[0036] The fused feature data after normalization and consistency processing is input into the fracture risk prediction model to obtain the risk probability of the subject to be analyzed occurring a fracture event within the preset time period.

[0037] The risk probability is compared with a preset risk threshold.

[0038] When the risk probability is less than a first preset threshold, the fracture risk level of the object to be analyzed is determined to be low risk;

[0039] When the risk probability is greater than or equal to the first preset threshold and less than the second preset threshold, the fracture risk level of the object to be analyzed is determined to be medium risk.

[0040] When the risk probability is greater than or equal to the second preset threshold, the fracture risk level of the object to be analyzed is determined to be high risk;

[0041] Output the risk probability and the fracture risk level.

[0042] Preferably, the method for constructing the joint assisted diagnostic model includes:

[0043] Obtain a training sample set containing clinical structured data, quantitative CT image data, and diagnostic labels;

[0044] Preprocessing is performed on the clinical structured data in the training sample set to obtain clinical feature data for training.

[0045] Feature analysis is performed on the quantitative CT image data in the training sample set to obtain training image feature data;

[0046] Perform a feature fusion operation on the training clinical feature data and the training image feature data to obtain training fused feature data;

[0047] The joint auxiliary diagnostic model is trained based on the fusion feature data used for training and the diagnostic labels, so that the joint auxiliary diagnostic model outputs bone metabolism status results and sarcopenia status results.

[0048] Another objective of this application is to provide an artificial intelligence-based device for diagnosing sarcopenia and osteoporosis and predicting fracture risk, the device comprising:

[0049] The raw data acquisition unit is used to acquire the clinical structured data and quantitative CT image data of the object to be analyzed;

[0050] The feature extraction unit is used to preprocess the clinical structured data to obtain clinical feature data; and to perform feature analysis on the quantitative CT image data to obtain image feature data.

[0051] A feature fusion unit is used to perform feature fusion on the clinical feature data and the image feature data to obtain fused feature data;

[0052] The diagnostic result acquisition unit is used to input the fused feature data into the joint auxiliary diagnostic model, output the bone metabolism status result and sarcopenia status result, and obtain the sarcopenia-osteoporosis auxiliary diagnostic result based on the bone metabolism status result and sarcopenia status result.

[0053] The risk level prediction unit is used to input the fused feature data into the fracture risk prediction model, output the risk probability of a fracture event occurring within a preset time period, and obtain the fracture risk level based on the risk probability.

[0054] Another objective of this application is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk as described above.

[0055] Another objective of this application is to provide an artificial intelligence-based system for diagnosing sarcopenia and predicting fracture risk, comprising a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk as described above.

[0056] This application provides an AI-based method for diagnosing sarcopenia and predicting fracture risk. Its key advantage lies in its ability to enhance the combined diagnostic capabilities for bone metabolism and sarcopenia by collaboratively processing structured clinical data and quantitative CT imaging data to form fused feature data, thereby reducing information gaps caused by relying on a single data source. Furthermore, by predicting fracture risk using the fused feature data, it can output the probability and level of fracture events within a preset time period, providing clinicians with a more forward-looking basis for risk stratification. Compared to traditional methods that separate muscle and bone assessments, this application is more conducive to improving diagnostic objectivity, consistency, and risk prediction capabilities. Attached Figure Description

[0057] Figure 1 A flowchart illustrating an artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk, provided for embodiments of this application;

[0058] Figure 2 A structural block diagram of an artificial intelligence-based device for diagnosing sarcopenia and predicting fracture risk, provided for an embodiment of this application;

[0059] Figure 3 This is a block diagram of the internal structure of a computer device in one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] It is understood that the terms "first," "second," etc., used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are used only to distinguish the first unit or module from another unit or module. For example, without departing from the scope of this application, the first unit may be referred to as the second unit, and similarly, the second module may be referred to as the first module.

[0062] like Figure 1 As shown, in one embodiment, an artificial intelligence-based method for diagnosing sarcopenia-osteoporosis and predicting fracture risk is proposed, which may specifically include the following steps:

[0063] Step S10: Obtain the clinical structured data and quantitative CT image data of the subject to be analyzed.

[0064] In this embodiment, the clinical structured data includes at least one or more of the following: demographic information, clinical symptoms and signs, bone mineral density (BMD) test results, and biochemical indicators related to bone metabolism; the quantitative CT imaging data is preferably L3 lumbar spine images. Demographic information may include sex, age, height, and weight; symptoms and signs may include lower back pain, fatigue, decreased height, history of falls, and history of fragility fractures; BMD test results may include BMD values, T-scores, Z-scores, osteopenia or osteoporosis grading conclusions obtained from DXA testing, and volumetric BMD values ​​obtained from QCT testing; biochemical indicators related to bone metabolism may include one or more of ALP, BGP, TRACP, D-Pyr, PINP, and Irisin. By simultaneously acquiring clinical structured data and quantitative CT imaging data, a data foundation is provided for subsequent multimodal collaborative analysis.

