Osteoporosis diagnosis method based on image recognition

Abstract

The present invention relates to the technical field of osteoporosis diagnosis, and discloses an osteoporosis diagnosis method based on image recognition. The method uses medical imaging equipment to obtain bone density image data of a target object and bone structure image data of a target area, and determines the distribution of trabeculae in the target area based on the bone structure image data; then, a three-dimensional bone density model of the target object is constructed based on the bone density image data, and an image analysis strategy is determined in combination with the three-dimensional bone density model and the distribution of trabeculae; an image analysis module obtains bone structure feature data of the target area based on the strategy, and identifies the bone density grade and degree of osteoporosis in the target area based on the feature data to obtain diagnostic data. By integrating bone density and bone structure-related data and combining image recognition technology for comprehensive analysis, the method provides a new, efficient and reliable method for clinical diagnosis, which helps to improve the accuracy and objectivity of osteoporosis diagnosis and reduce the occurrence of misdiagnosis and missed diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of osteoporosis diagnosis, and in particular to an osteoporosis diagnosis method based on image recognition. Background Art

[0002] Osteoporosis is a common bone disease characterized by decreased bone mass and disrupted bone microarchitecture, leading to increased bone fragility and susceptibility to fracture. With the aging population, the incidence of osteoporosis is increasing annually, severely impacting patients' quality of life and placing a significant burden on the healthcare burden. Therefore, early and accurate diagnosis of osteoporosis is of great practical significance. Currently, commonly used methods for diagnosing osteoporosis in clinical practice include bone density measurement and bone structure assessment. Dual-energy X-ray absorptiometry (DEXA) is the gold standard for bone density measurement, which measures bone density at specific locations to determine the presence of osteoporosis. However, this method has limitations. It only reflects overall bone density and cannot accurately assess changes in bone microarchitecture, a key pathological basis for the development and progression of osteoporosis. Furthermore, DEXA requires high equipment requirements and is relatively expensive, limiting its widespread use in primary healthcare settings. In addition to bone density measurement, methods for assessing bone structure are also evolving. Trabecular bone structure analysis is a key indicator of bone quality. Traditional analysis of trabecular bone structure relies primarily on tissue sectioning and microscopic observation. This method is invasive, painful, and incapable of in vivo testing, limiting its application in clinical diagnosis. In recent years, with the advancement of medical imaging technology, techniques such as computed tomography (CT) and magnetic resonance imaging (MRI) have been used for non-invasive assessment of trabecular bone structure. However, these methods still have limitations in terms of image resolution, analysis efficiency, and accuracy. While CT offers high spatial resolution, it incurs a high radiation dose, making it unsuitable for frequent testing. MRI, while capable of resolving soft tissue, lacks optimal visualization of trabecular bone structure, requires long imaging times, and is susceptible to motion artifacts. Existing osteoporosis diagnostic methods also present challenges in image analysis and data processing. Traditional image analysis relies primarily on manual interpretation, resulting in diagnostic results that are significantly influenced by physician experience and subjective judgment, leading to low inter-doctor consistency. Furthermore, manual interpretation is inefficient and cannot meet the demands of large-scale screening and rapid diagnosis. Although some automated analysis algorithms have been introduced into the diagnosis of osteoporosis, these algorithms often only analyze a single bone density or bone structure feature, ignoring the relationship between bone density and bone structure, resulting in the accuracy and reliability of the diagnostic results needing to be improved. Existing diagnostic methods also have defects in data integration and analysis. Bone density data and bone structure data are usually processed separately, and it is impossible to achieve collaborative analysis of the two, which affects the comprehensiveness and accuracy of the diagnosis. Moreover, the existing analysis algorithms are not comprehensive enough in extracting bone structure features, making it difficult to accurately reflect the degree of osteoporosis, resulting in frequent misdiagnosis and missed diagnosis. Existing osteoporosis diagnostic methods have many problems in bone microstructure assessment, equipment cost, analysis efficiency, data integration, etc., and it is difficult to meet the clinical needs for early and accurate diagnosis of osteoporosis.Therefore, developing a new, efficient and accurate method for diagnosing osteoporosis has important clinical value. Summary of the Invention

[0003] The purpose of the present invention is to provide an osteoporosis diagnosis method based on image recognition to solve the problems raised in the above background technology.

[0004] To achieve the above objectives, the present invention provides an osteoporosis diagnosis method based on image recognition, the method comprising:

[0005] Acquiring bone density image data of a target object and bone structure image data of a target area based on medical imaging equipment, and determining the distribution of trabecular bone in the target area according to the bone structure image data;

[0006] constructing a three-dimensional bone density model of the target object based on the bone density image data, and determining an image analysis strategy based on the three-dimensional bone density model and trabecular distribution;

[0007] A preset image analysis module is loaded, and the image analysis module is used to obtain bone structure feature data of the target area according to the image analysis strategy, identify the bone density grade and osteoporosis degree of the target area according to the feature data, and obtain diagnostic data.

[0008] Preferably, the bone density image data of the target object and the bone structure image data of the target area are obtained based on a medical imaging device, and the trabecular distribution of the target area is determined according to the bone structure image data, specifically:

[0009] Acquiring bone density image data of a target object and bone structure image data of a target area in the target object based on a medical imaging device;

[0010] performing a noise suppression operation on the bone structure image data based on an adaptive median filter algorithm to obtain denoised bone structure image data, introducing a pyramid decomposition algorithm, and setting an initial number of layers of the pyramid decomposition algorithm parameters;

[0011] Performing multi-scale decomposition on the denoised bone structure image data according to the pyramid decomposition algorithm to obtain m sub-images with different resolutions;

[0012] Performing grayscale statistical analysis on the sub-images, constructing a grayscale histogram of each sub-image, calculating an average grayscale value, grayscale variance, and entropy value of each sub-image based on the grayscale histogram, and constructing a feature statistical matrix using the average grayscale value, grayscale variance, and entropy value;

[0013] Calculating the contrast, energy, and correlation of each pixel point collected by the medical imaging device according to the characteristic statistical matrix to obtain a trabecular bone characteristic spectrum;

[0014] Acquiring sampling point position information of the trabecular bone characteristic spectrum according to the bone structure image data, mapping the trabecular bone characteristic spectrum of each sampling point according to the sampling point position information, and constructing a three-dimensional trabecular bone characteristic spectrum;

[0015] The trabecular gap at each sampling point is determined based on the three-dimensional trabecular characteristic spectrum, a trabecular distribution heat map and a trabecular gap gradient field of the target area are constructed based on the trabecular gap at each sampling point, and the trabecular distribution of the target area is determined based on the trabecular distribution heat map and the trabecular gap gradient field.

