Bone tissue compressive strength detection data analysis method and system

Abstract

The invention provides a bone tissue compressive strength detection data analysis method and system, and the method comprises the steps: debugging a bone tissue form recognition neural network in a debugging process of the bone tissue form recognition neural network on the premise that an annotated bone tissue form mark in a bone tissue medical image training set is adjusted, and carrying out the debugging of the bone tissue form recognition neural network. The stability of the bone tissue form recognition neural network is improved, the precision of the debugged bone tissue form recognition neural network is improved, and accurate bone tissue form detection on the target bone tissue medical image through the debugged bone tissue form recognition neural network is realized. Therefore, bone tissue compressive strength detection is indirectly helped to be completed.

Description

Technical Field

[0001] The present application relates to the field of machine learning, and specifically, to a method and system for analyzing bone tissue compressive strength detection data. Background Art

[0002] In the field of medical image analysis, the detection of bone tissue compressive strength is of great significance for assessing bone health, predicting fracture risk, and formulating personalized treatment plans. Traditional methods for detecting bone tissue compressive strength mostly rely on laboratory tests, which are not only costly and time-consuming, but also difficult to be widely used in clinical practice. In recent years, with the rapid development of vision and deep learning technology in detection systems, methods for detecting bone tissue compressive strength using medical image data combined with neural networks have gradually attracted attention. However, there are still some problems with the existing bone tissue compressive strength detection methods based on medical images. First, in the network training stage, medical image data often have the problem of inaccurate labeling, that is, the actual morphology of bone tissue in medical images may not be completely consistent with its labeled morphology, which will affect the subsequent training effect of the neural network for bone tissue morphology recognition, and thus lead to insufficient accuracy in bone tissue morphology recognition. Summary of the invention

[0003] The object of the present invention is to provide a method and system for analyzing bone tissue compressive strength detection data. The embodiment of the present application is implemented as follows: In a first aspect, an embodiment of the present application provides a method for analyzing bone tissue compressive strength detection data, including: obtaining a first training branch, the first training branch including a bone tissue medical image training set and a bone tissue medical image verification set, the bone tissue medical image training set including A categories of first bone tissue morphology, each category of first bone tissue morphology including S training images, each training image including respective actual bone tissue morphology labels and annotated bone tissue morphology labels, the actual bone tissue morphology labels of the training images represent the first bone tissue morphology actually corresponding to the training images, and the annotated bone tissue morphology labels of the training images represent the annotated first bone tissue morphology of the training images; the bone tissue medical image verification set including the A categories of first bone tissue morphology, each category of first bone tissue morphology including T verification images, each verification image including respective actual bone tissue morphology labels, the actual bone tissue morphology labels of the verification images represent the first bone tissue morphology actually corresponding to the verification image; wherein A≥1; T≥1; S≥1; a bone tissue morphology recognition neural network is used to identify the first bone tissue morphology according to the image relationships between the training images in the bone tissue medical image training set. According to the association situation, the annotated bone tissue morphology marks of each training image in the bone tissue medical image training set are adjusted to obtain the adjustment marks, and the morphology reference representation vector of each type of first bone tissue morphology is generated according to the adjustment marks and the image representation vectors of each training image in the bone tissue medical image training set; the bone tissue morphology recognition neural network is used to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set according to the commonality measurement results between the image representation vectors of each verification image in the bone tissue medical image verification set and the morphology reference representation vector of each type of first bone tissue morphology, and obtain the first bone tissue morphology estimation mark of each verification image in the bone tissue medical image verification set; according to the actual bone tissue morphology marks of each training image in the bone tissue medical image training set, as well as the actual bone tissue morphology marks and the first bone tissue morphology estimation marks of each verification image in the bone tissue medical image verification set, the bone tissue morphology recognition neural network is debugged; the bone tissue morphology recognition neural network after debugging is used to identify the bone tissue morphology corresponding to the target bone tissue medical image.

[0004] In a second aspect, the present application provides a detection system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the above-mentioned method is implemented.

[0005] The technical effects of the present application at least include: the bone tissue compressive strength detection data analysis method and system provided by the present application, by obtaining a first training branch, the first training branch includes a bone tissue medical image training set and a bone tissue medical image verification set; using a bone tissue morphology recognition neural network to adjust the annotated bone tissue morphology labels of each training image in the bone tissue medical image training set according to the image association between each training image in the bone tissue medical image training set, and obtain an adjustment label. Based on this, the bone tissue morphology recognition neural network performs label adjustment on the bone tissue medical image training set, and the adjustment process can adaptively adjust the inaccurate annotated bone tissue morphology labels in the bone tissue medical image training set, thereby reducing the disturbance of the bone tissue morphology recognition neural network caused by the training images in the bone tissue medical image training set whose actual corresponding bone tissue morphology is different from the annotated bone tissue morphology in the bone tissue medical image training set. Based on the adjustment mark and the image representation vector of each training image in the bone tissue medical image training set, a morphological reference representation vector of each type of first bone tissue morphology is generated; and a bone tissue morphology recognition neural network is used to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the image representation vectors of each verification image in the bone tissue medical image verification set and the morphological reference representation vector of each type of first bone tissue morphology, so as to obtain the first bone tissue morphology estimation mark of each verification image in the bone tissue medical image verification set; based on the actual bone tissue morphology mark of each training image in the bone tissue medical image training set, as well as the actual bone tissue morphology mark and the first bone tissue morphology estimation mark of each verification image in the bone tissue medical image verification set, the bone tissue morphology recognition neural network is debugged; the bone tissue morphology recognition neural network after debugging is used to identify the bone tissue morphology corresponding to the target bone tissue medical image. Based on this, under the premise that the annotated bone tissue morphology markers in the bone tissue medical image training set are adjusted, the bone tissue morphology recognition neural network is debugged to help improve the stability of the bone tissue morphology recognition neural network, thereby improving the accuracy of the bone tissue morphology recognition neural network after debugging, and helping to achieve accurate bone tissue morphology detection of the target bone tissue medical images through the bone tissue morphology recognition neural network after debugging, thereby indirectly helping to complete the detection of bone tissue compressive strength. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 It is a flow chart of a bone tissue compressive strength detection data analysis method provided in an embodiment of the present application.

[0007] Figure 2 It is a schematic diagram of the composition of a detection system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] The execution subject of the bone tissue compressive strength detection data analysis method provided in the embodiment of the present application is a computer device, including but not limited to a server, a personal computer, a laptop computer, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes: Step S110, obtaining a first training branch, wherein the first training branch includes a bone tissue medical image training set and a bone tissue medical image verification set, wherein the bone tissue medical image training set includes A categories of first bone tissue morphology, each category of the first bone tissue morphology includes S training images, each training image includes respective actual bone tissue morphology labels and annotated bone tissue morphology labels, the actual bone tissue morphology labels of the training images represent the first bone tissue morphology actually corresponding to the training images, and the annotated bone tissue morphology labels of the training images represent the annotated first bone tissue morphology of the training images; the bone tissue medical image verification set includes the A categories of first bone tissue morphology, each category of the first bone tissue morphology includes T verification images, each verification image includes respective actual bone tissue morphology labels, and the actual bone tissue morphology labels of the verification images represent the first bone tissue morphology actually corresponding to the verification image; wherein, A≥1; T≥1; S≥1.

[0009] In the embodiment of the present application, the first training branch or learning task refers to a set of data and related configurations for training a neural network. The first training branch here specifically refers to a data set and configuration for training a neural network for bone tissue morphology recognition. The bone tissue medical image training set is a set of medical image data for training a neural network for bone tissue morphology recognition. These data are, for example, bone tissue images acquired by medical imaging techniques (such as X-rays, CT scans, or MRI). For example, the training set may contain hundreds or even thousands of bone tissue medical images from different patients, each image corresponding to one or more bone tissue morphologies. The bone tissue medical image validation set is an independent data set used to verify the debugging effect of the neural network. During the neural network debugging (i.e., training), the validation set does not participate in the training, but will be used to evaluate the performance of the neural network so as to adjust parameters or select the neural network during the training process. For example, the validation set contains bone tissue medical images similar to the training set, but the images and data points should not overlap with the training set. Through the validation set, researchers can understand the performance of the neural network on unseen data. The first bone tissue morphology refers to the different bone tissue types or states contained in the training set and the validation set. These morphologies may represent different health states, pathological conditions, or age stages. For example, the first bone tissue morphology may include normal bone tissue, osteoporotic bone tissue, bone tissue healed after fracture, etc.

[0010] Training images are specific data points in the training set, i.e., medical images of bone tissue used to train the neural network. For example, a training image may be an X-ray or CT scan image showing a patient's specific bone tissue morphology. Actual bone tissue morphology labels are labels or markers that are determined by expert doctors based on medical knowledge and accurately reflect the true morphology of bone tissue in the image. For example, for an X-ray showing normal bone tissue, its actual bone tissue morphology label may be "normal bone tissue". Annotated bone tissue morphology labels are automatically generated through preliminary analysis or annotation tools, which may have certain errors and require subsequent neural network learning to correct. For example, during the automatic annotation process, an X-ray showing osteoporosis may be initially labeled as "normal bone tissue", but this label may be inaccurate and needs to be optimized through neural network learning.

[0011] Specifically, in step S110, the detection system obtains a specific learning task, namely, a first training branch, which consists of two parts: a bone tissue medical image training set and a bone tissue medical image verification set.

[0012] First, the bone tissue medical image training set is the basis of neural network learning. It contains A different categories of first bone tissue morphology, which may represent bone tissue in different health states, different ages or different pathological conditions. For example, A can be 3, representing three categories: normal bone tissue, osteoporotic bone tissue and bone tissue healing after fracture. For each category of bone tissue morphology, the training set contains S training images, which are bone tissue images obtained by medical imaging technology (such as X-ray, CT scan or MRI). Each training image is equipped with two labels: actual bone tissue morphology label and annotated bone tissue morphology label. The actual bone tissue morphology label is determined by expert doctors based on medical knowledge, and it accurately reflects the true morphology of bone tissue in the image. The annotated bone tissue morphology label may be automatically generated by preliminary analysis or annotation tools, which may have certain errors and require subsequent neural network learning to correct. Then, the role of the bone tissue medical image verification set is to verify the performance of the debugged neural network. This verification set also contains all A bone tissue morphology categories in the training set, but there are only T verification images under each category. These validation images are similar to the training images and are also obtained through medical imaging technology, and each image is marked with actual bone tissue morphology. The main function of the validation set is to regularly evaluate the neural network during the neural network debugging process to ensure that the neural network can maintain good performance on unseen data. In practical applications, for example, in a project to study the effect of osteoporosis on the compressive strength of bone tissue, the detection system obtains a large number of bone tissue medical images of osteoporosis patients and normal people from the medical database, and then divides these images into different bone tissue morphology categories (such as normal bone tissue, mild osteoporotic bone tissue, severe osteoporotic bone tissue, etc.) according to the judgment of experts, and forms training sets and validation sets respectively. Then, the detection system will use these data sets to train a bone tissue morphology recognition neural network so that it can automatically and accurately identify the bone tissue morphology in medical images in the future, and then analyze the compressive strength of bone tissue.