[0065] Step S20: Preprocess the clinical structured data to obtain clinical feature data; perform feature analysis on the quantitative CT image data to obtain image feature data.

[0066] In this embodiment, the preprocessing of clinical structured data includes: structuring the original case information according to a preset field dictionary; retaining continuous indicators as numerical variables, converting binary symptoms and signs into 0 / 1 coded variables, converting grading conclusions into ordinal or unique heat coded variables, and converting text-based examination conclusions into predefined category codes; and performing data cleaning, outlier identification, missing value handling, and numerical normalization on the encoded data to obtain clinical feature data. The clinical feature data can be represented as a case-feature two-dimensional matrix. ,in Indicates the number of cases. Indicates the number of clinical features.

[0067] In a preferred embodiment, missing continuous biochemical indicators or measurements can be handled using median imputation, mean imputation, or imputation based on similar samples; fields indicating "not examined" or "not measured" can be retained as independent missing status codes, or corresponding missing indicator variables can be added; blank text fields or cases without valid conclusions can be first marked as "unknown" before encoding. Through the above processing, clinical data from different sources and of different types can be uniformly mapped to computable features.

[0068] In this embodiment, the feature analysis of quantitative CT image data includes: converting the quantitative CT image data into model input tensors, performing target-level localization on the images, automatically locating and segmenting the vertebral body region and paraspinal muscle region based on an artificial intelligence segmentation model, and extracting image quantitative features based on the segmentation results to obtain image feature data. Image quantitative features may include one or more of the following: vertebral body region bone density features, paraspinal muscle cross-sectional area, fat infiltration rate, and muscle grayscale statistical features.

[0069] In a preferred embodiment, the cross-sectional area of ​​the paraspinal muscles can be calculated based on the number of pixels and pixel spacing in the paraspinal muscle region, and the fat infiltration rate can be obtained based on the area ratio of the fat infiltration region to the paraspinal muscle region. The bone mineral density features of the vertebral body region in quantitative CT images differ from the existing bone mineral density values ​​in clinical structured data. The former are image features obtained from automatic segmentation and quantification calculations of the original images, while the latter are structured detection values ​​or grading conclusions from existing examination reports. Both types of information participate in subsequent fusion modeling.

[0070] Step S30: Perform feature fusion on the clinical feature data and the imaging feature data to obtain fused feature data.

[0071] In this embodiment, clinical feature data is first encoded into clinical feature vectors, and image feature data is encoded into image feature vectors. Then, feature space mapping is performed on the clinical feature vectors and image feature vectors respectively to obtain a unified dimension of clinical latent representation and image latent representation. After dimensional alignment of the two, fused feature data is obtained by splicing or weighted summation.

[0072] In a preferred embodiment, the clinical feature vector can be represented as: The image feature vector can be represented as Clinical and radiographic potential representations can be represented as follows: and ,in and They have the same dimensions and correspond to the same fused feature space. Fusion feature data can be adopted... The above method can be used to obtain the modalities through splicing, or through weighted summation in specific implementations. Through this process, the differences in dimensions, dimensionalities, and representation spaces between different modalities are unified.

[0073] Step S40: Input the fused feature data into the joint auxiliary diagnostic model, output the bone metabolism status result and the sarcopenia status result, and obtain the sarcopenia-osteoporosis auxiliary diagnostic result based on the bone metabolism status result and the sarcopenia status result.

[0074] In this embodiment, the combined auxiliary diagnostic model includes at least a bone metabolism status determination branch and a sarcopenia status determination branch. The bone metabolism status determination branch outputs a bone metabolism status result of normal, osteopenia, or osteoporosis; the sarcopenia status determination branch outputs a sarcopenia status result of either the presence or absence of sarcopenia.

[0075] Step S50: Input the fused feature data into the fracture risk prediction model, output the risk probability of a fracture event occurring within a preset time period, and obtain the fracture risk level based on the risk probability.