[0016] Preferably, the three-dimensional bone density model of the target object is constructed based on the bone density image data, and the image analysis strategy is determined based on the three-dimensional bone density model and the distribution of trabeculae, specifically:

[0017] Performing multi-dimensional morphological analysis on the bone density image data, constructing a bone density feature tensor through block processing, and identifying bone tissue types based on the grayscale signals of the bone density image data using a random forest classifier to generate a three-dimensional bone density model of the target object that includes bone density values, morphological features, and tissue type labels;

[0018] Performing a position mapping operation on the trabecular distribution and the three-dimensional bone density model to construct a three-dimensional spatial distribution model of bone density resources, and dividing the three-dimensional spatial distribution model of bone density resources according to a preset block size to construct M subspace regions;

[0019] Obtaining trabecular gap information of each subspace region according to the trabecular distribution, and clustering subspace regions with similar trabecular gaps in the three-dimensional spatial distribution model of bone density resources based on a hierarchical clustering algorithm to obtain a clustering result;

[0020] Determining the distribution probability of trabeculae in each subspace region based on the clustering result, obtaining corresponding position information of each subspace region in the target object, and performing a bone density detection importance assessment on each position in the target region in the target object based on the distribution probability and the corresponding position information to obtain a detection importance score for each position in the target region;

[0021] Determining analysis requirement information for each location in the target area according to the detection importance score of each location, wherein the analysis requirement information includes whether to perform analysis and analysis accuracy requirement information;

[0022] Obtaining feature extraction clarity data of bone structures at different distances by the image analysis module, determining feature collection distance information for each position of the target area by the image analysis module based on the feature extraction clarity data and analysis requirement information, and determining a feature collection point set for the image analysis module based on the feature collection distance information;

[0023] An image analysis strategy of an image analysis module is determined according to the feature acquisition point set.

[0024] Preferably, the image analysis strategy of the image analysis module determined according to the feature acquisition point set is specifically:

[0025] Obtaining three-dimensional coordinate information of each feature acquisition point and initial position information of the image analysis module, and determining structural obstacles in the target object based on the three-dimensional bone density model;

[0026] Optimizing the path of the initial position information and the three-dimensional coordinate information of each feature acquisition point based on the Dijkstra algorithm, taking the structural obstacle as a path optimization restriction area, and outputting the shortest initial analysis path of the image analysis module;

[0027] acquiring in real time posture change data of the image analysis module during analysis by the image analysis module according to the shortest initial analysis path, and determining the motion noise direction and motion noise intensity of the real-time analysis position of the image analysis module according to the posture change data;

[0028] Acquiring operational stability data of the image analysis module, wherein the operational stability data includes data on the image analysis module's resistance to motion noises of different directions and intensities;

[0029] Acquiring recognition response time data of the image analysis module to the direction and intensity of the motion noise, and determining an analysis path compensation hysteresis amount of the image analysis module according to the recognition response time data;

[0030] Analyzing the direction and intensity of the motion noise at the real-time analysis position of the image analysis module according to the resistance data and the analysis path compensation hysteresis, and determining the cumulative amount of the analysis path deviation within the recognition response time of the image analysis module;

[0031] If the cumulative amount of the analysis path offset is less than a preset value, determining an analysis offset direction and an analysis offset distance for analysis by the image analysis module according to the shortest initial analysis path based on the cumulative amount of the analysis path offset, and determining a compensation direction and a compensation distance of the image analysis module based on the analysis offset direction and the analysis offset distance to obtain compensation data;

[0032] Performing a compensation operation on the shortest initial analysis path of the image analysis module during real-time operation according to the compensation data;

[0033] If the cumulative amount of the analysis path deviation is greater than a preset value, obtaining motion noise intensity and motion noise direction data of the real-time running path of the image analysis module, constructing a motion noise change map based on the motion noise intensity and motion noise direction data of the real-time running path, and determining a motion noise change trend of a target area in the target object based on the motion noise change map;

[0034] An interpolation operation is performed on the motion noise change trend based on bilinear interpolation to determine the motion noise information within the preset range of the shortest initial analysis path. The operational stability of the image analysis module within the preset range of the shortest initial analysis path is determined based on the motion noise information and resistance data. Based on the operational stability, the shortest initial analysis path section with a cumulative analysis path offset greater than a preset value is adjusted to obtain an updated analysis path. The image analysis module performs an analysis operation based on the updated analysis path.

[0035] Preferably, the image analysis module obtains bone structure feature data of the target area according to the image analysis strategy, identifies the bone density grade and osteoporosis degree of the target area according to the feature data, and obtains diagnostic data, specifically:

[0036] Obtaining standard feature data of different bone density levels of the target object, and labeling the standard feature data with the bone density level to obtain labeled feature data;

[0037] Building a bone density recognition model based on a residual network, and importing the labeled feature data into the bone density recognition model for training;

[0038] The image analysis module obtains bone structure feature data of the target area according to the image analysis strategy, imports the bone structure feature data into the trained bone density recognition model to identify the bone density level, and evaluates the degree of osteoporosis of each bone density level to obtain diagnostic data.

[0039] Preferably, the transmitting of the diagnostic data to a terminal display device according to digital communication technology is specifically as follows:

[0040] Building a layered data communication protocol stack based on digital communication technology, encoding the diagnostic data, and building a diagnostic data digital transmission signal;

[0041] The diagnostic data digital transmission signal is digitally transmitted to a terminal display device according to the layered data communication protocol stack, and a decoding operation is performed on the diagnostic data digital transmission signal to obtain osteoporosis diagnostic data of the target area.

[0042] Preferably, the bone density recognition model is constructed based on the residual network, and the labeled feature data is imported into the bone density recognition model for training, specifically:

[0043] Determine the number of basic convolutional layers and skip connection layers of the residual network, and set the initial learning rate and batch size parameters;

[0044] The labeled feature data is divided into a training set and a validation set, and the error between the model prediction value and the true value is calculated using a cross entropy loss function;

[0045] The model parameters are iteratively updated through the stochastic gradient descent optimizer. When the accuracy of the validation set reaches the preset threshold, the training is stopped to obtain the trained bone density recognition model.

[0046] Preferably, the layered data communication protocol stack is constructed based on digital communication technology, the diagnostic data is encoded, and a diagnostic data digital transmission signal is constructed, specifically:

[0047] Define the modulation mode and transmission rate of the physical layer, set the frame header format and verification rules of the data link layer, and determine the routing strategy and address allocation of the network layer;

[0048] The diagnostic data is converted into binary form, and a start flag, a length field and a check code are added according to the frame format of the data link layer to generate a diagnostic data digital transmission signal.

[0049] Preferably, the path optimization is performed on the initial position information and the three-dimensional coordinate information of each feature acquisition point based on the Dijkstra algorithm, and the structural obstacle is used as the path optimization restriction area to output the shortest initial analysis path of the image analysis module, specifically:

[0050] Construct a three-dimensional coordinate grid map of the target area and map the three-dimensional coordinates of each feature collection point into a grid node;

[0051] Initialize the distance value of each node to infinity, set the distance value of the initial position node to zero, and use the priority queue to store the nodes to be processed;

[0052] Take the node with the smallest distance value in turn, update the distance value of its adjacent node, and skip it if the adjacent node is a structural obstacle;

[0053] When the distance values ​​of all feature acquisition point nodes are updated, the shortest initial analysis path is obtained by backtracking.