[0013] Step S120, using a bone tissue morphology recognition neural network to adjust the annotated bone tissue morphology marks of each training image in the bone tissue medical image training set according to the image correlation between each training image in the bone tissue medical image training set to obtain an adjustment mark, and based on the adjustment mark and the image representation vector of each training image in the bone tissue medical image training set, generate a morphology reference representation vector for each type of first bone tissue morphology.

[0014] In step S120, the detection system uses the bone tissue morphology recognition neural network to perform in-depth processing on the data in the bone tissue medical image training set, aiming to improve the recognition accuracy of the neural network for bone tissue morphology. This step involves two key sub-processes: the adjustment of the annotated bone tissue morphology markers and the generation of the morphological reference representation vector. First, the detection system adjusts the annotated bone tissue morphology markers by analyzing the image association between each training image in the bone tissue medical image training set. The image association here can refer to the similarity between the images in visual features, the continuity of the bone tissue structure, the diffusion pattern of the lesion area, etc. For example, if two training images are very similar visually, but their annotated bone tissue morphology markers are different, this may be caused by errors in the initial annotation process. In this case, the neural network can learn the similarity between these images and adjust the annotation markers of the current image according to the markers of other similar images. The process of adjusting the annotated bone tissue morphology markers can be implemented by a variety of machine learning techniques, such as Graph Neural Network (GNN). GNN can process the association relationship between images, regard the training images as vertices (i.e., nodes) in the graph, and the edges between the vertices represent the similarity between the images. Through iterative learning of GNN, the detection system can generate an adjusted label for each training image, which is closer to the actual bone tissue morphology of the image.

[0015] Then, the detection system generates a morphological reference representation vector for each first bone tissue morphology based on the adjusted labels and the image representation vectors of the training images. The image representation vector is a feature vector extracted from the training images by a deep learning neural network (such as a convolutional neural network, CNN), which captures key information in the image, such as the texture, shape, and density of the bone tissue. The adjusted labels provide information about the bone tissue morphology category to which each training image belongs. In order to generate the morphological reference representation vector, the detection system first selects the training images belonging to the same bone tissue morphology category and extracts their image representation vectors. Then, these vectors are aggregated (such as average pooling) to obtain a morphological reference representation vector representing the bone tissue morphology category. This vector can be regarded as a center point of the bone tissue morphology category in the feature space, which represents the representative features of the bone tissue morphology of this category. For example, suppose that the training set contains three bone tissue morphology categories: normal bone tissue, osteoporotic bone tissue, and fracture healing bone tissue. For each bone tissue morphology, the detection system selects the training images belonging to the category, extracts their image representation vectors, and generates a morphological reference representation vector. These morphological reference representation vectors will serve as a benchmark for subsequent bone tissue morphology estimation, helping the neural network to more accurately identify bone tissue morphology in new bone tissue medical images.

[0016] Step S130, using the bone tissue morphology recognition neural network to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the image representation vectors of each verification image in the bone tissue medical image verification set and the morphological reference representation vectors of each type of first bone tissue morphology, to obtain the first bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set.

[0017] In step S130, the detection system uses the debugged bone tissue morphology recognition neural network to estimate the bone tissue morphology of the images in the bone tissue medical image verification set.

[0018] First, the detection system extracts features from each verification image in the verification set and generates a corresponding image representation vector. This vector contains, for example, key information in the image, such as the texture, shape, density and other features of the bone tissue, which are essential for distinguishing different bone tissue morphologies. These image representation vectors are automatically extracted from the original image by a deep learning neural network (such as a convolutional neural network CNN), and they capture the complex patterns and structures in the image. Then, the detection system compares the image representation vector of each verification image with the morphological reference representation vector of each type of first bone tissue morphology. The morphological reference representation vector is generated during the training process and represents the representative features of each type of bone tissue morphology in the feature space. By calculating the commonality measurement results between the image representation vector of the verification image and the morphological reference representation vector (i.e., the commonality evaluation results of the two, such as cosine similarity, Euclidean distance, etc.), the detection system can evaluate the similarity between the verification image and each type of bone tissue morphology.

[0019] In this process, if the commonality measurement result between the image representation vector of a certain verification image and the morphological reference representation vector of a certain type of bone tissue morphology is the highest, then the detection system will estimate the verification image as belonging to this type of bone tissue morphology and assign it the corresponding first bone tissue morphology estimation tag. This estimation tag represents the judgment result of the detection system on the bone tissue morphology in the verification image. For example, suppose that the verification set contains a medical image of bone tissue suspected of osteoporosis. The detection system extracts the image representation vector of the image and then compares it with the morphological reference representation vectors of three types of bone tissue morphology: normal bone tissue, osteoporotic bone tissue, and fracture healing bone tissue. If the calculation result shows that the commonality measurement result between the image representation vector of the image and the morphological reference representation vector of osteoporotic bone tissue is the highest, then the detection system will estimate the image as belonging to osteoporotic bone tissue morphology and assign it the corresponding estimation tag. Through this process, the detection system can estimate the bone tissue morphology of each image in the verification set and generate the corresponding estimation tag. These estimation tags can not only help researchers understand the performance of neural networks on unseen data, but also serve as an important basis for subsequent neural network optimization and adjustment.

[0020] Step S140, debugging the bone tissue morphology recognition neural network based on the actual bone tissue morphology labels of each training image in the bone tissue medical image training set, as well as the actual bone tissue morphology labels and the first bone tissue morphology estimation labels of each verification image in the bone tissue medical image verification set; the bone tissue morphology recognition neural network after debugging is used to identify the bone tissue morphology corresponding to the target bone tissue medical image.

[0021] In step S140, the detection system uses the actual bone tissue morphology markers in the bone tissue medical image training set and the validation set, as well as the first bone tissue morphology estimation marker on the validation set, to debug and optimize the bone tissue morphology recognition neural network. This step is a key part of the machine learning process, which aims to improve the accuracy and generalization ability of the neural network in practical applications. First, the detection system obtains the actual bone tissue morphology markers of each image in the training set and the validation set. These markers are obtained by medical experts after analyzing the bone tissue medical images based on their professional knowledge, and represent the true morphology of the bone tissue in the image. For example, in the scenario of bone tissue compressive strength test data analysis, the actual bone tissue morphology markers may include categories such as "normal bone tissue", "osteoporotic bone tissue" and "healing bone tissue after fracture". Then, the detection system uses the bone tissue morphology recognition neural network to estimate the bone tissue morphology of the images in the validation set and generates the first bone tissue morphology estimation marker. These estimation markers represent the judgment results of the neural network on the bone tissue morphology of the images in the validation set. However, since there may be errors or deviations in the training process of the neural network, the estimation markers may not be completely consistent with the actual bone tissue morphology markers. In order to correct these errors and improve the performance of the neural network, the detection system compares the actual bone tissue morphology markers and the first bone tissue morphology estimated markers of each image in the validation set and calculates the difference or loss between them. This loss function measures the gap between the predicted results of the neural network and the actual results, and is a key indicator for neural network optimization. Based on this loss function, the detection system will use the back propagation algorithm and the gradient descent algorithm to debug and optimize the bone tissue morphology recognition neural network. The back propagation algorithm is used to calculate the gradient of the loss function with respect to the neural network parameters, while the gradient descent algorithm updates the neural network parameters according to these gradients to minimize the loss function. Through this iterative process, the neural network parameters will gradually adjust to the optimal state, making the predicted results of the neural network closer to the actual results.

[0022] After debugging, the detection system will obtain a bone tissue morphology recognition neural network with better performance. This network can not only perform well on the training set, but also maintain a high accuracy on the unseen validation set. Therefore, it will be used to identify the bone tissue morphology corresponding to the target bone tissue medical image, and provide accurate and reliable input for the subsequent bone tissue compressive strength test data analysis. For example, suppose there is a bone tissue medical image showing osteoporosis in the validation set, and its actual bone tissue morphology is marked as "osteoporotic bone tissue". However, in the preliminary estimation, the bone tissue morphology recognition neural network may mistakenly estimate it as "normal bone tissue". Through the debugging process of step S140, the detection system will discover this error and correct it by optimizing the neural network parameters. In the end, the debugged network will be able to more accurately identify this image as "osteoporotic bone tissue", thereby providing more valuable information for the subsequent bone tissue compressive strength test data analysis.

[0023] In some feasible implementation schemes, the bone tissue morphology recognition neural network includes a label adjustment network component, and the label adjustment network component includes a graph representation branch component and an iterative adjustment branch component. At this time, the step S120, using the bone tissue morphology recognition neural network to adjust the annotated bone tissue morphology labels of each training image in the bone tissue medical image training set according to the image association between each training image in the bone tissue medical image training set to obtain the adjustment label, includes: Step S121, using the graph representation branch component, and mapping each training image in the bone tissue medical image training set to the bone tissue morphological feature space according to the image association between each training image in the bone tissue medical image training set.

[0024] In step S121, the detection system uses a graph representation branch component, which is a special machine learning neural network component, for example, based on the principle of a graph neural network (GNN). Its core purpose is to map these images into a high-dimensional space called "bone tissue morphological feature space" based on the image association between each training image in the bone tissue medical imaging training set. Each point in this space represents a training image, and the similarity between images is reflected in the space by the distance between the points. Specifically, the graph representation branch component first regards each training image in the training set as a vertex in the graph. The edges between these vertices represent the association between the images. In the scenario of bone tissue compressive strength detection data analysis, the association between images may be based on a variety of factors, such as the similarity of bone tissue texture, shape, density and other features, or based on the annotation information of medical experts.

[0025] The graph representation branch component then uses a graph neural network to learn the relationships represented by these edges. A graph neural network is a type of neural network that can process graph-structured data. It can update the feature representation of a vertex by iteratively aggregating the neighbor information of the vertex. In this process, the feature representation of each vertex (i.e., each training image) gradually fuses the information of its neighbor vertices to form a global representation that contains the image association information. As the graph neural network learns, each training image in the training set is mapped to a point in the bone tissue morphology feature space. The location of this point depends on the characteristics of the image itself and its association with other images. For example, if two training images are very similar visually, their positions in the bone tissue morphology feature space will also be very close. On the contrary, if there are large differences in the features of the two images, they will be farther apart in space.