[0076] In this embodiment, the fused feature data is first normalized and consistency-processed according to the feature processing rules corresponding to the training phase of the fracture risk prediction model. Then, the processed fused feature data is input into the fracture risk prediction model to obtain the probability of a fracture event occurring in the analyzed object within a preset time period. Subsequently, this probability is compared with a preset risk threshold range. When the probability is less than a first preset threshold, the fracture risk level is determined to be low; when the probability is greater than or equal to the first preset threshold and less than a second preset threshold, the fracture risk level is determined to be medium; when the probability is greater than or equal to the second preset threshold, the fracture risk level is determined to be high. Finally, the risk probability and fracture risk level are output.

[0077] The key advantage of this application lies in its ability to enhance the combined auxiliary diagnostic capabilities for bone metabolic status and sarcopenia by collaboratively processing clinical structured data and quantitative CT imaging data to form fused feature data, thereby reducing information gaps caused by relying on a single data source. Furthermore, by using the fused feature data for fracture risk prediction, it can output the probability and severity of fracture events within a preset time period, providing clinicians with a more forward-looking basis for risk stratification. Compared to the traditional approach of treating muscle and bone assessments separately, this application is more conducive to improving diagnostic objectivity, consistency, and risk prediction capabilities.

[0078] In a preferred embodiment, a method for preprocessing the clinical structured data to obtain clinical feature data includes:

[0079] The original case information is structured and organized according to a preset field dictionary;

[0080] Continuous indicators are retained as numerical variables, binary symptoms and signs are converted into 0 / 1 coded variables, and graded conclusions are converted into ordinal or unique heat coded variables.

[0081] Convert the inspection results in text form into predefined category codes;

[0082] After the data has been encoded, data cleaning, outlier identification, missing value handling, and numerical normalization are performed to obtain clinical feature data.

[0083] In this embodiment, original case information from different sources, of varying types, and with inconsistent expression methods is converted into standardized clinical feature data that can be stably processed by subsequent artificial intelligence models. Specifically, the preset field dictionary is used to predefine the name, data type, value range, coding rules, and missing data marking method of each field in the clinical data, so as to ensure that similar information from different cases, different examination sources, and different input formats can be mapped to a unified field. The original case information can come from one or more of electronic medical records, physical examination records, biochemical test reports, bone density test reports, and manual input forms. After being organized by the preset field dictionary, it can form a structured data table with cases as rows and features as columns. Binary symptoms and signs refer to variables that only indicate the presence or absence of two states, such as back pain, fatigue, height shortening, history of falls, and history of fragility fractures. These can be uniformly converted into 0 / 1 coded variables, where 1 indicates the presence of the corresponding symptom or sign, and 0 indicates the absence of the corresponding symptom or sign. Grading conclusions refer to variables with a clear grading order, such as "normal," "osteoporosis," and "osteoporosis." These can be converted to ordinal or one-hot coded variables depending on the actual model type. Ordinal coding is preferred when preserving the grading order; one-hot coding is preferred when the model mistakenly interprets gradations as having a linear distance relationship. For text-based inspection conclusions, such as natural language descriptions like "consider osteoporosis," they can be converted to predefined category codes through keyword extraction, terminology normalization, rule mapping, and manual review to reduce the noise impact of free text on subsequent modeling. After field preparation and coding, the resulting structured data undergoes further data cleaning, outlier identification, missing value handling, and numerical normalization. Data cleaning includes deleting duplicate records, standardizing units of measurement, correcting obvious input errors, and eliminating inconsistent field formats. The outlier identification feature is used to detect abnormal records that exceed physiologically reasonable ranges, statistical distribution ranges, or equipment measurement ranges, such as negative age, significantly distorted height and weight, or biochemical indicators exceeding the equipment's detection limit. Identified outliers can be removed, truncated, corrected, or marked. The missing value handling feature addresses common clinical situations where some examinations are incomplete, results are not recorded, or text is empty. For continuous indicators, median imputation, mean imputation, or imputation based on similar samples can be used.

[0084] In a preferred embodiment, a method for performing feature analysis on the quantitative CT image data to obtain image feature data includes:

[0085] The quantitative CT image data is converted into a model input tensor.

[0086] The target layer is located by performing target layer localization on the input tensor of the model to obtain the anatomical layer to be analyzed;

[0087] The vertebral body region and paravertebral muscle region in the anatomical plane are located and segmented based on an artificial intelligence segmentation model.

[0088] Based on the localization and segmentation results, image quantitative features are extracted. The image quantitative features include at least one or more of the following: bone density features of the vertebral body region, cross-sectional area of ​​the paraspinal muscles, fat infiltration rate, and muscle grayscale statistical features.

[0089] The image feature data is generated based on the image quantization features.