[0054] Preferably, the noise suppression operation is performed on the bone structure image data based on the adaptive median filtering algorithm to obtain denoised bone structure image data, specifically:

[0055] Set the minimum and maximum sizes of the filter window and calculate the median, maximum and minimum values ​​of the pixels in the current window;

[0056] If the difference between the maximum and minimum values ​​in the current window is less than the noise threshold, the median is retained as the filtered pixel value; if the difference is greater than or equal to the noise threshold, the filter window size is expanded and the calculation is repeated until the window reaches the maximum size or meets the noise suppression condition, thereby obtaining the denoised bone structure image data.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention's image recognition-based osteoporosis diagnostic method uses medical imaging equipment to obtain bone density image data of a target subject and bone structure image data of a target area. The method then combines these two data to determine an image analysis strategy, thereby identifying the bone density level and osteoporosis severity of the target area. This method effectively overcomes many drawbacks of existing diagnostic methods. In existing methods, bone density measurements often only reflect the overall density of the bone and fail to fully reflect changes in bone microstructure. Furthermore, bone structure assessments often struggle to effectively integrate bone density data due to technical limitations. This method incorporates both bone density image data and bone structure image data into the analysis system. Trabecular bone distribution is first determined based on the bone structure image data, and then the analysis strategy is determined based on a three-dimensional bone density model, resulting in a more comprehensive assessment of bone condition. Trabecular bone, as an important component of bone microstructure, is closely related to the development and progression of osteoporosis. Combining this with bone density data can reflect the health of the bone from multiple dimensions, avoiding the one-sided diagnosis caused by a single data source. During the analysis process, the image analysis module acquires bone structure feature data based on the established strategy, thereby identifying the bone density level and osteoporosis severity. This feature-based identification method can more accurately capture subtle changes associated with osteoporosis. Traditional manual image interpretation is subject to significant subjective factors, and judgments may vary between doctors. However, this method, through a standardized image analysis process, reduces human interference, resulting in more objective and consistent diagnostic results. Furthermore, the construction of a three-dimensional bone density model intuitively presents the three-dimensional structure of the bone, providing richer spatial information for developing analytical strategies, facilitating more accurate localization of target areas and extraction of feature data. This method eliminates the need for costly specialized equipment and can acquire the required data using existing medical imaging equipment, lowering the diagnostic hardware barrier and facilitating its widespread adoption in a wider range of medical institutions. For patients, this non-invasive diagnostic method avoids the pain associated with invasive tests, improving patient acceptance. In terms of data processing, bone density and bone structure data are analyzed collaboratively, eliminating the need for separate processing and subsequent manual integration. This simplifies the diagnostic process, improves efficiency, and provides clinicians with diagnostic data more quickly, enabling timely treatment for patients. In terms of diagnostic accuracy, the comprehensive consideration of multiple features of bone density and bone structure allows for earlier detection of signs of osteoporosis. Early diagnosis is crucial for the treatment of osteoporosis, allowing patients to gain more time for treatment, slowing disease progression, and reducing the risk of complications such as fractures. Furthermore, accurate identification of bone density grade and osteoporosis severity provides valuable insights for doctors to develop personalized treatment plans, making treatment more targeted. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a diagram showing the working principle of the osteoporosis diagnosis method based on image recognition according to the present invention;

[0060] Figure 2 Flowchart for identification of bone density grade and degree of osteoporosis;

[0061] Figure 3 Flowchart for transmitting diagnostic data to terminal display device;

[0062] Figure 4 Flowchart for training a bone density recognition model. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] See also Figures 1-4 The present invention provides an osteoporosis diagnosis method based on image recognition, and the specific implementation steps are as follows:

[0065] Acquiring bone density image data of a target object and bone structure image data of a target area based on medical imaging equipment, and determining the distribution of trabecular bone in the target area according to the bone structure image data;

[0066] constructing a three-dimensional bone density model of the target object based on the bone density image data, and determining an image analysis strategy based on the three-dimensional bone density model and trabecular distribution;

[0067] A preset image analysis module is loaded, and the image analysis module is used to obtain bone structure feature data of the target area according to the image analysis strategy, identify the bone density grade and osteoporosis degree of the target area according to the feature data, and obtain diagnostic data.

[0068] Example 1:

[0069] The target subject's bone density imaging data and bone structure imaging data of the target area within the target subject are acquired using medical imaging equipment. The medical imaging equipment here can include dual-energy X-ray absorptiometry, high-resolution CT, etc. The target area can be selected based on diagnostic needs, such as the lumbar spine and hip, which are prone to osteoporosis. The acquired bone density imaging data can reflect the density of the target subject's bones, while the bone structure imaging data can provide detailed information about the bone structure in the target area.

[0070] The minimum filter window size was set to 3×3 and the maximum to 7×7. The median, maximum, and minimum values ​​of the pixels within the window were calculated. For each pixel in the bone structure imaging data, a filter window with an initial size of 3×3 was constructed with that pixel as the center. All pixel values ​​within the window were traversed, and the maximum, minimum, and median values ​​were determined.

[0071] If the difference between the maximum and minimum values ​​in the current window is less than the noise threshold, the median is retained as the filtered pixel value. If the difference is greater than or equal to the noise threshold, the filter window size is expanded and the calculation is repeated until the window reaches the maximum size or the noise suppression condition is met, thus obtaining the denoised bone structure image data. Assuming the noise threshold is set to 15, when the difference between the maximum and minimum values ​​of the pixels in the 3×3 window is 20, which is greater than 15, the window size is expanded to 5×5, and the maximum, minimum, and median values ​​of the pixels in the window are recalculated. If the difference is now 12, which is less than 15, the median value is used as the filtered value of the pixel point. If the difference in the 5×5 window is still greater than or equal to 15, the window size is expanded to 7×7 and the calculation method is repeated until the pixel value that meets the noise suppression condition is obtained, completing the noise suppression operation for the entire bone structure image data.

[0072] A pyramid decomposition algorithm was introduced, and the initial number of layers of the pyramid decomposition algorithm parameters was set to 4. The pyramid decomposition algorithm can decompose images into sub-images of different resolutions. The initial number of layers should be determined based on the resolution and detail of the bone structure imaging data. The initial setting of 4 layers ensures the decomposition effect while avoiding information redundancy caused by excessive decomposition.

[0073] The denoised bone structure image data is decomposed at multiple scales using a pyramid decomposition algorithm, yielding four sub-images of varying resolutions. During this decomposition process, the denoised bone structure image data is subjected to Gaussian filtering and downsampling, yielding sub-images of progressively lower resolutions. The first sub-image is the result of a single filtering and downsampling operation on the original denoised image, the second sub-image is the result of the same operation on the first sub-image, and so on, until four sub-images of varying resolutions are obtained.