[0026] In this way, the graph representation branch component successfully maps the images in the training set to a bone tissue morphological feature space that can reflect the image association. This space provides the basis for iterative adjustments in subsequent steps, allowing the detection system to optimize their annotated bone tissue morphological labels based on the position of the images in space. For example, suppose the training set contains two medical images of bone tissue showing osteoporosis characteristics. Since the two images are very similar in texture, shape, etc., the graph representation branch component will map them to similar positions in the bone tissue morphological feature space. This means that in subsequent steps, the detection system can more easily identify the similarities between the two images and adjust their annotation labels accordingly.

[0027] Step S122, using the iterative adjustment branch component to adjust the annotated bone tissue morphology mark of each training image in the bone tissue medical image training set according to the spatial coordinates of each training image in the bone tissue morphology feature space to obtain an adjustment mark.

[0028] In step S122, the detection system uses the iterative adjustment branch component to adjust the annotated bone tissue morphology labels of each training image in the bone tissue medical image training set according to the spatial coordinates of the training images in the bone tissue morphology feature space (i.e., their distribution positions in the high-dimensional space) to obtain more accurate adjustment labels. This step is intended to optimize the labels of the training data and improve the accuracy of subsequent bone tissue morphology recognition.

[0029] In step S121, the role of the branch component is represented by a graph, and each training image in the training set is mapped to this feature space. The position of each image in the feature space (i.e., the spatial coordinates) reflects the similarity relationship between it and other images. Specifically, images with similar positions are more similar in bone tissue morphology, while images with farther positions are more different. Then, the iterative adjustment branch component comes into play. This component is an algorithmic architecture that does not require parameter optimization, but performs iterative adjustments based on existing feature space coordinates and annotated bone tissue morphology markers. The goal of the adjustment is to obtain consistent labels for images with similar bone tissue morphology, while correcting those erroneous labels caused by initial annotation errors or image feature differences.

[0030] The specific implementation of the adjustment process may include the following steps: Compute spatial similarity: First, the iterative conditioning branch computes the similarity between each training image and its neighbor images (i.e., images that are close in position in the feature space). This can be done by computing the Euclidean distance, cosine similarity, or other metrics between their spatial coordinates.

[0031] Identify potential mislabeling: The component then identifies images that are similar to neighboring images in feature space but have inconsistent annotated bone tissue morphology. These images are likely to be mislabeled during the initial annotation process.

[0032] Adjusting annotation labels: For the identified potentially mislabeled images, the iterative adjustment branch component re-adjusts the annotation labels of the image based on the labels of its neighboring images and the similarity between them. The adjustment method may be weighted averaging, voting decision or other appropriate strategies.

[0033] Iterative optimization: The above steps are repeated to form an iterative optimization process. In each iteration, the component adjusts the mark according to the current feature space coordinates and annotation marks until a certain convergence condition or number of iterations is reached.

[0034] Through the iterative adjustment of step S122, the detection system can obtain more accurate and reliable adjustment marks, providing a high-quality data basis for subsequent bone tissue morphology estimation and neural network debugging.

[0035] In some feasible implementation schemes, each training image in the bone tissue medical image training set includes respective morphological representation information in the bone tissue morphological feature space. Then, the step S122, using the iterative adjustment branch component to adjust the annotated bone tissue morphological mark of each training image in the bone tissue medical image training set according to the spatial coordinates of each training image in the bone tissue morphological feature space in the bone tissue morphological feature space, to obtain the adjustment mark, includes: Step S1221, determines the first importance factor of each training image in the bone tissue medical image training set based on the spatial coordinates of each training image in the bone tissue medical image training set and other training images with the same annotated bone tissue morphological markers in the bone tissue morphological feature space, wherein the first importance factor of the training image represents the adaptation support coefficient between the annotated bone tissue morphological markers contained in the training image and the actual bone tissue morphological markers.

[0036] In step S1221, the detection system performs an in-depth analysis of each training image in the bone tissue medical image training set to determine their first importance factor. This factor is a weight parameter that reflects the adaptation support coefficient between the annotated bone tissue morphological markers on the training image and the actual bone tissue morphological markers, that is, the accuracy of the markers or the possibility of matching. First, the detection system needs to clarify which training images have the same annotated bone tissue morphological markers. In the scenario of bone tissue compressive strength detection data analysis, these markers may represent different bone tissue states, such as "normal bone tissue", "osteoporotic bone tissue" or "healing bone tissue after fracture". For each set of images with the same annotation markers, the detection system will further analyze their positional relationship in the bone tissue morphological feature space.

[0037] The bone tissue morphology feature space is a high-dimensional space, in which each point represents a training image, and the position of the point (i.e., spatial coordinates) is determined by the morphological representation information of the image. The morphological representation information is, for example, a feature vector extracted from the original medical image by a deep learning neural network (such as a convolutional neural network CNN), which captures key information in the image, such as the texture, shape, and density of the bone tissue. In step S1221, the detection system compares the positions of images with the same annotation mark in the feature space. If an image is very close to other images with the same annotation mark in spatial position, it means that they have a high similarity in bone tissue morphology. Based on this similarity, the detection system can assign a higher first importance factor to the image, indicating that the adaptation support coefficient between its annotation mark and the actual bone tissue morphology is high, that is, the accuracy of the mark is high. On the contrary, if an image is far away from other images with the same annotation mark in spatial position, this indicates that they have differences in bone tissue morphology. In this case, the detection system may assign a lower first importance factor to the image, indicating that the adaptation support coefficient between its annotation mark and the actual bone tissue morphology is low, that is, the accuracy of the mark is questionable.

[0038] For example, assume that there are two images A and B annotated as "osteoporotic bone tissue" in the training set. In the bone tissue morphological feature space, image A is close to multiple other images annotated as "osteoporotic bone tissue", while image B is farther away from these images. In this case, the detection system may assign a higher first importance factor to image A because it is highly similar in morphology to images of the same category; and assign a lower first importance factor to image B because it may be a mislabeled sample. Through step S1221, the detection system can determine a quantitative first importance factor for each training image, which provides important reference information for labeling adjustments in subsequent steps.

[0039] Step S1222: Generate a basic reference representation vector of each type of annotated bone tissue morphological markers according to the first importance factor and morphological representation information of all training images containing the same annotated bone tissue morphological markers.

[0040] In step S1222, the detection system generates a basic reference representation vector for each type of annotated bone tissue morphological marker based on the first importance factor of all training images containing the same annotated bone tissue morphological marker and their morphological representation information. This process is a key step in the iterative adjustment branch component, which aims to establish a stable reference point for fine-tuning the annotated bone tissue morphological markers of the training images in subsequent steps. In the scenario of bone tissue compressive strength detection data analysis, each type of bone tissue morphology (such as normal bone tissue, osteoporotic bone tissue, etc.) needs to have a representative vector that can summarize the main features of this type of bone tissue morphology. This representative vector is the basic reference representation vector.

[0041] Then, see the specific implementation process of step S1222. The detection system identifies all training images with the same annotated bone tissue morphology markers. These images each have a morphological representation information (for example, a high-dimensional vector) in the bone tissue morphology feature space, and this vector captures the key features of the bone tissue in the image. At the same time, each image also has a first importance factor, which reflects the adaptation support coefficient between the image's annotation mark and the actual bone tissue morphology. With this information, the detection system begins to construct a basic reference representation vector for each type of annotated bone tissue morphology marker. This process, for example, involves a weighted average operation. Specifically, the detection system weights and sums the morphological representation information of all training images under the same type of bone tissue morphology according to their first importance factors, and then divides it by the sum of the importance factors to obtain an averaged vector. This averaged vector is the basic reference representation vector of the type of bone tissue morphology. For example, suppose there are 10 training images that are all annotated as "osteoporotic bone tissue." These 10 images each have a morphological representation vector in the bone tissue morphology feature space, and each vector has a corresponding first importance factor. The detection system performs weighted average of these 10 vectors according to their importance factors to obtain an averaged vector, which represents the basic reference representation vector of the category "osteoporotic bone tissue".

[0042] By generating such a basic reference representation vector, the detection system provides a stable reference point for the subsequent label adjustment step. This reference point can help the system more accurately determine whether the annotation labels of each training image are accurate, and make necessary adjustments.

[0043] Step S1223, determining the reference bone tissue morphological markers of each training image in the bone tissue medical image training set based on the commonality measurement results between the morphological representation information of each training image in the bone tissue medical image training set and the basic reference representation vector of each type of the annotated bone tissue morphological marker.

[0044] In step S1223, the detection system further utilizes the function of the iterative adjustment branch component to determine the reference bone tissue morphology label of each training image by comparing the commonality measurement results between the morphological representation information of each training image in the bone tissue medical image training set and the basic reference representation vector of each type of annotated bone tissue morphology label. This step is a key link in the label adjustment process, which aims to assign a more reliable bone tissue morphology label to each image in the training set based on the established basic reference representation vector. The morphological representation information is a high-dimensional feature vector extracted from the original medical image by a deep learning neural network, which captures the key information of the bone tissue in the image, such as texture, shape, and density. The basic reference representation vector is generated in step S1222 and represents the representative features of each type of annotated bone tissue morphology in the feature space. In step S1223, the detection system calculates the similarity between the morphological representation information of each training image and the basic reference representation vector of each type of annotated bone tissue morphology. This similarity can be calculated by various metrics, such as cosine similarity, Euclidean distance, etc. The higher the similarity, the higher the degree of matching between the training image and the type of annotated bone tissue morphology. Then, the detection system assigns a reference bone tissue morphology marker to each training image based on the calculated similarity value. Specifically, the detection system finds the annotated bone tissue morphology marker corresponding to the basic reference representation vector with the highest similarity to the training image morphology representation information, and uses it as the reference bone tissue morphology marker for the training image. This reference marker provides an important reference basis for subsequent marker adjustments.

[0045] Let's take an example to illustrate this process. Suppose there is a medical image of bone tissue in the training set, and its morphological representation information is obtained after extraction through a deep learning neural network. Then, the detection system calculates the similarity between this morphological representation information and the three basic reference representation vectors of "normal bone tissue", "osteoporotic bone tissue" and "bone tissue healing after fracture". If the calculation result shows that the morphological representation information has the highest similarity with the basic reference representation vector of "osteoporotic bone tissue", then the system will determine the reference bone tissue morphological marker of the image as "osteoporotic bone tissue". Through the processing of step S1223, the detection system can assign a reference bone tissue morphological marker to each image in the training set. This marker is based on the feature information of the image itself and the similarity with the representative features of various types of bone tissue morphology, so it is more accurate and reliable. This provides strong support for the subsequent marker adjustment step.