[0090] In this embodiment, the purpose of feature analysis on quantitative CT image data is to extract objective quantitative indicators reflecting bone and muscle states from the original images, thereby avoiding subjective errors and efficiency problems caused by relying solely on manual visual inspection or measurement. Quantitative CT image data can be two-dimensional layered images or three-dimensional image volume data composed of continuous layers. Converting the quantitative CT image data into a model input tensor means organizing the original images into an array format that can be directly received by the artificial intelligence model according to a preset resolution, channel format, and numerical range. For example, converting a single-layer grayscale image into a two-dimensional tensor, or converting multi-layer images stacked into a three-dimensional tensor, so as to facilitate subsequent automatic localization, segmentation, and feature extraction.

[0091] In one specific implementation, target layer localization is used to determine the anatomical layer corresponding to the analysis task from quantitative CT images. Preferably, the anatomical layer to be analyzed is the L3 lumbar spine layer or its adjacent layer. Target layer localization can be achieved through template matching, key anatomical landmark recognition, convolutional neural network classification, or regression localization. By using target layer localization, the processing range of subsequent segmentation models can be narrowed while maintaining the integrity of image information, thereby improving model computational efficiency and region extraction accuracy. For multi-layer image data, layer selection can be performed first, followed by joint analysis of continuous slices near the target layer to improve segmentation stability.

[0092] In this embodiment, the artificial intelligence segmentation model is used to automatically locate and segment the vertebral body region and the paraspinal muscle region. The artificial intelligence segmentation model can employ convolutional neural networks, encoder-decoder networks, U-Net-like networks, or other deep learning models capable of semantic segmentation of medical images. The model input is a preprocessed quantitative CT image tensor, and the model output is a segmentation mask corresponding to the input image. In the segmentation mask, the vertebral body region reflects bone density-related information, and the paraspinal muscle region reflects skeletal muscle quantity and quality-related information. For the paraspinal muscle region, further identification of intramuscular fat infiltration areas can be performed to calculate the fat infiltration rate. After completing the above image quantification feature extraction, various image features can be combined into an image feature vector in a predetermined order, for example, represented as... Each dimension corresponds to bone mineral density features, paraspinal muscle cross-sectional area, fat infiltration rate, and grayscale statistical features, respectively. This image feature vector serves as the image feature data, used in conjunction with clinical feature data for subsequent feature fusion and joint modeling. In this way, raw image information is transformed into structured quantitative features with clear clinical and computational significance.

[0093] In a preferred embodiment, a method for fusing the clinical feature data and the imaging feature data to obtain fused feature data includes:

[0094] The clinical feature data is encoded into a clinical feature vector;

[0095] The image feature data is encoded into an image feature vector;

[0096] The clinical feature vector and the image feature vector are respectively mapped to the feature space to obtain a unified dimension of clinical potential representation and image potential representation;

[0097] Dimensional alignment is performed between the clinical latent representation and the image latent representation;

[0098] The dimensionally aligned clinical latent representation and the image latent representation are concatenated or weighted and summed to obtain the fused feature data.

[0099] In this embodiment, the purpose of feature fusion is to map clinical feature data and imaging feature data from different sources, with different physical meanings, and different dimensions into a unified analysis space, enabling the model to simultaneously utilize clinical and imaging information for joint discrimination. Specifically, clinical feature data can be represented as clinical feature vectors. Each dimension corresponds to age, height, weight, symptoms and signs, bone density-related examination results, and biochemical indicators, respectively; the imaging feature data can be represented as an imaging feature vector. Each dimension corresponds to bone mineral density features, cross-sectional area, fat infiltration rate, and grayscale statistical features calculated from quantitative CT images. Because and The dimensions, numerical ranges, and statistical distributions of these features are usually not consistent, so they cannot be simply superimposed directly. Instead, feature space mapping is required first.

[0100] In one embodiment, clinical feature vectors and image feature vectors can be linearly or nonlinearly mapped to obtain a latent representation with a unified dimension. For example, a fully connected layer mapping method can be used to respectively... and Mapped to:

[0101] ; ;

[0102] in, and The feature mapping matrix, and For bias terms, and These represent the clinical latent representation and the imaging latent representation, respectively. In this way, two types of features that originally had different dimensions can be mapped to latent vectors of the same dimension, thus achieving dimension alignment. After dimension alignment, the clinical latent representation and the imaging latent representation are in a unified feature space, facilitating subsequent fusion operations.

[0103] In this embodiment, the dimension-aligned clinical latent representation and the image latent representation can be fused using either a stitching method or a weighted summation method. When using the stitching method, the two can be concatenated along the feature dimension to obtain fused feature data. The concatenation method can completely preserve the information of both modalities, making it suitable for subsequent model learning of intermodal relationships. When using a weighted summation method, it can be done according to... To integrate, among which To preset the fusion weights, satisfy The weighted summation method is suitable for scenarios where the clinical latent representation and the imaging latent representation are already in the same semantic subspace, which helps to control the fusion dimension and reduce model complexity.