[0074] Grayscale statistical analysis is performed on each sub-image to construct a grayscale histogram for each sub-image. The grayscale histogram reflects the distribution of pixels of different grayscale values ​​within the sub-image. By counting the frequency of occurrence of each grayscale value in the sub-image, a corresponding grayscale histogram is formed. Based on the grayscale histogram, the average grayscale value, grayscale variance, and entropy value of each sub-image are calculated. The average grayscale value, grayscale variance, and entropy value are used to construct a feature statistical matrix. The average grayscale value is obtained by dividing the sum of the products of the grayscale value and the corresponding frequency by the total number of pixels. The grayscale variance is calculated by calculating the square of the difference between each grayscale value and the average grayscale value and multiplying it by the sum of the frequencies. The entropy value is calculated by calculating the information entropy based on the frequency of the grayscale values. These three values ​​are arranged in the order of the sub-images to form a feature statistical matrix.

[0075] The trabecular bone characteristic spectrum is obtained by calculating the contrast, energy, and correlation of each pixel captured by the medical imaging device based on the characteristic statistical matrix. Using the data in the characteristic statistical matrix and the calculation method of the gray-level co-occurrence matrix, the contrast (reflecting the local variation of the pixel grayscale), energy (reflecting the uniformity of the image grayscale distribution), and correlation (reflecting the degree of correlation between adjacent pixel grayscales) of each pixel are calculated. These parameters together constitute the trabecular bone characteristic spectrum, which reflects the characteristic information of trabecular bone.

[0076] The sampling point locations of the trabecular bone spectrum are obtained from the bone structure imaging data. Uniform sampling is used to select sampling points at regular intervals across the bone structure imaging data, and the coordinates of each sampling point are recorded. The trabecular bone spectrum of each sampling point is mapped based on the sampling point location information to construct a three-dimensional trabecular bone spectrum. The trabecular bone spectrum of each sampling point is mapped to its position in three-dimensional space, and the areas between the sampling points are filled in using methods such as interpolation to form a three-dimensional trabecular bone spectrum. This more comprehensively reflects the characteristic distribution of trabecular bone in three-dimensional space.

[0077] The trabecular interstices at each sampling point are determined based on the three-dimensional trabecular bone characteristic spectrum. The distribution of trabecular features in the three-dimensional trabecular bone characteristic spectrum is analyzed to identify the interstices between trabeculae and calculate the size of the interstices at each sampling point. A trabecular distribution heatmap and a trabecular interstice gradient field are constructed based on the trabecular interstices at each sampling point. The trabecular distribution heatmap uses color to represent the size distribution of trabecular interstices, with darker colors indicating larger interstices. The trabecular interstice gradient field reflects the spatial rate and direction of change of trabecular interstices. The trabecular distribution heatmap and trabecular interstice gradient field determine the trabecular distribution in the target area. By combining the color distribution range in the heatmap and the changing trend of the gradient field, the density and arrangement of trabeculae within the target area can be determined.

[0078] Example 2:

[0079] Dual-energy X-ray absorptiometry was used to collect bone density imaging data from the target subjects, covering the entire skeletal area. Simultaneously, high-resolution CT was used to perform tomography scans of the selected target area, acquiring bone structural imaging data for that region. The target area was identified as the femoral neck and intertrochanteric region of the hip. The bone density imaging data displayed differences in bone mineral content distribution at varying grayscale levels, while the bone structural imaging data revealed the microscopic arrangement and spatial connectivity of the trabeculae in that region.

[0080] Bone density imaging data were preprocessed to remove non-bone tissue regions at the image edges. An adaptive threshold segmentation algorithm was used to automatically determine the segmentation threshold by analyzing the bimodal distribution characteristics of the image grayscale histogram. Pixels above the threshold were defined as bone regions, while the remaining pixels were marked as background and removed. During preprocessing, a 3×3 median filter was used to smooth the bone regions, eliminating salt-and-pepper noise generated during image acquisition while preserving detailed information about the bone edges.

[0081] The preprocessed bone density imaging data was divided into multiple subregions. Based on the midline of the femoral neck, the area extending medially to the edge of the medullary cavity was designated the medial region, the area extending laterally to the edge of the cortical bone was designated the lateral region, and the area between the medial and lateral regions was designated the intermediate region. The subregions were divided using an edge detection algorithm to extract the contour of the femoral neck. Anatomical landmarks were then used to define the boundaries of each region, ensuring that the geometry of each subregion matched the skeletal anatomy.

[0082] Calculate the peak bone density and distribution range for each subregion. For the medial region, traverse the grayscale values ​​of all pixels and select the maximum value as the peak bone density. The difference between the minimum and maximum grayscale values ​​is calculated to determine the bone density distribution range. The peak bone density and distribution range are calculated using the same method for the lateral and middle regions, and the results are recorded for each subregion separately.

[0083] Noise reduction was performed on bone structure imaging data using a non-local means filtering algorithm. The similarity window size was set to 7×7, and the search window size was set to 21×21. Similar pixels were determined by calculating the Euclidean distance between pixel neighborhoods. The grayscale values ​​of similar pixels were weighted averaged as the filtering result for the current pixel. The filtering parameters were dynamically adjusted based on the image noise level. When image noise was significant, the similarity weight coefficient was increased to improve the noise reduction effect.

[0084] The de-noised bone structure image data were reconstructed in three dimensions using a surface rendering algorithm. First, contours were extracted from the two-dimensional slice images. The Canny operator was used to detect the edge contours of the trabeculae. The contour lines of adjacent slices were connected using a contour matching algorithm to construct a three-dimensional mesh model of the trabeculae. During the three-dimensional reconstruction process, the voxel spacing was set to 0.3 mm × 0.3 mm × 0.5 mm to ensure that the reconstructed model accurately reflected the spatial distribution characteristics of the trabeculae.

[0085] Trabecular morphological parameters were extracted from the three-dimensional bone structural model, including trabecular connection density, structural model index, and degree of anisotropy. The connection density was calculated by counting the number of trabecular connection points per unit volume, with each connection point requiring at least three trabeculae to intersect. The structural model index was calculated by calculating the surface area to volume ratio of the trabeculae and reflects the rod-like or plate-like structure of the trabeculae. The degree of anisotropy was determined by calculating the variance of the main direction distribution of the trabecular orientation, reflecting the directional characteristics of the trabecular arrangement.

[0086] A correlation model between bone density parameters and trabecular morphology parameters was established. Data from the medial, lateral, and intermediate zones were integrated, along with the peak and distribution range of bone density and the corresponding trabecular connectivity density and structural model index. Multiple linear regression was used to construct correlation equations, using bone density parameters as independent variables and trabecular morphology parameters as dependent variables. The regression coefficients were calculated using the least squares method to obtain correlation model expressions for each subregion.

[0087] Based on the association model, a bone quality assessment atlas for the target region is generated. Using pseudo-color coding, the atlas maps the combined values ​​of bone density parameters and trabecular bone morphology onto a preset color gradient. The gradient from green to yellow to red represents the progression from good to poor bone quality. The atlas also marks the boundaries of each subregion, visually demonstrating the differences in bone quality across regions through the spatial distribution of color.

[0088] The bone quality assessment atlas fuses the grayscale information of bone density images with the texture features of bone structure images. Using a weighted fusion algorithm, the grayscale values ​​of the bone density images and the texture energy values ​​of the bone structure images are superimposed at a weight ratio of 3:1 to generate a fused atlas that combines both density and structure information. During the fusion process, histogram equalization is used to enhance details in low-contrast areas, allowing weak signals in the atlas to be clearly presented.