[0046] Step S1224, diffusing the reference bone tissue morphology mark of each training image in the bone tissue medical image training set according to the mark transfer mechanism to obtain an adjustment mark.

[0047] In step S1224, the detection system uses a label transfer mechanism to diffuse the reference bone tissue morphology labels of each training image in the bone tissue medical imaging training set, thereby obtaining an adjustment label. The label transfer mechanism is a semi-supervised learning algorithm that can use known labeled data (here, reference bone tissue morphology labels) to infer the labels of unknown labeled data during the training process, thereby optimizing the labeling quality of the entire data set. In the scenario of bone tissue compressive strength test data analysis, this step is crucial to improving the accuracy and reliability of the training data set. In the label transfer mechanism, the training data is regarded as a graph structure, in which each training image is regarded as a vertex in the graph, and the edges between the vertices represent the similarity relationship between the images. Vertices with known labels (i.e., images with reference bone tissue morphology labels) are regarded as known vertices in the graph, while vertices with unlabeled or uncertain labels are regarded as unknown vertices. The label transfer mechanism propagates the label information of known vertices in the graph in an iterative manner, so that the labels of unknown vertices gradually stabilize, and finally obtain adjustment labels. When implementing step S1224, the detection system constructs a label diffusion table (essentially, it can be understood as a transfer matrix), which describes the probability of label transfer between each vertex in the graph. This probability is calculated based on a similarity measure between vertices (such as cosine similarity, Gaussian kernel function, etc.). In the context of a bone tissue medical image training set, the similarity between vertices can be determined by comparing their morphological representation information.

[0048] Then, the detection system updates the label probability distribution of each vertex according to the label diffusion table and the reference bone tissue morphological label. Specifically, for each unknown vertex, the detection system updates the label probability distribution of the unknown vertex according to the label information of its neighbor vertices (vertices with known labels) and the similarity relationship between them. This process is carried out through iterations, and each iteration will further diffuse the label information in the graph, so that the label of the unknown vertex gradually tends to be stable. For example, suppose there are two bone tissue medical images A and B with inconsistent initial labels in the training set, but their positions in the morphological feature space are very close, and they are similar to the images known to be labeled "osteoporotic bone tissue". In step S1224, the detection system identifies the neighbor vertices of A and B (i.e., images with known labels) and constructs a label diffusion table based on the similarity relationship between them. Then, the detection system updates the label probability distribution of A and B based on this table and the label information of the neighbor vertices. Since A and B have a high similarity with the known “osteoporotic bone tissue” image, after multiple iterations, their labeling probability distributions will gradually converge to “osteoporotic bone tissue”, thereby obtaining the adjusted labeling.

[0049] Through the label transfer mechanism of step S1224, the detection system can make full use of the known labeled data in the training set to optimize the labels of the unknown labeled data, thereby improving the labeling quality and accuracy of the entire data set. This is of great significance for the subsequent use of these data to train the bone tissue morphology recognition neural network and perform bone tissue compressive strength test data analysis.

[0050] In some feasible implementations, the bone tissue medical image training set includes an annotated bone tissue morphology label a, and the bone tissue medical image training set has a total of B training images corresponding to the annotated bone tissue morphology label a, and any training image in the B training images is represented as training image b. Wherein, a is not greater than A, and b is not greater than B.

[0051] Then, the step S1221, according to the spatial coordinates of each training image in the bone tissue medical image training set and other training images annotated with the same bone tissue morphology mark in the bone tissue morphology feature space, determines the first importance factor of each training image in the bone tissue medical image training set, including: Step S12211, determining the spatial difference between the morphological characterization information of the training image b and the morphological characterization information of other B-1 training images, wherein the spatial difference is used to indicate the distribution of two training images in the bone tissue morphological feature space.

[0052] In step S12211, the detection system determines the spatial difference between the morphological representation information of the training image b and the other B-1 training images having the same annotated bone tissue morphological markers.

[0053] In the scenario of bone tissue compressive strength test data analysis, each bone tissue medical image will be feature extracted through a deep learning neural network (such as a convolutional neural network CNN) to obtain a high-dimensional feature vector, which represents the morphological representation information of the bone tissue in the image. This vector contains multiple features such as the texture, shape, and density of the bone tissue in the image, and is the basis for subsequent analysis. Then, the detection system calculates the spatial difference between the morphological representation information of the training image b and the other B-1 images with the same annotated bone tissue morphological markers. The spatial difference here actually refers to the distance between these feature vectors in the bone tissue morphological feature space. This distance can be calculated in a variety of ways, such as Euclidean distance or cosine distance. Euclidean distance measures the straight-line distance between two points, while cosine distance measures the cosine value of the angle between two points, which reflects more directional similarity.

[0054] Taking Euclidean distance as an example, assuming that the morphological representation information of the training image b is an n-dimensional vector , and the morphological representation information of another training image is a same n-dimensional vector , then the Euclidean distance between them for: , The detection system performs such calculations on the training image b and the other B-1 images one by one, thereby obtaining a set of B-1 spatial differences. Each value in this set reflects the distribution of the training image b and the corresponding image in the bone tissue morphology feature space. If the spatial difference between the two images is small, it means that their positions in the feature space are close, which implies that they have a high similarity in bone tissue morphology. Through such calculations, the detection system can provide important reference information for subsequent steps, help determine the similarity relationship between training images, and then optimize the annotation of bone tissue morphology markers and improve the accuracy of the bone tissue morphology recognition neural network.

[0055] Step S12212: The smallest spatial difference among the determined spatial differences is used as the second importance factor of the training image b, and the second importance factor is inversely correlated with the smallest spatial difference.

[0056] In step S12212, the detection system analyzes the various spatial differences calculated in step S12211 to determine the second importance factor of the training image b. This step is an important step in evaluating the similarity between the training image b and its similar images (i.e., images with the same annotated bone tissue morphology markers), and plays a key role in the subsequent optimization of the bone tissue morphology recognition neural network.

[0057] In step S12211, the detection system has calculated the spatial difference between the training image b and the other B-1 images in the training set with the same annotated bone tissue morphology markers. These differences are measured by feature distances (such as Euclidean distance or cosine distance), reflecting the distribution of images in the bone tissue morphology feature space. Then, in step S12212, the detection system finds the minimum value from these spatial differences. This minimum value represents the minimum distance between the training image b and a specific image in the feature space, which means that the two images have the highest similarity in bone tissue morphology. Since it is hoped that the importance factor can reflect the similarity between images, the minimum spatial difference is given a special meaning here. In order to convert the minimum spatial difference into the second importance factor, the detection system adopts a reverse correlation strategy. Specifically, since the smaller the spatial difference, the higher the similarity, the second importance factor is negatively correlated with the minimum spatial difference. This means that when the minimum spatial difference is small, the second importance factor will be large, and vice versa. To understand this relationship more intuitively, we can assume that the spatial difference between training image b and another image is 0.1 (assuming that the smaller the distance value, the more similar it is), which is the minimum value of all calculated spatial differences. Then, according to the principle of reverse association, the second importance factor of training image b will be a relatively large value, such as 0.9 (assuming that the importance factor ranges from 0 to 1). This large value reflects the high similarity between training image b and its most similar image, thus emphasizing its importance in the process of optimizing the annotation of bone tissue morphology labels. In this way, the detection system can accurately evaluate the similarity level of each training image among its peers and provide valuable reference information for subsequent steps. This not only helps to improve the accuracy of the bone tissue morphology recognition neural network, but also optimizes the data quality of the entire training set, providing a more reliable basis for the analysis of bone tissue compressive strength detection data.

[0058] Step S12213, normalize the second importance factor of the training image b by the second importance factor of each of the B training images to obtain the first importance factor of the training image b; wherein the basic reference representation vector of the annotated bone tissue morphological marker a is the result of weighted adjustment of the morphological representation information of the B training images based on the first importance factors of the B training images.

[0059] In step S12213, the detection system normalizes the second importance factors of B training images with the same annotated bone tissue morphology marker a to obtain the first importance factor of each training image. This step is the key to ensure the comparability and consistency of the importance factors, and provides a basis for the subsequent generation of the basic reference representation vector of the annotated bone tissue morphology marker a.

[0060] In step S12212, the detection system has calculated the second importance factor for each training image, which is based on the minimum spatial difference between the image and its similar images in the bone tissue morphological feature space. The second importance factor directly reflects the similarity between the image and its similar images, but because the second importance factors of different images may be at different orders of magnitude, normalization processing is required. Normalization processing is also called data normalization. In step S12213, the detection system divides the second importance factor of each training image by the sum of the second importance factors of all B images to obtain the first importance factor of each image. This processing ensures that the sum of the first importance factors of all training images is 1, so that these factors have a unified scale, which is convenient for subsequent comparison and analysis.

[0061] To illustrate with a specific example, suppose there are three training images (image 1, image 2, and image 3) with the same annotated bone tissue morphological marker a. Their second importance factors are 0.3, 0.6, and 0.1, respectively. After normalization, the first importance factor of image 1 is 0.3 / (0.3+0.6+0.1)=0.3, the first importance factor of image 2 is 0.6 / (0.3+0.6+0.1)=0.6, and the first importance factor of image 3 is 0.1 / (0.3+0.6+0.1)=0.1. In this way, the first importance factor of each image reflects its importance in the image class. Finally, the basic reference representation vector of the annotated bone tissue morphological marker a is the result of weighted adjustment of the morphological representation information of B training images according to these first importance factors. Specifically, the morphological representation information of each training image is multiplied by its corresponding first importance factor, and then all the weighted morphological representation information is added together, and the result is the basic reference representation vector of the annotated bone tissue morphological marker a. This vector represents the representative features of this type of bone tissue morphology in the feature space and is an important reference for subsequent training and prediction of the neural network for bone tissue morphology recognition.

[0062] In some feasible implementation schemes, the label transfer mechanism includes a label diffusion table (a table, i.e., a matrix) and a morphological label table. Then, the step S1224, diffuses the reference bone tissue morphological labels of each training image in the bone tissue medical image training set according to the label transfer mechanism to obtain the adjustment label, including: Step S12241, optimizing the labeled diffusion table according to the morphological characterization information of each training image in the bone tissue medical image training set; the table element at position (x, y) in the optimized labeled diffusion table is determined by the spatial difference between the morphological characterization information of the training image x and the morphological characterization information of the training image y; the training image x represents a random training image in the bone tissue medical image training set, and the training image y represents a random training image in the bone tissue medical image training set, and both x and y are greater than or equal to 1.