[0104] In a preferred embodiment, the fused feature data is obtained by stitching together the data to preserve as much complementary information as possible between clinical and imaging features. For example, clinical indicators can better reflect the overall metabolic status and medical history, while imaging features can more directly reflect the local status of the vertebral body and paravertebral muscles. After fusion, the two can jointly characterize the overall musculoskeletal status of the subject under analysis. The fused feature data can be further input into a combined auxiliary diagnostic model and a fracture risk prediction model to complete subsequent status determination and risk assessment.

[0105] In a preferred embodiment, the method of inputting the fused feature data into a joint auxiliary diagnostic model and outputting bone metabolism status results and sarcopenia status results, and obtaining sarcopenia-osteoporosis auxiliary diagnostic results based on the bone metabolism status results and the sarcopenia status results includes:

[0106] The fused feature data is input into the bone metabolism status determination branch, and the bone metabolism status result is output as normal, reduced bone mass, or osteoporosis.

[0107] The fused feature data is input into the sarcopenia status determination branch, and the sarcopenia status result is output as either sarcopenia present or sarcopenia absent.

[0108] When the bone metabolism status result is osteoporosis and the sarcopenia status result is sarcopenia, the auxiliary diagnostic result is determined to be sarcopenia-osteoporosis.

[0109] When the bone metabolism status result is osteopenia and the sarcopenia status result is sarcopenia, the auxiliary diagnostic result is determined to be sarcopenia combined with low bone mass.

[0110] In this embodiment, the joint auxiliary diagnostic model outputs both bone metabolism status and sarcopenia status results based on the same fused feature data, avoiding the need for completely independent analysis processes for skeletal and muscle states. The joint auxiliary diagnostic model can employ a dual-branch output structure, sharing the same fused feature data as input but with separate branches for determining bone metabolism status and sarcopenia status. The bone metabolism status branch outputs one of the following results: normal, reduced bone mass, or osteoporosis; the sarcopenia status branch outputs one of the following results: presence of sarcopenia or absence of sarcopenia. This setup allows the model to learn the collaborative features between muscle and bone using shared multimodal feature representations, while also optimizing the output results for different tasks.

[0111] In one embodiment, the bone metabolism status determination branch can employ a multi-class output structure, where the output can be category labels or probability distributions corresponding to each category. For example, a Softmax layer can output probability values ​​for three categories: normal, osteopenia, and osteoporosis, with the category with the highest probability being taken as the bone metabolism status result. The sarcopenia status determination branch can employ a binary output structure, for example, outputting the probability of sarcopenia through a Sigmoid layer, and then determining the final sarcopenia status result based on a preset determination threshold. The determination threshold can be preset or optimized based on training set validation results.

[0112] In a preferred embodiment, the fused feature data is input into a fracture risk prediction model, which outputs the probability of a fracture event occurring within a preset time period. The method for obtaining the fracture risk level based on the probability of the fracture is as follows:

[0113] The fused feature data is then normalized and subjected to consistency processing.

[0114] The fused feature data after normalization and consistency processing is input into the fracture risk prediction model to obtain the risk probability of the subject to be analyzed occurring a fracture event within the preset time period.

[0115] The risk probability is compared with a preset risk threshold.

[0116] When the risk probability is less than a first preset threshold, the fracture risk level of the object to be analyzed is determined to be low risk;

[0117] When the risk probability is greater than or equal to the first preset threshold and less than the second preset threshold, the fracture risk level of the object to be analyzed is determined to be medium risk.

[0118] When the risk probability is greater than or equal to the second preset threshold, the fracture risk level of the object to be analyzed is determined to be high risk;

[0119] Output the risk probability and the fracture risk level.

[0120] In this embodiment, the fracture risk prediction model is used to predict the probability of a fracture event occurring in the subject of analysis within a preset time period based on fused feature data. Unlike the joint auxiliary diagnostic model, which mainly focuses on current state identification, the fracture risk prediction model emphasizes the assessment of the probability of future events. Therefore, its output is not a simple disease state category, but a continuous risk probability and a risk level mapped from the risk probability. The preset time period can be set according to actual clinical management needs, such as 6 months, 12 months, or 24 months.