[0089] Edge enhancement was performed on the fused bone quality assessment atlas. The Sobel operator was used to calculate the color gradient of the atlas, extracting the boundaries of different bone quality areas. The grayscale values ​​of these boundaries were enhanced to create a clear contrast with the surrounding areas. The intensity of edge enhancement was controlled by a gradient threshold. When the gradient value exceeded the threshold, the color values ​​of the boundary pixels were adjusted to ensure clear boundaries and avoid artifacts caused by over-enhancement.

[0090] The processed bone quality assessment atlas is stored in a standard medical imaging format, using the "patient ID_examination date_target area" file format to facilitate data management and retrieval. The stored file contains metadata such as the atlas' pixel size, color map, and associated model parameters. A data dictionary details the meaning of each metadata element, ensuring accurate data exchange between different systems.

[0091] The various parameters involved in the above process, such as filter window size, reconstruction voxel spacing, fusion weight ratio, etc., can be adjusted according to specific clinical needs and imaging characteristics to adapt to the bone quality assessment scenarios of different target objects, providing intuitive and detailed visual information for the comprehensive assessment of bone quality.

[0092] Example 3:

[0093] The three-dimensional coordinates of each feature acquisition point and the initial position of the image analysis module are acquired using the positioning system of the medical imaging device. A three-dimensional rectangular coordinate system is established with an anatomical landmark on the target bone (e.g., the apex of the greater trochanter of the femur) as the origin. The x-axis points distally along the long axis of the bone, the y-axis is perpendicular to the x-axis and points medially, and the z-axis is perpendicular to the x-y plane and points upward. The coordinates (x, y, z) of each feature acquisition point are determined based on this coordinate system. The image analysis module's initial position is set to (50 mm, 30 mm, 20 mm) in the coordinate system. This position is chosen to avoid obstructing structures on the bone surface and ensure that the initial field of view covers the primary feature acquisition area. Structural obstructions within the target object are identified based on a three-dimensional bone density model. These obstructions include dense cortical bone, joint spaces, and metal artifacts in the image. The spatial extent of these obstructions is determined using grayscale threshold segmentation (defined as areas with grayscale values ​​greater than 2000 HU as dense cortical bone) and morphological recognition (joint spaces are characterized as continuous low-grayscale strips).

[0094] A three-dimensional coordinate grid map of the target area is constructed, dividing the three-dimensional space into a cubic grid with a side length of 0.5 mm. The coordinates of each grid node are (i×0.5, j×0.5, k×0.5), where i, j, and k are integers. The three-dimensional coordinates of each feature acquisition point are mapped to a grid node. The Euclidean distance between the feature acquisition point coordinate and the grid node coordinate is calculated, and the grid node with the smallest distance is selected as the mapping node for that point. The distance value of each node is initialized to infinity, and the distance value of the node at the initial position is set to zero. A priority queue is used to store the nodes to be processed. The priority queue is sorted in ascending order of node distance value and initially contains only the node at the initial position. The node with the smallest distance value is sequentially selected and its six adjacent grid nodes (neighboring nodes along the positive and negative directions of the X, Y, and Z axes) are checked. If the adjacent node is a structural obstacle, it is skipped. Otherwise, the distance value of the adjacent node is calculated (the current node distance value plus the grid side length of 0.5 mm). If the distance value is less than the adjacent node's current distance value, the distance value of the adjacent node is updated and added to the priority queue. After the distance values ​​of all feature collection point nodes are updated, starting from the last feature collection point node, the shortest initial analysis path consisting of a series of grid nodes is obtained by backtracking to the predecessor node of each node (that is, the node whose distance value is updated) until the initial position node is reached.

[0095] The image analysis module acquires attitude change data in real time during analysis based on the shortest initial analysis path. Angular velocity and acceleration data are collected via the module's built-in six-axis gyroscope (sampling frequency: 100Hz). After Kalman filtering, the module's real-time attitude angles (pitch angle α, roll angle β, and heading angle γ) are obtained. The attitude change data is used to determine the direction and intensity of motion noise at the image analysis module's real-time analysis location. The direction of motion noise is determined by the sign of the rate of change of the attitude angle. For example, when the rate of change of the pitch angle α is positive, the noise direction is in the direction of increasing pitch angle. The intensity of motion noise is calculated using the root mean square of the attitude angle rate of change using the following formula:

[0096] ;

[0097] in, 、 、 The time intervals (Take 0.1s) the change in pitch angle, roll angle and heading angle, The unit is rad / s.

[0098] Obtain the operational stability data of the image analysis module by applying sinusoidal disturbances of different directions (positive and negative directions of the X, Y, and Z axes) and different intensities (0.1 rad / s to 1 rad / s) to the module in a simulated environment, and record the module's analysis error under each disturbance. The disturbance range with an error less than 0.1 mm is the module's resistance to motion noise of that direction and intensity. Obtain the recognition response time data of the image analysis module to the direction and intensity of motion noise by inputting a step-changing noise signal into the module and recording the time interval from signal input to module output of the recognition result. After multiple measurements, the average value is taken as the recognition response time. (Usually 0.2s). The image analysis module's analysis path compensation lag is determined based on the recognition response time data. The compensation lag is the recognition response time. With module movement speed (set to 5mm / s) is the product of the hysteresis distance .

[0099] The motion noise direction and intensity of the image analysis module's real-time analysis position are analyzed based on the resistance data and the analysis path compensation hysteresis, and the cumulative amount of the image analysis module's analysis path deviation within the recognition response time is determined. The cumulative amount of path deviation is obtained by integrating the change in motion noise intensity within the recognition response time, i.e.

[0100] ;

[0101] in: is the integral symbol; subscript and superscript Indicates the lower and upper limits of the integral (definite integral interval); is the integrand; Represents a variable Integration is performed. If the cumulative analysis path offset is less than a preset value (set to 0.05 rad), the image analysis module determines the analysis offset direction (consistent with the direction of motion noise) and analysis offset distance (the product of the cumulative offset and the distance from the module to the target) based on the shortest initial analysis path. Based on the analysis offset direction and analysis offset distance, the image analysis module determines the compensation direction (opposite to the offset direction) and compensation distance (equal to the offset distance), generating compensation data. Based on the compensation data, the image analysis module compensates for the shortest initial analysis path during real-time operation. For example, when the module moves along the positive X-axis, if the cumulative offset in the positive Y-axis direction is detected to be 0.03 rad, and the distance from the module to the current feature acquisition point is 10 mm, the offset distance is 0.03 × 10 = 0.3 mm. In this case, the control module moves 0.3 mm in the negative Y-axis direction for compensation.