[0063] In step S12241, the detection system uses the morphological representation information of each training image in the bone tissue medical image training set to optimize (or update) the label diffusion number table. This step is a key link in the label transfer mechanism. It constructs and updates a matrix that can reflect the probability of label transfer between images based on the similarity between images. In the analysis of bone tissue compressive strength detection data, each bone tissue medical image will be feature extracted through a deep learning neural network (such as a convolutional neural network CNN) to obtain a high-dimensional feature vector, which represents the morphological representation information of the bone tissue in the image. This vector captures multiple features of the bone tissue in the image, such as texture, shape, density, etc., and is the basis for subsequent analysis.

[0064] In step S12241, the detection system extracts the morphological representation information of all images in the training set. Then, it calculates the spatial difference between each pair of images, which actually measures the distance between the images in the bone tissue morphological feature space. This distance can be calculated in many ways, such as Euclidean distance or cosine distance. Here, it is assumed that Euclidean distance is used as the measurement standard.

[0065] Taking two training images x and y as an example, the detection system extracts their morphological representation information, represented as vector (V_x) and vector (V_y) respectively. Then, it calculates the Euclidean distance between the two vectors, that is: , Where n is the dimension of the vector, and are the i-th element of vectors (V_x) and (V_y) respectively. This distance value It represents the difference between images x and y in the bone tissue morphological feature space.

[0066] The detection system then fills the spatial differences between all image pairs into the label diffusion table. The label diffusion table is a matrix where the element at position (x, y) represents the label transfer probability from image x to image y. Since smaller spatial differences indicate higher similarity between images, the label transfer probability should also be larger. Therefore, the detection system will use the inverse of the spatial difference (or a value after some transformation) as the label transfer probability, for example.

[0067] Specifically, for the element at position (x, y) in the label diffusion number table, the detection system can use the following formula to calculate: , in, is a small positive number (set according to actual needs) to avoid division by zero. This formula converts the spatial difference into a probability value between 0 and 1, which represents the probability of label transfer from image x to image y.

[0068] Through this process, the detection system obtains an optimized label diffusion table, which accurately reflects the similarity relationship between all images in the training set and the probability distribution of label transfer. This table will be used to calculate the adjustment label of each image in the subsequent steps, thereby improving the label quality and accuracy of the entire training set.

[0069] Step S12242, optimizing the morphological marker table according to the reference bone tissue morphological markers of each training image in the bone tissue medical image training set; the x-th row in the optimized morphological marker table is a one-bit effective encoding of A dimensions constructed according to the reference bone tissue morphological markers of the training image x.

[0070] In step S12242, the detection system optimizes the morphological marker table according to the reference bone tissue morphological marker of each training image in the bone tissue medical image training set. The morphological marker table is a matrix used to record the relationship between the training image and different bone tissue morphological categories, and its optimization process provides basic information for subsequent marker diffusion. In the previous step, the detection system has determined the reference bone tissue morphological marker for each training image. This marker is calculated based on the similarity between the image and its similar images in the feature space, and represents the most likely bone tissue morphological category of the image. Then, the detection system optimizes the morphological marker table according to these reference bone tissue morphological markers. The morphological marker table is a matrix with A rows and N columns, where A represents the number of bone tissue morphological categories and N represents the number of images in the training set. Each row of the matrix corresponds to a bone tissue morphological category, and each column corresponds to a training image. In the process of optimizing the morphological marker table, the detection system walks through each training image in the training set. For each training image x, it will find its corresponding reference bone tissue morphological marker, assuming that this marker is the a-th bone tissue morphology. Then, the detection system sets the element in the ath column in the xth row of the morphological marker table to 1, indicating that the training image x belongs to the ath type of bone tissue morphology; at the same time, all other elements in the xth row are set to 0, indicating that the training image x does not belong to other types of bone tissue morphology. This process actually constructs an A-dimensional one-bit effective code for each training image, in which only the position corresponding to the reference bone tissue morphological marker is 1, and the rest of the positions are 0. For example, assuming that there are 100 images in the training set and there are three types of bone tissue morphology (normal bone tissue, osteoporotic bone tissue, and fracture healing bone tissue), then the morphological marker table is a matrix with 3 rows and 100 columns. For training image 1, if its reference bone tissue morphological marker is "normal bone tissue" (assuming it is the first type), then the first row and first column element of the morphological marker table will be set to 1, and the rest of the elements will be 0. For training image 2, if its reference bone tissue morphological marker is "osteoporotic bone tissue" (assuming it is the second type), then the second row and second column element of the morphological marker table will be set to 1, and the rest of the elements will be 0, and so on. Through this process, the detection system obtains an optimized morphological labeling table, which clearly records the relationship between each training image and different bone tissue morphology categories.

[0071] Step S12243, multiply the optimized label diffusion table and the optimized morphological label table, sum the temporary table obtained after the product and the morphological label table to obtain a target table; wherein the table element at position (x, a) in the target table represents that the actual bone tissue morphological label of the training image x is the support coefficient of the first bone tissue morphology u in the first bone tissue morphology of the A categories, and u is not greater than A.

[0072] In step S12243, the detection system will use two key tables, the optimized marker diffusion table and the morphological marker table, to obtain a target table through a series of calculation steps. This target table is crucial for the subsequent determination of the adjustment markers of the training images, because it provides a quantitative support coefficient for the possibility that each training image belongs to a different bone tissue morphology category. The marker diffusion table reflects the similarity between the training images by calculating the spatial difference between them, and then constructs a probability distribution of marker transfer. The morphological marker table constructs an A-dimensional one-bit effective code for each image based on the reference bone tissue morphology marker of each training image, clarifying the relationship between the image and different bone tissue morphology categories.

[0073] In step S12243, the first operation is to calculate the product of the two tables. This operation actually combines the probability of label transfer in the label diffusion table with the one-bit effective code in the morphological label table to calculate the probability of each training image receiving a specific bone tissue morphology category label from other images. Specifically, for each element (P_xy) in the label diffusion table (indicating the probability of label transfer from image x to image y) and the element (M_xa) at the corresponding position (x, a) in the morphological label table (indicating whether image x belongs to the a-th bone tissue morphology, with a value of 1 or 0), their product represents the degree of support of image y for image x belonging to the a-th bone tissue morphology.

[0074] By walking (i.e., traversing) the entire label diffusion table and the morphological label table, the detection system can obtain a temporary table, that is, a table of intermediate states, in which each element is a support value obtained based on the above-mentioned product operation. However, this temporary table only considers the label transfer relationship between some images. In order to obtain a more comprehensive and accurate support coefficient, it is necessary to perform a summation operation with the original morphological label table. The purpose of the summation operation is to add up all the element values ​​(i.e., the support values ​​from different images) of the corresponding rows of each image in the temporary table to obtain the total support of the image belonging to each bone tissue morphology category. Specifically, for the element at position (x, a) in the temporary table and the element at the corresponding position (x, a) in the morphological label table, adding them together will obtain the element value at position (x, a) in the target table. This value represents the support coefficient (such as probability or confidence) that the actual bone tissue morphology label of the training image x is the a-th bone tissue morphology. For example, suppose there are 3 images in the training set (image 1, image 2, and image 3), and there are 2 types of bone tissue morphology categories (normal bone tissue and osteoporotic bone tissue). After the product operation, the temporary number table may contain the following data: the support level of image 1 from image 2 for normal bone tissue is 0.6, and the support level from image 3 is 0.3; the initial support level of image 1 itself marked as normal bone tissue (i.e., the value in the morphology labeling number table) is 1. Adding these values, the total support coefficient of image 1 belonging to normal bone tissue is 1.9. Similarly, the total support coefficient of image 1 belonging to osteoporotic bone tissue can be calculated. Finally, the target number table will contain the support coefficients of each image belonging to different bone tissue morphology categories, which will provide important basis for the subsequent determination of regulatory markers.

[0075] Step S12244, obtaining an adjustment mark based on the target number table, wherein the adjustment mark includes the adjustment bone tissue morphology mark of each training image in the bone tissue medical image training set, and the adjustment bone tissue morphology mark of the training image x is the first bone tissue morphology corresponding to the maximum number table element in the xth row of the target number table.

[0076] In step S12244, the detection system uses the target number table generated in the previous step to determine the adjusted bone tissue morphology marker for each training image. This step is the last link of the marker transfer mechanism. By comprehensively analyzing the support coefficients in the target number table, a most likely bone tissue morphology category is assigned to each training image. As mentioned above, the target number table is a matrix with A rows and N columns, where A represents the number of bone tissue morphology categories and N represents the number of images in the training set. Each element in the target number table represents the support coefficient (used to represent the possibility, which can be expressed as probability or confidence) that the corresponding training image belongs to a certain type of bone tissue morphology. This support coefficient is calculated based on the similarity relationship between the training images and their reference bone tissue morphology markers. In step S12244, the detection system will walk each row of the target number table (i.e., the support coefficient row of each training image). For each row, the detection system will find the element with the largest value, and the column index corresponding to this largest element represents the most likely bone tissue morphology category of the training image. Specifically, if the largest element in the xth row of the target number table is located in the ath column, then the adjusted bone tissue morphology marker of the training image x is determined to be the ath type of bone tissue morphology.

[0077] Through this process, the detection system is able to assign a regulated bone tissue morphology marker to each image in the training set. This regulated marker is calculated based on the similarity relationship between images and the information of known markers, so it is more accurate and reliable than the initial annotation marker. These regulated markers will be used for subsequent neural network debugging and data analysis to improve the accuracy and reliability of bone tissue morphology recognition.

[0078] In some feasible implementations, the adjustment mark includes the adjustment bone tissue morphology mark of each training image in the bone tissue medical image training set, the first bone tissue morphology u is any one of the first bone tissue morphologies of the A categories, u is not greater than A, and the bone tissue morphology recognition neural network includes an image feature extraction component and a bone tissue morphology classification component, and the bone tissue morphology classification component includes a feature extraction branch component. Then, in the step S120, based on the adjustment mark and the image representation vector of each training image in the bone tissue medical image training set, a morphology reference representation vector of each type of first bone tissue morphology is generated, including: Step S123, using the image feature extraction component to extract features from one or more first training images to obtain image representation vectors of each first training image; the first training image is a training image in which the adjusted bone tissue morphology marker is the first bone tissue morphology u in the bone tissue medical image training set.