[0121] In this embodiment, the fused feature data is normalized and consistent according to the feature processing rules corresponding to the training phase of the fracture risk prediction model. This means that the same feature processing method is used in the prediction phase as in the training phase to ensure the consistency of the input distribution. For example, if min-max normalization or standardization is used for each dimension of the features in the training phase, the same normalization parameters should be used in the prediction phase; if a specific encoding method is used for certain missing fields in the training phase, the same rules should also be used in the prediction phase. This processing can reduce the impact of data distribution offset between the training and application phases on the prediction results. The fracture risk prediction model can use logistic regression, support vector machine, gradient boosting tree, random forest, deep neural network, or other supervised learning models. Regardless of the model used, the input is the consistent fused feature data, and the output is the risk probability of the analyzed object experiencing a fracture event within a preset time period, denoted as . This risk probability is typically a continuous value between 0 and 1, with a higher value indicating a higher likelihood of a fracture.

[0122] In a preferred embodiment, the method for constructing the joint assisted diagnostic model includes:

[0123] Obtain a training sample set containing clinical structured data, quantitative CT image data, and diagnostic labels;

[0124] Preprocessing is performed on the clinical structured data in the training sample set to obtain clinical feature data for training.

[0125] Feature analysis is performed on the quantitative CT image data in the training sample set to obtain training image feature data;

[0126] Perform a feature fusion operation on the training clinical feature data and the training image feature data to obtain training fused feature data;

[0127] The joint auxiliary diagnostic model is trained based on the fusion feature data used for training and the diagnostic labels, so that the joint auxiliary diagnostic model outputs bone metabolism status results and sarcopenia status results.

[0128] In this embodiment, the joint auxiliary diagnostic model and the fracture risk prediction model correspond to different tasks, and therefore are constructed and trained separately. The training method for the fracture risk prediction model can be similar to that for the joint auxiliary diagnostic model, but the joint auxiliary diagnostic model uses diagnostic labels, including bone metabolism status labels and sarcopenia status labels; the fracture risk prediction model uses fracture event labels, indicating whether a fracture event has occurred within a preset time period.

[0129] In the above embodiments, the construction of the joint auxiliary diagnostic model involves learning the correspondence between fused feature data and diagnostic labels, wherein the diagnostic labels include at least bone metabolism status labels and sarcopenia status labels. Bone metabolism status labels can originate from clinical diagnostic conclusions, bone mineral density grading results, or expert annotation results; sarcopenia status labels can originate from clinical judgment criteria, muscle strength or muscle mass assessment results, and expert review conclusions. By collecting a training sample set containing clinical structured data, quantitative CT image data, and the aforementioned diagnostic labels, the training data foundation for the joint auxiliary diagnostic model can be established. Subsequently, the clinical structured data and quantitative CT image data in the training sample set undergo preprocessing, feature analysis, and feature fusion operations consistent with the aforementioned application stage to obtain fused feature data for training. Based on this fused feature data and diagnostic labels, the joint auxiliary diagnostic model is trained, enabling the model to learn the mapping relationship between multimodal features and bone metabolism status and sarcopenia status. The fracture risk prediction model can adopt a binary classification probability output structure, outputting risk probabilities through a Sigmoid layer or other probability mapping methods; alternatively, it can adopt a survival analysis-based modeling approach, transforming this to obtain the probability of fracture events occurring within a preset time period. Since the joint auxiliary diagnostic model and the fracture risk prediction model have different task objectives, although they can have similar preprocessing and feature fusion processes, their training sample labels, loss functions and evaluation indicators are preferably set separately to improve the relevance of model construction and the reliability of output results.

[0130] like Figure 2 As shown, in one embodiment, an artificial intelligence-based device for diagnosing sarcopenia and predicting fracture risk is provided, which may specifically include:

[0131] The raw data acquisition unit 510 is used to acquire the clinical structured data and quantitative CT image data of the object to be analyzed;

[0132] The feature extraction unit 520 is used to preprocess the clinical structured data to obtain clinical feature data; and to perform feature analysis on the quantitative CT image data to obtain image feature data.

[0133] The feature fusion unit 530 is used to perform feature fusion on the clinical feature data and the image feature data to obtain fused feature data;

[0134] The diagnostic result acquisition unit 540 is used to input the fused feature data into the joint auxiliary diagnostic model, output the bone metabolism status result and the sarcopenia status result, and obtain the sarcopenia-osteoporosis auxiliary diagnostic result based on the bone metabolism status result and the sarcopenia status result.

[0135] The risk level prediction unit 550 is used to input the fused feature data into the fracture risk prediction model, output the risk probability of a fracture event occurring within a preset time period, and obtain the fracture risk level based on the risk probability.