[0102] If the cumulative deviation of the analysis path exceeds a preset value (0.05 rad), the image analysis module acquires motion noise intensity and direction data for the real-time path. A motion noise change graph is constructed based on a time series (recorded every 0.1 s), with time on the X-axis and noise intensity on the Y-axis, and different colored curves representing different noise directions. The motion noise change trend of the target area within the target object is determined based on the motion noise change graph. For example, a linear fit is used to determine the slope of the noise intensity over time. A positive slope indicates an increasing trend, while a negative slope indicates a decreasing trend. The motion noise trend is interpolated using bilinear interpolation. Evenly distributed interpolation points are selected within the preset range of the shortest initial analysis path (5 mm on either side of the path). The noise values ​​at these interpolation points are calculated based on the noise data of adjacent measurement points to obtain continuous motion noise information. The motion noise information and the noise resistance data are used to determine the operational stability of the image analysis module within the preset range of the shortest initial analysis path. If the noise intensity in a particular area is within the module's resistance range (σ < 0.3 rad / s), the stability of that area is good; otherwise, the stability is poor. Based on operational stability, the shortest initial analysis path segments with cumulative analysis path deviations greater than a preset value are adjusted, avoiding areas of poor stability and replanning the path in adjacent areas of good stability to obtain an updated analysis path. The image analysis module then performs analysis operations based on the updated analysis path. For example, if the noise intensity σ=0.5 rad / s in a section of the original path (exceeding the resistance capacity of 0.3 rad / s), a search is performed within 5mm on both sides of the section for areas with noise intensity σ<0.3 rad / s. These areas are reconnected to form a new path segment, replacing the original one.

[0103] Example 4:

[0104] Obtain standard feature data of different bone density levels of the target object. These data come from cases with clear bone density levels in previous clinical diagnoses, covering multiple parts such as the lumbar spine and hip. Standard feature data include parameters such as bone density value, trabecular thickness, trabecular spacing, and cortical thickness. The standard feature data are labeled with bone density levels. Bone density levels are divided into four levels according to the standards of the World Health Organization: normal, osteopenia, osteoporosis, and severe osteoporosis. Each level corresponds to a different bone density T value range. Normal is T value ≥ -1.0, osteopenia is -2.5 < T value < -1.0, osteoporosis is T value ≤ -2.5, and severe osteoporosis is T value ≤ -2.5 and accompanied by one or more brittle fractures. The labeled feature data forms labeled feature data.

[0105] The minimum size of the filtering window was set to 3×3 and the maximum size was set to 9×9. For the bone structure image data included in the annotated feature data, the median, maximum, and minimum pixel values ​​within the current window were calculated. If the difference between the maximum and minimum values ​​within the current window was less than the noise threshold (set to 20 based on the image signal-to-noise ratio), the median was retained as the filtered pixel value. If the difference was greater than or equal to the noise threshold, the filtering window size was expanded by 2 pixels at a time. The calculation was repeated until the window reached the maximum size of 9×9 or the noise suppression condition was met. De-noised bone structure image data were obtained and incorporated into the annotated feature data.

[0106] The residual network was designed with 18 base convolutional layers and 8 skip connection layers. The initial learning rate was set to 0.001, and the batch size was set to 32. The base convolutional layers used a 3×3 convolution kernel. Each convolutional layer was followed by a batch normalization layer and a Reluctant Unit (ReLU) activation function. The skip connection layer element-wise summed the outputs of the previous and next layers to mitigate the vanishing gradient problem in deep networks. The labeled feature data was divided into a training set and a validation set with a ratio of 7:3. The training set was used for model parameter learning, while the validation set was used to evaluate the model's generalization ability. The cross-entropy loss function was used to calculate the error between the model's predicted values ​​and the true values. The cross-entropy loss function effectively measures the difference between two probability distributions and is suitable for loss calculation in multi-classification tasks. The model parameters were iteratively updated using a stochastic gradient descent optimizer. Each iteration adjusted the parameters based on the gradient direction of the loss function. The learning rate was decayed with the number of iterations using a cosine annealing method. Training was terminated when the validation set accuracy did not improve for five consecutive epochs and reached a preset threshold (e.g., 85%). This resulted in a trained bone density recognition model.

[0107] The image analysis module acquires bone structural feature data from the target region, specifically the lumbar vertebrae L1-L4, based on an image analysis strategy. This strategy includes the distribution of feature acquisition points (20 evenly distributed points per vertebra) and the acquisition sequence (collection proceeds sequentially from the superior edge of the L1 vertebra to the inferior edge of the L4 vertebra). The acquired bone structural feature data includes parameters such as bone density (g / cm²), trabecular thickness (measured using a grayscale gradient method, in mm), trabecular count (number of trabeculae per unit area), and cortical thickness (measured using edge detection, in mm) at each point. This bone structural feature data is then fed into a trained bone density recognition model for bone density classification. The model extracts deep features from the feature data through multiple layers of convolution. A fully connected layer then outputs a probability distribution of four bone density levels, with the highest probability level being the classification result.

[0108] The degree of osteoporosis is assessed within each bone density category. Within the normal category, bone density values ​​between -1.0 and 0, with regular, continuous trabeculae, are considered stable. Values ​​close to -1.0 indicate slightly sparse trabeculae, indicating a potential risk of osteopenia. Within the osteopenia category, values ​​close to -1.0 or -2.5 are categorized as mild or moderate osteoporosis. Mild osteoporosis corresponds to values ​​between -1.0 and -1.5, with slightly increased intertrabecular spacing; moderate osteoporosis corresponds to values ​​between -1.5 and -2.5, with decreased trabecular continuity. Within the osteoporosis category, the presence of trabecular fractures is taken into account to categorize patients as mild (occasional trabecular fractures) or moderate (multiple trabecular fractures). In the severity of osteoporosis, based on the number and location of fragility fractures, a single vertebral compression fracture is considered severe osteoporosis, while multiple vertebral fractures or hip fractures are considered extreme osteoporosis. The results of bone density grade identification and osteoporosis severity assessment are integrated to generate diagnostic data that includes the location of the target area, bone density value, bone density grade, and description of osteoporosis severity.

[0109] Example 5:

[0110] Based on the acquired bone density and bone structure image data of the target region, the bone density image data was preprocessed, including grayscale normalization and edge enhancement. Grayscale normalization linearly mapped the image grayscale values ​​to a range of 0-255. Edge enhancement used the Sobel operator to calculate the gradient amplitude, enhancing pixels with gradient values ​​greater than a threshold (set to 50). The bone structure image data was segmented, using the Otsu thresholding method to automatically determine the segmentation threshold, separating bone tissue from surrounding soft tissue, and generating a binary bone structure image.

[0111] The preprocessed bone density imaging data were reconstructed in three dimensions using the MarchingCubes algorithm. The isosurface threshold was set to 120 (corresponding to the average grayscale value of bone tissue), dividing the three-dimensional space into a grid of cubes with a side length of 0.5 mm. The intersection points between the cubes and the isosurfaces were determined based on the relationship between the grayscale value of each cube vertex and the isosurface threshold. These intersection points were then connected to form triangular facets, creating a three-dimensional surface model of bone density. The segmented bone structure imaging data were then reconstructed in three dimensions using voxel rendering. Binarized two-dimensional slices were stacked layer by layer, with each voxel assigned a corresponding pixel value (1 for bone tissue and 0 for background). A ray casting algorithm was then used to calculate the intersection of a ray from the observation point to each voxel. The color and transparency of the voxel at the intersection were determined based on the voxel value at the intersection, resulting in a realistic three-dimensional bone structure model.