[0079] In step S123, the detection system uses the image feature extraction component to select and process those training images that adjust the bone tissue morphology and mark it as a specific category (first bone tissue morphology u) from the bone tissue medical image training set. These selected images are called first training images, which are key data for generating the corresponding bone tissue morphology category morphology reference representation vector. First, the image feature extraction component is an important component of the bone tissue morphology recognition neural network, which is responsible for extracting key features related to bone tissue morphology from medical images. This component is, for example, a trained deep learning neural network, such as a convolutional neural network (CNN), which can automatically learn how to extract useful information from raw image data. In step S123, the detection system identifies all training images in the training set that adjust the bone tissue morphology and mark it as the first bone tissue morphology u. These images are, for example, selected by expert annotation or preliminary classification algorithms, and they represent typical samples of the bone tissue morphology of this category. For example, if the first bone tissue morphology u represents "osteoporotic bone tissue", then these first training images are those medical images that have been labeled or identified as osteoporosis. Then, for each first training image, the detection system will input it into the image feature extraction component for processing. This component will perform convolution, pooling and other operations on the image, and gradually extract higher-level features from the original pixel data. These features may be information such as the texture, shape, and density of bone tissue, which together constitute the image representation vector. This image representation vector is a high-dimensional numerical array that contains information about all key features in the image. Each vector corresponds to a first training image, and only images belonging to the same bone tissue morphology category will be used to generate the corresponding morphological reference representation vector. Therefore, the accuracy and richness of the image representation vector are crucial to the generation of the morphological reference representation vector in subsequent steps. For example, suppose there are 100 training images with the bone tissue morphology marked as "osteoporotic bone tissue". In step S123, the detection system inputs these 100 images one by one into the image feature extraction component to obtain the image representation vector of each image. These vectors will then be used to generate the morphological reference representation vector for the category of osteoporotic bone tissue, thereby helping the neural network to more accurately identify new osteoporotic medical images.

[0080] Step S124: Use the feature extraction branch component to perform feature extraction on the image representation vector of each first training image to obtain a refined representation vector of each first training image.

[0081] In step S124, the detection system continues to deepen the processing of the bone tissue medical image data, especially after determining the image representation vector of the first training image (i.e., the training image that adjusts the bone tissue morphology to be labeled as a specific category a). This step further extracts more refined and discriminative features from these image representation vectors through the feature extraction branch component, which are called refined representation vectors. First, it is necessary to understand the role of the feature extraction branch component. In the bone tissue morphology recognition neural network, the feature extraction branch component is, for example, a carefully designed neural network layer or module that is specifically responsible for extracting information that is critical for bone tissue morphology classification from the input feature vector. This component may contain multiple convolutional layers, pooling layers, fully connected layers, etc., which are combined to form a complex nonlinear transformation to capture the deep features in the image representation vector. In step S124, the detection system passes the image representation vector of each first training image as input to the feature extraction branch component. These image representation vectors already contain key information in the original medical image, such as the texture, shape, and density of bone tissue. However, in order to further improve the accuracy of classification, more representative features need to be extracted from these representation vectors. The feature extraction branch component generates a refined representation vector for each first training image by performing a series of complex transformations and calculations on the input image representation vector. This refinement process may include a convolution operation, which uses multiple learnable convolution kernels to perform local weighted summation on the image representation vector to extract features at different spatial locations; a pooling operation, which reduces the feature dimension by downsampling the feature map while retaining the most significant feature information; and a fully connected operation, which maps the extracted features to a new feature space to better adapt to subsequent classification tasks. Finally, the feature extraction branch component outputs a refined representation vector for each first training image. These vectors, for example, have higher dimensions and more complex structures, and they capture deep features that are difficult to directly observe in the image representation vector. These deep features are crucial for distinguishing different bone tissue morphology categories because they reflect the subtle differences in bone tissue in different states.

[0082] Step S125, performing mean calculation on the refined representation vectors of the first training images to obtain a morphological reference representation vector of the first bone tissue morphology u.

[0083] In step S125, the detection system performs mean calculation on the refined representation vectors of all first training images belonging to the same bone tissue morphology category (first bone tissue morphology u), thereby generating a morphology reference representation vector of the bone tissue morphology category. This step is an important part of constructing the morphology reference representation in the bone tissue morphology recognition neural network, which helps the neural network to more accurately identify the bone tissue morphology in new bone tissue medical images. In step S124, the detection system has extracted features from the image representation vectors of each first training image through the feature extraction branch component to obtain their respective refined representation vectors. These refined representation vectors capture the deep features of the bone tissue in the image and are key information for distinguishing different bone tissue morphology categories. In step S125, the detection system first identifies all first training images belonging to the first bone tissue morphology u. These images represent typical samples of the bone tissue morphology of this category, and their refined representation vectors contain the common features of the bone tissue morphology of this category. Then, the detection system performs mean calculation on the refined representation vectors of these first training images. This process actually averages all sample features of the same category of bone tissue morphology to obtain a morphological reference characterization vector that comprehensively reflects the representative features of the bone tissue morphology of this category. This morphological reference characterization vector is a high-dimensional vector, and each element of it is the average value of the corresponding feature on all samples. For example, suppose there are 100 first training images belonging to the category of "osteoporotic bone tissue", and each image obtains a 1000-dimensional refined characterization vector through the feature extraction branch component. In step S125, the detection system calculates the mean of these 100 1000-dimensional vectors, that is, averages the values ​​on each dimension to obtain an average vector. This average vector is the morphological reference characterization vector of the "osteoporotic bone tissue" category, which represents the typical position of the bone tissue morphology of this category in the feature space.

[0084] By calculating the morphological reference representation vector, the detection system provides important reference information for subsequent bone tissue morphological recognition. When a new bone tissue medical image is input into the neural network, the neural network can calculate the similarity between the image representation vector of the image and each morphological reference representation vector, thereby determining which bone tissue morphological category the image belongs to. Since the morphological reference representation vector is calculated based on a large number of samples, it has high representativeness and stability, which helps to improve the accuracy and reliability of bone tissue morphological recognition.

[0085] In some feasible implementation schemes, the bone tissue morphology classification component further includes a classification branch component, and the classification branch component includes a morphology reference characterization vector of each type of first bone tissue morphology. Then, the step S130 uses the bone tissue morphology recognition neural network to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the image characterization vectors of each verification image in the bone tissue medical image verification set and the morphology reference characterization vector of each type of first bone tissue morphology, and obtains the first bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set, including: Step S130A: using the image feature extraction component to perform feature extraction on each verification image in the bone tissue medical image verification set to obtain an image representation vector for each verification image in the bone tissue medical image verification set.

[0086] In step S130A, the detection system uses the debugged image feature extraction component to extract features from each verification image in the bone tissue medical image verification set, thereby obtaining an image representation vector for each verification image. The image feature extraction component is, for example, a trained deep learning neural network, such as a convolutional neural network (CNN). It can automatically extract key features related to bone tissue morphology from the original medical image and encode these features into high-dimensional image representation vectors. These vectors contain key information such as the texture, shape, and density of the bone tissue in the image, and are the basis for subsequent classification.

[0087] In step S130A, the detection system first passes each verification image in the verification set as input to the image feature extraction component. The image feature extraction component then performs a series of processing and transformations on the input image to extract key features in the image. This process may include operations such as convolution, pooling, and activation functions, which together constitute the core of deep learning.

[0088] Taking the convolution operation as an example, the convolution layer in the image feature extraction component uses multiple learnable convolution kernels to perform convolution operations on the input image to generate a series of feature maps. These feature maps capture different levels of feature information in the image, from low-level edges and textures to high-level shapes and structures. Then, the pooling layer downsamples the feature map to reduce the dimension of the feature and retain the most significant feature information. Finally, after multiple layers of convolution and pooling, the image feature extraction component outputs a high-dimensional image representation vector that contains information about all key features in the input image.

[0089] Through this process, the detection system is able to generate a corresponding image representation vector for each verification image in the verification set. These vectors will be used as input in subsequent steps to calculate the similarity between the verification image and different bone tissue morphologies, and finally determine the bone tissue morphology estimation label of the verification image.

[0090] For example, suppose there is a medical image of osteoporotic bone tissue in the verification set. In step S130A, the detection system inputs this image into the image feature extraction component, and after operations such as convolution and pooling, a high-dimensional image representation vector is obtained. This vector may contain key feature information describing the morphology of osteoporotic bone tissue, such as sparse trabeculae and cortical thinning. This image representation vector will be used in subsequent steps to calculate the similarity with the "osteoporotic bone tissue" morphology reference representation vector, so as to estimate the bone tissue morphology category of this verification image.

[0091] Step S130B: using the feature extraction branch component to perform feature extraction on the image representation vector of each verification image in the bone tissue medical image verification set, to obtain a refined representation vector of each verification image in the bone tissue medical image verification set.

[0092] In step S130B, the detection system further uses the feature extraction branch component in the bone tissue morphology recognition neural network to perform deeper feature extraction on the image representation vector of each verification image in the bone tissue medical image verification set, thereby obtaining a refined representation vector for each verification image. This process is performed after the preliminary feature representation (i.e., image representation vector) of the verification image has been obtained, and is intended to extract more refined and discriminative features from these preliminary features to improve the accuracy of bone tissue morphology estimation.

[0093] The feature extraction branch component is, for example, a neural network module, which may contain multiple convolutional layers, pooling layers, and fully connected layers. These network layers can capture deeper feature information from the input image representation vector through complex nonlinear transformations. These feature information is crucial for distinguishing different bone tissue morphologies because they reflect the subtle differences in bone tissue in different states.

[0094] Specifically, in step S130B, the detection system takes the image representation vector of each verification image in the verification set as input and passes it to the feature extraction branch component. The feature extraction branch component then further processes and transforms these image representation vectors. This process may include convolution operations of multiple convolutional layers, each of which uses multiple learnable convolution kernels to perform convolution operations on the input feature map to generate new feature maps. These new feature maps capture feature information at different scales and levels in the image.

[0095] Then, the pooling layer in the feature extraction branch downsamples the feature map output by the convolutional layer to reduce the dimension of the feature and retain the most significant feature information. This process helps reduce the computational complexity of the neural network and enhances the robustness of the neural network to small changes in the input image.

[0096] Finally, after multiple layers of convolution and pooling, the feature extraction branch component will output a refined representation vector, which contains more detailed and discriminative feature information in the verification image. This feature information will be used in the subsequent bone tissue morphology classification step to help the detection system more accurately estimate the bone tissue morphology category of the verification image.