[0136] In the embodiments of this application, the explanation and description of the above-mentioned artificial intelligence-based sarcopenia-osteoporosis diagnosis and fracture risk prediction device can be referred to the explanation and description of the corresponding method above. For the description of the artificial intelligence-based sarcopenia-osteoporosis diagnosis and fracture risk prediction method, please refer to the above text, which will not be repeated here.

[0137] Figure 3 An internal structural diagram of a computer device in one embodiment is shown. Figure 3 As shown, the computer device includes a processor, memory, network interface, input device, and display screen connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and may also store computer programs. When executed by the processor, these computer programs enable the processor to implement an artificial intelligence-based method for diagnosing sarcopenia and osteoporosis and predicting fracture risk. The internal memory may also store computer programs, which, when executed by the processor, enable the processor to implement the artificial intelligence-based method for diagnosing sarcopenia and osteoporosis and predicting fracture risk.

[0138] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In one embodiment, the AI-based sarcopenia-osteoporosis diagnosis and fracture risk prediction device provided in this application can be implemented as a computer program, which can be implemented as follows: Figure 3 The device operates on the device shown. The device's memory can store the various program modules that make up this AI-based sarcopenia-osteoporosis diagnosis and fracture risk prediction device, for example, Figure 2 The raw data acquisition unit 510 shown is an example. The computer program, composed of various program modules, causes the processor to execute the steps in the artificial intelligence-based method for diagnosing sarcopenia and osteoporosis and predicting fracture risk, as described in the various embodiments of this application.

[0140] For example, Figure 3 The computer device shown can be used as follows Figure 2 The raw data acquisition unit 510 in the AI-based sarcopenia-osteoporosis diagnosis and fracture risk prediction device shown executes step S10. And so on.

[0141] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, causes the processor to perform the steps of the artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk as described above.

[0142] In the embodiments of this application, please refer to the above description of the artificial intelligence-based method for diagnosing sarcopenia-osteoporosis and predicting fracture risk, which will not be repeated here.

[0143] In one embodiment, an artificial intelligence-based sarcopenia-osteoporosis diagnosis and fracture risk prediction system is provided. The system includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the artificial intelligence-based sarcopenia-osteoporosis diagnosis and fracture risk prediction method as described above.

[0144] In this embodiment, the system can be a computer hardware system that executes its corresponding method when running. For a description of the artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk, please refer to the above text; it will not be repeated here.

[0145] It should be understood that although the steps in the flowcharts of the various embodiments of this application are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0146] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for diagnosing sarcopenia-osteoporosis and predicting fracture risk based on artificial intelligence, characterized in that, The method includes: Acquire the clinical structured data and quantitative CT image data of the subjects to be analyzed; The clinical structured data is preprocessed to obtain clinical feature data; the quantitative CT image data is subjected to feature analysis to obtain image feature data. The clinical feature data and the imaging feature data are fused to obtain fused feature data; The fused feature data is input into the joint auxiliary diagnostic model, which outputs bone metabolism status results and sarcopenia status results. Based on the bone metabolism status results and the sarcopenia status results, an auxiliary diagnostic result for sarcopenia-osteoporosis is obtained. The fused feature data is input into the fracture risk prediction model, which outputs the probability of a fracture event occurring within a preset time period. The fracture risk level is obtained based on the probability of the fracture.

2. The method for diagnosing sarcopenia-osteoporosis and predicting fracture risk based on artificial intelligence according to claim 1, characterized in that, Methods for preprocessing the clinical structured data to obtain clinical feature data include: The original case information is structured and organized according to a preset field dictionary; Continuous indicators are retained as numerical variables, binary symptoms and signs are converted into 0 / 1 coded variables, and graded conclusions are converted into ordinal or unique heat coded variables; text-based examination conclusions are converted into predefined category codes. After the data has been encoded, data cleaning, outlier identification, missing value handling, and numerical normalization are performed to obtain clinical feature data.

3. The method for diagnosing sarcopenia-osteoporosis and predicting fracture risk based on artificial intelligence according to claim 1, characterized in that, Methods for performing feature analysis on the quantitative CT image data to obtain image feature data include: The quantitative CT image data is converted into a model input tensor. The target layer is located by performing target layer localization on the input tensor of the model to obtain the anatomical layer to be analyzed; The vertebral body region and paravertebral muscle region in the anatomical plane are located and segmented based on an artificial intelligence segmentation model. Based on the localization and segmentation results, image quantitative features are extracted. The image quantitative features include at least one or more of the following: bone density features of the vertebral body region, cross-sectional area of ​​the paraspinal muscles, fat infiltration rate, and muscle grayscale statistical features. The image feature data is generated based on the image quantization features.