[0112] Bone density distribution characteristics were extracted from the 3D bone density model, including mean, standard deviation, maximum, minimum, and volume fractions of different density intervals (e.g., 0-200 HU, 200-400 HU, etc.). Trabecular morphology characteristics were extracted from the 3D bone structure model, including trabecular thickness (calculated using distance transformation), trabecular spacing (measured after skeletonization), trabecular number density (number of trabeculae per unit volume), and trabecular connection density (number of trabecular connection points per unit volume).

[0113] Based on the extracted bone density distribution and trabecular morphological features, a eigenvector was constructed. The eigenvector had 15 dimensions: the first 5 dimensions represented bone density distribution characteristics (mean, standard deviation, maximum, minimum, and volume fraction of high-density areas), and the last 10 dimensions represented trabecular morphological characteristics (mean thickness, thickness standard deviation, mean spacing, spacing standard deviation, number density, and connection density). The weights of each eigenvalue in the eigenvector were calculated and determined using the analytic hierarchy process. A judgment matrix was constructed, in which the elements represent the relative importance of two features. For example, the relative importance of mean bone density and trabecular thickness was set to 3, indicating that mean bone density is slightly more important than trabecular thickness. The maximum eigenvalue of the judgment matrix and its corresponding eigenvector were normalized to obtain the weight vector for each feature. The weight vector elements ranged from 0 to 1, and the sum of all elements was 1.

[0114] Bone density and bone structure image data were registered using a mutual information-based registration method. The spatial transformation parameters (translation and rotation angle) of the two images were initialized to zero, and the mutual information between the two images under the current transformation was calculated. The mutual information reflects the statistical dependence between the two images. The transformation parameters were iteratively adjusted using an optimization algorithm (such as the Powell algorithm) to maximize the mutual information. Iterations were terminated when the change in mutual information between two consecutive iterations was less than a threshold (set to 0.001). The optimal spatial transformation parameters were then obtained, and the bone structure image data were transformed into the coordinate system of the bone density image data.

[0115] The registered bone density and bone structure image data were fused using a weighted averaging method. Fusion weights were set to 0.6 (for the bone density image) and 0.4 (for the bone structure image). Pixel values ​​at the same location after registration were weighted averaged to generate the fused image data. The fused image data were post-processed using a Gaussian filter to remove noise generated during the fusion process. The standard deviation of the Gaussian filter was set to 0.8, and the window size was 3×3.

[0116] Feature extraction was performed on the post-processed fused image data using a multi-scale analysis method. A Gaussian pyramid was used to decompose the fused image into four sub-images of different scales, with each sub-image having half the resolution of the previous layer. Texture features were extracted for each sub-image, including gray-level co-occurrence matrix features (contrast, energy, entropy, and correlation) and local binary pattern features (homogeneity and amount of inhomogeneity).

[0117] The extracted texture features are integrated with the previously established feature vector to form a comprehensive feature vector. The comprehensive feature vector has 25 dimensions: the first 15 dimensions represent bone density distribution and trabecular morphology, and the last 10 dimensions represent texture features from the fused image. Osteoporosis is diagnosed based on the comprehensive feature vector using a support vector machine classifier. The comprehensive feature vector is input into a trained support vector machine model, which outputs a diagnosis of osteoporosis, which includes four categories: normal, osteopenia, osteoporosis, and severe osteoporosis.

[0118] Diagnostic results are visualized by color-coding areas of varying osteoporosis severity on the 3D bone model: normal areas are green, areas with osteopenia are yellow, areas with osteoporosis are orange, and areas with severe osteoporosis are red. The range and proportion of each area are also displayed, along with the value and weight of each feature in the comprehensive feature vector, to form an intuitive diagnostic report.

[0119] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0120] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing osteoporosis based on image recognition, characterized in that: The following steps are involved: Acquiring bone density image data of a target object and bone structure image data of a target area based on medical imaging equipment, and determining the distribution of trabecular bone in the target area according to the bone structure image data; constructing a three-dimensional bone density model of the target object based on the bone density image data, and determining an image analysis strategy based on the three-dimensional bone density model and trabecular distribution; Loading a preset image analysis module, the image analysis module is used to obtain bone structure feature data of the target area according to the image analysis strategy, identify the bone density grade and osteoporosis degree of the target area based on the feature data, and obtain diagnostic data; The three-dimensional bone density model of the target object is constructed based on the bone density image data, and the image analysis strategy is determined based on the three-dimensional bone density model and the distribution of trabeculae, specifically: Performing multi-dimensional morphological analysis on the bone density image data, constructing a bone density feature tensor through block processing, and identifying bone tissue types based on the grayscale signals of the bone density image data using a random forest classifier to generate a three-dimensional bone density model of the target object that includes bone density values, morphological features, and tissue type labels; Performing a position mapping operation on the trabecular distribution and the three-dimensional bone density model to construct a three-dimensional spatial distribution model of bone density resources, and dividing the three-dimensional spatial distribution model of bone density resources according to a preset block size to construct M subspace regions; Obtaining trabecular gap information of each subspace region according to the trabecular distribution, and clustering subspace regions with similar trabecular gaps in the three-dimensional spatial distribution model of bone density resources based on a hierarchical clustering algorithm to obtain a clustering result; Determining the distribution probability of trabeculae in each subspace region based on the clustering result, obtaining corresponding position information of each subspace region in the target object, and performing a bone density detection importance assessment on each position in the target region in the target object based on the distribution probability and the corresponding position information to obtain a detection importance score for each position in the target region; Determining analysis requirement information for each location in the target area according to the detection importance score of each location, wherein the analysis requirement information includes whether to perform analysis and analysis accuracy requirement information; Obtaining feature extraction clarity data of bone structures at different distances by the image analysis module, determining feature collection distance information for each position of the target area by the image analysis module based on the feature extraction clarity data and analysis requirement information, and determining a feature collection point set for the image analysis module based on the feature collection distance information; An image analysis strategy of an image analysis module is determined according to the feature acquisition point set.

2. The osteoporosis diagnosis method based on image recognition according to claim 1, characterized in that: The step of acquiring bone density image data of the target object and bone structure image data of the target area based on a medical imaging device, and determining the trabecular distribution of the target area according to the bone structure image data, is specifically as follows: Acquiring bone density image data of a target object and bone structure image data of a target area in the target object based on a medical imaging device; performing a noise suppression operation on the bone structure image data based on an adaptive median filter algorithm to obtain denoised bone structure image data, introducing a pyramid decomposition algorithm, and setting an initial number of layers of the pyramid decomposition algorithm parameters; Performing multi-scale decomposition on the denoised bone structure image data according to the pyramid decomposition algorithm to obtain m sub-images with different resolutions; Performing grayscale statistical analysis on the sub-images, constructing a grayscale histogram of each sub-image, calculating an average grayscale value, grayscale variance, and entropy value of each sub-image based on the grayscale histogram, and constructing a feature statistical matrix using the average grayscale value, grayscale variance, and entropy value; Calculating the contrast, energy, and correlation of each pixel point collected by the medical imaging device according to the characteristic statistical matrix to obtain a trabecular bone characteristic spectrum; Acquiring sampling point position information of the trabecular bone characteristic spectrum according to the bone structure image data, mapping the trabecular bone characteristic spectrum of each sampling point according to the sampling point position information, and constructing a three-dimensional trabecular bone characteristic spectrum; The trabecular gap at each sampling point is determined based on the three-dimensional trabecular characteristic spectrum, a trabecular distribution heat map and a trabecular gap gradient field of the target area are constructed based on the trabecular gap at each sampling point, and the trabecular distribution of the target area is determined based on the trabecular distribution heat map and the trabecular gap gradient field.