[0097] To illustrate with a specific example, suppose there is a medical image of bone tissue showing osteoporosis in the verification set. In step S130A, the detection system has obtained the image representation vector of this image through the image feature extraction component. Then, in step S130B, the detection system inputs this image representation vector into the feature extraction branch component, and after multi-layer convolution and pooling processing, obtains a refined representation vector. This refined representation vector may contain more detailed feature information describing the morphology of osteoporotic bone tissue, such as structural changes in trabeculae, decreased bone density, etc. These feature information will help the detection system to more accurately estimate the bone tissue morphology category of this verification image as "osteoporosis".

[0098] Step S130C, using the classification branch component, determines the first bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the refined representation vectors of each verification image in the bone tissue medical image verification set and the morphological reference representation vectors of each type of first bone tissue morphology.

[0099] In step S130C, the detection system uses the classification branch component in the bone tissue morphology recognition neural network to determine the bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set. This step is based on the commonality measurement results between the refined representation vector of the verification image and the morphological reference representation vector of each type of bone tissue morphology, thereby achieving accurate estimation of the bone tissue morphology of the verification image.

[0100] The classification branch component is an important part of the bone tissue morphology recognition neural network, which contains the morphological reference representation vectors of each type of bone tissue morphology. These morphological reference representation vectors are obtained by processing and calculating the training images in the previous steps, and they represent the representative features of different bone tissue morphologies in the feature space. In step S130C, the detection system takes the refined representation vector of each verification image in the verification set as input and passes it to the classification branch component. Then, the classification branch component calculates the commonality measurement result between the refined representation vector of each verification image and the morphological reference representation vector of each type of bone tissue morphology. This commonality measurement result is calculated, for example, using a similarity measurement method, which measures the degree of similarity between the verification image and different bone tissue morphologies. Specifically, for each verification image in the verification set, the detection system calculates the similarity between its refined representation vector and the morphological reference representation vectors of all categories of bone tissue morphology. These similarity values ​​will form a similarity vector, in which each element corresponds to a bone tissue morphology category. Then, the detection system finds the maximum value in this similarity vector, that is, the bone tissue morphology category that is most similar to the verification image. This category is considered to be the estimated bone tissue morphology of the verification image, that is, the first bone tissue morphology estimation label. To illustrate with a specific example, suppose there is a medical image of bone tissue showing osteoporosis in the verification set. In the previous step, the detection system has obtained the refined representation vector of this image through the feature extraction component and the feature extraction branch component. Then, in step S130C, the detection system calculates the similarity between this refined representation vector and the morphological reference representation vectors of the three types of bone tissue morphologies: "normal bone tissue", "osteoporotic bone tissue" and "bone tissue healing after fracture". Assuming that the calculation results show that the refined representation vector of this image has the highest similarity with the morphological reference representation vector of "osteoporotic bone tissue", then the detection system will mark the estimated bone tissue morphology of this verification image as "osteoporotic bone tissue".

[0101] Through this process, the detection system is able to assign an estimated bone tissue morphology marker to each validation image in the validation set. These estimated markers will be used to evaluate the performance of the neural network and provide feedback for subsequent neural network optimization and adjustment. At the same time, these estimated markers can also serve as an important basis for the analysis of bone tissue compressive strength test data, helping medical experts to more accurately assess the health of bone tissue.

[0102] In some feasible implementation schemes, the bone tissue morphology recognition neural network also includes an image feature extraction component. Then, the step S121, using the graph representation branch component, maps each training image in the bone tissue medical image training set to the bone tissue morphology feature space according to the image association between each training image in the bone tissue medical image training set, including: Step S121A, generating an undirected complete graph, wherein the undirected complete graph includes a plurality of vertices, the plurality of vertices are fully connected, and a training image in the bone tissue medical image training set corresponds to a vertex in the undirected complete graph.

[0103] In step S121A, the detection system captures the potential associations between the training images in the bone tissue medical image training set by generating an undirected complete graph. An undirected complete graph is a special graph structure in which all vertices (or nodes) are interconnected. In an embodiment of the present application, each vertex represents a training image in the bone tissue medical image training set. Since each training image has established a direct connection with all other images, this graph structure can fully reflect the associations between the images in the training set.

[0104] Specifically, when executing step S121A, the detection system first walks through the entire bone tissue medical image training set. For each image in the training set, the detection system creates a corresponding vertex in the undirected complete graph. Since the training set may contain hundreds or thousands of images, the generated undirected complete graph will have a corresponding number of vertices. Then, the detection system establishes connections between these vertices. Due to the nature of the undirected complete graph, each vertex is connected to all other vertices. These connections actually represent potential associations between the training images, although these associations may not be clear or significant in the initial stage. However, through subsequent learning and processing of algorithms such as graph neural networks, these potential associations will be gradually revealed and strengthened.

[0105] Taking a specific example, suppose that the bone tissue medical image training set contains 100 images, each of which represents a specific bone tissue morphology. In step S121A, the detection system will generate an undirected complete graph containing 100 vertices. Each vertex corresponds to a training image, and the connection between the vertices represents the potential association between these images. As subsequent steps are executed, these potential associations will be further analyzed and utilized to improve the accuracy and efficiency of bone tissue morphology recognition. Step S121A establishes a comprehensive association network for the images in the bone tissue medical image training set by generating an undirected complete graph. This network provides an important foundation for subsequent feature extraction, label adjustment and other steps, which helps to achieve more accurate bone tissue morphology recognition and compressive strength test data analysis.

[0106] Step S121B: extracting features from each vertex in the undirected complete graph through the image feature extraction component to obtain an image representation vector for each vertex.

[0107] In step S121B, the detection system uses the image feature extraction component to extract features from each vertex in the undirected complete graph (i.e., each training image in the bone tissue medical image training set) to obtain an image representation vector for each vertex. The image feature extraction component is, for example, a trained deep learning neural network, such as a convolutional neural network (CNN). This neural network has learned to extract key features related to bone tissue morphology from the original medical image. These features may include the texture, shape, density, etc. of the bone tissue, which together constitute important information describing the state of the bone tissue. In step S121B, the detection system passes each vertex in the undirected complete graph (representing a training image) as input to the image feature extraction component. For each vertex, the component performs a series of feature extraction operations. These operations include, for example, convolution, pooling, activation functions, etc., which work together to capture key features in the image.

[0108] Taking the convolution operation as an example, the convolution layer in the image feature extraction component uses multiple learnable convolution kernels to perform convolution operations on the input image to generate a series of feature maps. These feature maps contain the feature information of the image at different levels and scales. Then, the pooling layer downsamples the feature map to reduce the dimension of the feature and retain the most significant feature information. Finally, after multiple layers of convolution and pooling, the image feature extraction component outputs a high-dimensional image representation vector. The image representation vector is a numerical array that contains key feature information describing the morphology of bone tissue in the input image. For example, for a medical image of bone tissue showing osteoporosis, its image representation vector may contain numerical values ​​representing features such as sparse trabeculae and thinning of the cortex. These numerical values ​​encode the key information in the image, enabling the detection system to understand and identify different bone tissue morphologies. Through step S121B, the detection system generates a corresponding image representation vector for each training image. These vectors not only capture the feature information of the image itself, but also provide a basis for subsequent analysis of the correlation between images and morphological recognition. In subsequent steps, the detection system will use these image representation vectors to construct the bone tissue morphology feature space, and further adjust the annotated bone tissue morphology markers and optimize the neural network.

[0109] Step S121C: using the graph to represent the branch component, and mapping each vertex to morphological representation information in the bone tissue morphological feature space according to the image representation vector of each vertex, wherein one vertex corresponds to one morphological representation information.

[0110] In step S121C, the detection system uses a graph representation branch component (such as a graph neural network GNN) to map the image representation vector of each vertex in the undirected complete graph to the bone tissue morphological feature space to form the morphological representation information of each vertex. First of all, it is necessary to clarify the role of the graph representation branch component. The graph representation branch component, especially the graph neural network (GNN), is a neural network specifically designed to process graph structured data. It can capture the complex relationships between vertices in the graph (here, training images) and update the feature representation of the vertices by learning these relationships. In the scenario of bone tissue compressive strength test data analysis, GNN can effectively utilize the correlation between training images to generate more representative morphological representation information.

[0111] In step S121C, the detection system first takes the image representation vector of each vertex obtained in step S121B as input and passes it to the graph representation branch component (GNN). Then, the GNN iteratively updates the feature representation of each vertex based on the topological structure of the undirected complete graph (i.e., the connection relationship between vertices) and the features of the vertex (i.e., the image representation vector).

[0112] Specifically, each vertex in the GNN aggregates the feature information of its neighboring vertices and updates it in combination with its own feature information. This process includes two steps, for example, the aggregation function and the update function. The aggregation function is responsible for collecting the feature information of neighboring vertices, while the update function updates the feature representation of the vertex based on the aggregated information and the vertex's own feature information. Through multiple iterations, the GNN can learn the complex relationships between vertices and generate more accurate and rich morphological representation information.

[0113] To give a specific example, suppose there is a vertex in the undirected complete graph representing a medical image of bone tissue showing osteoporosis. In step S121B, the image representation vector of the image has been obtained. In step S121C, this vector is input into the GNN. The GNN will consider the correlation between the image and other similar images (i.e., neighboring vertices), and update the feature representation of the image by aggregating the feature information of these similar images and combining its own feature information. After multiple iterations, the GNN will output a new morphological representation information, which not only contains the key features of the original image, but also incorporates the correlation information with other similar images. This new morphological representation information will be used as the position representation of the image in the bone tissue morphological feature space for subsequent labeling adjustment and neural network optimization.

[0114] Through the processing of step S121C, the detection system generates a unique morphological representation information for each training image. This morphological representation information not only captures the characteristic information of the image itself, but also incorporates the correlation information between images, providing strong support for subsequent bone tissue morphological recognition and analysis.

[0115] In some feasible implementation schemes, the bone tissue morphology recognition neural network includes a label adjustment network component and a bone tissue morphology classification component, the bone tissue morphology classification component includes a feature extraction branch component, and the label adjustment network component includes a graph representation branch component. Then, the step S140, based on the actual bone tissue morphology labels of each training image in the bone tissue medical image training set, and the actual bone tissue morphology labels of each verification image in the bone tissue medical image verification set and the first bone tissue morphology estimation label, debugs the bone tissue morphology recognition neural network, including: Step S140A, debugging the feature extraction branch component in the bone tissue morphology classification component according to the loss between the actual bone tissue morphology mark of each verification image in the bone tissue medical image verification set and the first bone tissue morphology estimated mark, to obtain a debugged feature extraction branch component.