4. The method for diagnosing sarcopenia and predicting fracture risk based on artificial intelligence according to claim 1, characterized in that, The method for fusing the clinical feature data and the imaging feature data to obtain fused feature data includes: The clinical feature data is encoded into a clinical feature vector; The image feature data is encoded into an image feature vector; The clinical feature vector and the image feature vector are respectively mapped to the feature space to obtain a unified dimension of clinical potential representation and image potential representation; Dimensional alignment is performed between the clinical latent representation and the image latent representation; The dimensionally aligned clinical latent representation and the image latent representation are concatenated or weighted and summed to obtain the fused feature data.

5. The method for diagnosing sarcopenia-osteoporosis and predicting fracture risk based on artificial intelligence according to claim 1, characterized in that, The method for inputting the fused feature data into a joint auxiliary diagnostic model, outputting bone metabolism status results and sarcopenia status results, and obtaining sarcopenia-osteoporosis auxiliary diagnostic results based on the bone metabolism status results and the sarcopenia status results includes: The fused feature data is input into the bone metabolism status determination branch, and the bone metabolism status result is output as normal, reduced bone mass, or osteoporosis. The fused feature data is input into the sarcopenia status determination branch, and the sarcopenia status result is output as either sarcopenia present or sarcopenia absent. When the bone metabolism status result is osteoporosis and the sarcopenia status result is sarcopenia, the auxiliary diagnostic result is determined to be sarcopenia-osteoporosis. When the bone metabolism status result is osteopenia and the sarcopenia status result is sarcopenia, the auxiliary diagnostic result is determined to be sarcopenia combined with low bone mass.

6. The method for diagnosing sarcopenia-osteoporosis and predicting fracture risk based on artificial intelligence according to claim 1, characterized in that, The method for inputting the fused feature data into a fracture risk prediction model and outputting the risk probability of a fracture event occurring within a preset time period, and then obtaining the fracture risk level based on the risk probability, is as follows: The fused feature data is then normalized and subjected to consistency processing. The fused feature data after normalization and consistency processing is input into the fracture risk prediction model to obtain the risk probability of the subject to be analyzed occurring a fracture event within the preset time period. The risk probability is compared with a preset risk threshold. When the risk probability is less than a first preset threshold, the fracture risk level of the object to be analyzed is determined to be low risk; When the risk probability is greater than or equal to the first preset threshold and less than the second preset threshold, the fracture risk level of the object to be analyzed is determined to be medium risk. When the risk probability is greater than or equal to the second preset threshold, the fracture risk level of the object to be analyzed is determined to be high risk; Output the risk probability and the fracture risk level.

7. The method for diagnosing sarcopenia-osteoporosis and predicting fracture risk based on artificial intelligence according to claim 1, characterized in that, The method for constructing the joint assisted diagnostic model includes: Obtain a training sample set containing clinical structured data, quantitative CT image data, and diagnostic labels; Preprocessing is performed on the clinical structured data in the training sample set to obtain clinical feature data for training. Feature analysis is performed on the quantitative CT image data in the training sample set to obtain training image feature data; Perform a feature fusion operation on the training clinical feature data and the training image feature data to obtain training fused feature data; The joint auxiliary diagnostic model is trained based on the fusion feature data used for training and the diagnostic labels, so that the joint auxiliary diagnostic model outputs bone metabolism status results and sarcopenia status results.

8. An artificial intelligence-based device for diagnosing sarcopenia and osteoporosis and predicting fracture risk, characterized in that, The device includes: The raw data acquisition unit is used to acquire the clinical structured data and quantitative CT image data of the object to be analyzed; The feature extraction unit is used to preprocess the clinical structured data to obtain clinical feature data; and to perform feature analysis on the quantitative CT image data to obtain image feature data. A feature fusion unit is used to perform feature fusion on the clinical feature data and the image feature data to obtain fused feature data; The diagnostic result acquisition unit is used to input the fused feature data into the joint auxiliary diagnostic model, output the bone metabolism status result and sarcopenia status result, and obtain the sarcopenia-osteoporosis auxiliary diagnostic result based on the bone metabolism status result and sarcopenia status result. The risk level prediction unit is used to input the fused feature data into the fracture risk prediction model, output the risk probability of a fracture event occurring within a preset time period, and obtain the fracture risk level based on the risk probability.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the steps of the artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk as described in any one of claims 1 to 7.

10. An artificial intelligence-based system for diagnosing sarcopenia and osteoporosis and predicting fracture risk, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the artificial intelligence-based method for diagnosing sarcopenia and predicting fracture risk as described in any one of claims 1 to 7.

Images