3. The osteoporosis diagnosis method based on image recognition according to claim 1, characterized in that: The image analysis strategy of the image analysis module is determined according to the feature acquisition point set, specifically: Obtaining three-dimensional coordinate information of each feature acquisition point and initial position information of the image analysis module, and determining structural obstacles in the target object based on the three-dimensional bone density model; Optimizing the path of the initial position information and the three-dimensional coordinate information of each feature acquisition point based on the Dijkstra algorithm, taking the structural obstacle as a path optimization restriction area, and outputting the shortest initial analysis path of the image analysis module; acquiring in real time posture change data of the image analysis module during analysis by the image analysis module according to the shortest initial analysis path, and determining the motion noise direction and motion noise intensity of the real-time analysis position of the image analysis module according to the posture change data; Acquiring operational stability data of the image analysis module, wherein the operational stability data includes data on the image analysis module's resistance to motion noises of different directions and intensities; Acquiring recognition response time data of the image analysis module to the direction and intensity of the motion noise, and determining an analysis path compensation hysteresis amount of the image analysis module according to the recognition response time data; Analyzing the direction and intensity of the motion noise at the real-time analysis position of the image analysis module according to the resistance data and the analysis path compensation hysteresis, and determining the cumulative amount of the analysis path deviation within the recognition response time of the image analysis module; If the cumulative amount of the analysis path offset is less than a preset value, determining an analysis offset direction and an analysis offset distance for analysis by the image analysis module according to the shortest initial analysis path based on the cumulative amount of the analysis path offset, and determining a compensation direction and a compensation distance of the image analysis module based on the analysis offset direction and the analysis offset distance to obtain compensation data; Performing a compensation operation on the shortest initial analysis path of the image analysis module during real-time operation according to the compensation data; If the cumulative amount of the analysis path deviation is greater than a preset value, obtaining motion noise intensity and motion noise direction data of the real-time running path of the image analysis module, constructing a motion noise change map based on the motion noise intensity and motion noise direction data of the real-time running path, and determining a motion noise change trend of a target area in the target object based on the motion noise change map; An interpolation operation is performed on the motion noise change trend based on bilinear interpolation to determine the motion noise information within the preset range of the shortest initial analysis path. The operational stability of the image analysis module within the preset range of the shortest initial analysis path is determined based on the motion noise information and resistance data. Based on the operational stability, the shortest initial analysis path section with a cumulative analysis path offset greater than a preset value is adjusted to obtain an updated analysis path. The image analysis module performs an analysis operation based on the updated analysis path.

4. The osteoporosis diagnosis method based on image recognition according to claim 1, characterized in that: The image analysis module obtains bone structure feature data of the target area according to the image analysis strategy, identifies the bone density level and osteoporosis degree of the target area according to the feature data, and obtains diagnostic data, specifically: Obtaining standard feature data of different bone density levels of the target object, and labeling the standard feature data with the bone density level to obtain labeled feature data; Building a bone density recognition model based on a residual network, and importing the labeled feature data into the bone density recognition model for training; The image analysis module obtains bone structure feature data of the target area according to the image analysis strategy, imports the bone structure feature data into the trained bone density recognition model to identify the bone density level, and evaluates the degree of osteoporosis of each bone density level to obtain diagnostic data.

5. The osteoporosis diagnosis method based on image recognition according to claim 4, characterized in that: The diagnostic data is transmitted to the terminal display device according to digital communication technology, specifically: Building a layered data communication protocol stack based on digital communication technology, encoding the diagnostic data, and building a diagnostic data digital transmission signal; The diagnostic data digital transmission signal is digitally transmitted to a terminal display device according to the layered data communication protocol stack, and a decoding operation is performed on the diagnostic data digital transmission signal to obtain osteoporosis diagnostic data of the target area.

6. The osteoporosis diagnosis method based on image recognition according to claim 4, characterized in that: The bone density recognition model is constructed based on the residual network, and the labeled feature data is imported into the bone density recognition model for training, specifically: Determine the number of basic convolutional layers and skip connection layers of the residual network, and set the initial learning rate and batch size parameters; The labeled feature data is divided into a training set and a validation set, and the error between the model prediction value and the true value is calculated using a cross entropy loss function; The model parameters are iteratively updated through the stochastic gradient descent optimizer. When the accuracy of the validation set reaches the preset threshold, the training is stopped to obtain the trained bone density recognition model.

7. The osteoporosis diagnosis method based on image recognition according to claim 5, characterized in that: The layered data communication protocol stack is constructed based on digital communication technology, the diagnostic data is encoded, and a diagnostic data digital transmission signal is constructed, specifically: Define the modulation mode and transmission rate of the physical layer, set the frame header format and verification rules of the data link layer, and determine the routing strategy and address allocation of the network layer; The diagnostic data is converted into binary form, and a start flag, a length field and a check code are added according to the frame format of the data link layer to generate a diagnostic data digital transmission signal.

8. The osteoporosis diagnosis method based on image recognition according to claim 3, characterized in that: The path optimization is performed on the initial position information and the three-dimensional coordinate information of each feature acquisition point based on the Dijkstra algorithm, and the structural obstacle is used as the path optimization restriction area to output the shortest initial analysis path of the image analysis module, specifically: Construct a three-dimensional coordinate grid map of the target area and map the three-dimensional coordinates of each feature collection point into a grid node; Initialize the distance value of each node to infinity, set the distance value of the initial position node to zero, and use the priority queue to store the nodes to be processed; Take the node with the smallest distance value in turn, update the distance value of its adjacent node, and skip it if the adjacent node is a structural obstacle; When the distance values ​​of all feature acquisition point nodes are updated, the shortest initial analysis path is obtained by backtracking.

9. The osteoporosis diagnosis method based on image recognition according to claim 2, characterized in that: The noise suppression operation is performed on the bone structure image data based on the adaptive median filtering algorithm to obtain the denoised bone structure image data, specifically: Set the minimum and maximum sizes of the filter window and calculate the median, maximum and minimum values ​​of the pixels in the current window; If the difference between the maximum and minimum values ​​in the current window is less than the noise threshold, the median value is retained as the filtered pixel value; If the difference is greater than or equal to the noise threshold, the filter window size is expanded and the calculation is repeated until the window reaches the maximum size or the noise suppression condition is met, thereby obtaining the denoised bone structure image data.

Images