[0116] In step S140A, the detection system debugs the feature extraction branch component, one of the key components in the bone tissue morphology recognition neural network, to ensure that it can more accurately extract key features related to bone tissue morphology from medical images. This process is based on the comparison between the actual bone tissue morphology markers of the validation image in the validation set and the first bone tissue morphology estimated markers estimated by the neural network, and aims to optimize the performance of the feature extraction branch component by reducing the difference (i.e., loss) between the two.

[0117] Specifically, each validation image in the validation set has an actual bone tissue morphology label annotated by a medical expert, which represents the true morphological category of the bone tissue in the image. At the same time, after the preliminary estimation of the bone tissue morphology recognition neural network, each validation image will also get a first bone tissue morphology estimation label. This estimation label is the prediction result given by the neural network based on its learned feature extraction and classification capabilities.

[0118] In step S140A, the detection system calculates the difference between the actual bone tissue morphology marker of each verification image and the first bone tissue morphology estimated marker. This difference is measured, for example, by a loss function, which can quantify the degree of inconsistency between the predicted result and the actual result. Commonly used loss functions include cross entropy loss, mean square error loss, etc., which are selected according to specific task requirements and data characteristics.

[0119] Once the loss values ​​for each validation image are calculated, the detection system uses these loss values ​​to debug the feature extraction branch component. This process is achieved, for example, by backpropagation algorithms and gradient descent methods. Specifically, the detection system calculates the gradient of each parameter in the feature extraction branch component (i.e., the derivative of the loss value with respect to the parameter) based on the loss value, and then updates the parameter value based on these gradients. Through multiple iterations of this process, the parameters of the feature extraction branch component are gradually adjusted to the optimal state, making the neural network's estimated results closer to the actual results. For example, suppose there is a medical image showing osteoporosis in the validation set, and its actual bone tissue morphology is marked as "osteoporosis". However, after the initial estimation of the neural network, the first bone tissue morphology of the image is estimated to be marked as "normal bone tissue". In this case, the detection system calculates a large loss value, indicating that there is a large difference between the estimated result and the actual result. Then, the detection system will debug the parameters of the feature extraction branch component based on this loss value, so that the features of osteoporosis can be more accurately identified in the next estimation. After multiple iterations of debugging, the performance of the feature extraction branch component will be significantly improved, thereby improving the accuracy of the entire bone tissue morphology recognition neural network.

[0120] Step S140B, using a reference estimation component to perform bone tissue morphology estimation on each verification image in the bone tissue medical image verification set, and obtaining a second bone tissue morphology estimation label for each verification image in the bone tissue medical image verification set; the reference estimation component includes a first classification branch component and a debugged feature extraction branch component, the first classification branch component contains a target reference representation vector for each type of first bone tissue morphology, and the target reference representation vector is generated based on the image representation vector of the training image containing the same actual bone tissue morphology label in the bone tissue medical image training set.

[0121] In step S140B, the detection system uses a reference estimation component to perform a second bone tissue morphology estimation on each verification image in the bone tissue medical image verification set to obtain a more accurate and reliable second bone tissue morphology estimation label. The reference estimation component includes a first classification branch component and a debugged feature extraction branch component, which work together to improve the accuracy of the estimation.

[0122] The first classification branch component in the reference estimation component contains the target reference representation vectors for each class of first bone tissue morphology. These target reference representation vectors are generated based on the image representation vectors of training images with the same actual bone tissue morphology labels in the bone tissue medical imaging training set. Specifically, the detection system first selects all training images belonging to the same bone tissue morphology category from the training set, then extracts the image representation vectors of these images, and merges these vectors into a target reference representation vector in some way (such as averaging, weighting, etc.). This vector represents the representative features of this type of bone tissue morphology in the feature space.

[0123] In step S140B, the detection system first uses the debugged feature extraction branch component to extract features from each verification image in the verification set to obtain an image representation vector for each verification image. This debugged feature extraction branch component has been optimized in step S140A and can more accurately extract key features related to bone tissue morphology from the image. Then, the detection system compares the image representation vectors of these verification images with the target reference representation vector in the first classification branch component. Specifically, the detection system calculates the similarity or distance between the image representation vector of each verification image and the target reference representation vector of each type of bone tissue morphology. This similarity or distance reflects the degree of association between the verification image and different bone tissue morphology categories. Based on these similarity or distance values, the detection system can assign a second bone tissue morphology estimation mark to each verification image. For example, the detection system selects the bone tissue morphology category corresponding to the target reference representation vector with the highest similarity or the smallest distance as the estimation mark. This estimation mark represents the most likely judgment of the detection system on the bone tissue morphology in the verification image.

[0124] For example, suppose there is a medical image showing osteoporosis in the verification set, and its image representation vector is obtained after being processed by the feature extraction branch component. Then, the detection system will compare this vector with the target reference representation vectors of three types of bone tissue morphology: "normal bone tissue", "osteoporotic bone tissue" and "bone tissue healed after fracture". If the calculation result shows that the image has the highest similarity with the target reference representation vector of "osteoporotic bone tissue", then the detection system will determine the second bone tissue morphology estimation marker of this verification image as "osteoporotic bone tissue". Through this process, the detection system can generate a more accurate and reliable second bone tissue morphology estimation marker for each verification image in the verification set. These estimation markers will be used in the subsequent step S140C to debug the representation branch component and further optimize the performance of the entire bone tissue morphology recognition neural network.

[0125] Step S140C, debugging the graph representation branch component according to the loss between the actual bone tissue morphology mark of each verification image in the bone tissue medical image verification set and the second bone tissue morphology estimated mark, to obtain a debugged graph representation branch component.

[0126] In step S140C, the detection system uses the loss between the actual bone tissue morphology label of each verification image in the bone tissue medical image verification set and the second bone tissue morphology estimated label obtained by referring to the estimated component to debug the graph representation branch component to optimize its mapping and association capabilities in the bone tissue morphology feature space. This process is intended to further improve the accuracy of the neural network in identifying bone tissue morphology.

[0127] The graph representation branch component, for example, is a graph neural network (GNN) neural network, which is responsible for processing the association information between images and mapping these images into the bone tissue morphology feature space. In this space, similar images are mapped to similar positions, thereby preserving the association relationship between the images. In step S140C, the detection system calculates the difference between the actual bone tissue morphology marker and the second bone tissue morphology estimated marker of each verification image in the verification set, and this difference is measured by a loss function. The loss function is selected according to the specific task requirements and data characteristics, such as cross entropy loss or mean square error loss. This loss value reflects the accuracy of the neural network when using the association information between images to estimate the bone tissue morphology. Then, the detection system uses this loss value to debug the graph representation branch component. This process is implemented, for example, by a back propagation algorithm and a gradient descent method. Specifically, the detection system calculates the gradient of each parameter in the graph representation branch component according to the loss value, and then updates the value of the parameter according to these gradients. Through multiple iterations of this process, the parameters of the graph representation branch component will gradually be adjusted to the optimal state, so that the neural network can better process the association information between images and improve the accuracy of bone tissue morphology estimation.

[0128] For example, suppose there is a medical image showing osteoporosis in the validation set, and its actual bone tissue morphology is marked as "osteoporosis". By referring to the estimated component, the neural network gives a second bone tissue morphology estimated label, but if there is a large difference between this estimated label and the actual label (that is, the loss value is large), then the detection system will debug the graph representation branch component based on this difference. Specifically, the detection system will calculate the gradient of the relevant parameters in the graph representation branch component according to the loss value, and update the parameter values ​​according to these gradients, so that the characteristics of osteoporosis can be more accurately identified in the next estimation. Through the debugging process of step S140C, the detection system can optimize the performance of the graph representation branch component, making it more accurate and effective in processing the correlation information between images and generating morphological representation information. This will help to improve the performance of the entire bone tissue morphology recognition neural network and provide more reliable support for subsequent bone tissue compressive strength test data analysis.

[0129] In some feasible implementation schemes, the first bone tissue morphology u is any one of the A categories of first bone tissue morphologies, u is not greater than A, and the bone tissue morphology recognition neural network further includes an image feature extraction component, then the method further includes: Step SI, using the image feature extraction component to extract features from one or more second training images to obtain image representation vectors of each second training image; the second training image is a training image in which the actual bone tissue morphology label in the bone tissue medical image training set is the first bone tissue morphology u.

[0130] In step SI, the detection system uses the image feature extraction component to perform in-depth processing on a portion of specific images in the bone tissue medical image training set in response to the specific needs of bone tissue compressive strength test data analysis. These specific images are the second training images, which are images in which the actual bone tissue morphology in the training set is marked as a specific category (first bone tissue morphology u). Through this step, the detection system aims to extract key feature information from these images to provide a basis for subsequent analysis and recognition. First, the image feature extraction component is an important part of the bone tissue morphology recognition neural network, which is, for example, a trained deep learning neural network, such as a convolutional neural network (CNN). This neural network has learned to extract key features related to bone tissue morphology from medical images, such as texture, shape, density, etc. These features are crucial for distinguishing different bone tissue morphologies.

[0131] In step SI, the detection system first identifies all the second training images in the training set that are labeled as the first bone tissue morphology u. For example, if the first bone tissue morphology u represents "osteoporosis", then the second training image is all the medical images in the training set that are labeled as "osteoporosis" by experts. Then, the detection system inputs these second training images into the image feature extraction component one by one. For each input image, the feature extraction component performs a series of complex calculations and operations, including convolution, pooling, activation functions, etc., to extract the key features in the image. These features are encoded into high-dimensional image representation vectors, each of which contains rich information about the bone tissue morphology in the image.

[0132] For example, suppose a medical image showing osteoporosis is input as the second training image into the feature extraction component. The component will first convolve the image through multiple convolutional layers to extract feature maps at different levels. Then, the feature map is downsampled through the pooling layer to reduce the feature dimension and retain the most significant feature information. Finally, after processing by the fully connected layer, the component will output a high-dimensional image representation vector, which represents the representation of this osteoporosis medical image in the feature space.

[0133] Through the processing of step SI, the detection system can generate a corresponding image representation vector for each second training image. These vectors not only capture the key feature information of bone tissue in the image, but also provide a basis for the generation of subsequent target reference representation vectors. These target reference representation vectors will play an important role in bone tissue morphology recognition, helping the detection system to more accurately identify bone tissue morphology in new medical images.

[0134] Step SII: Use the debugged feature extraction branch component to perform feature extraction on the image representation vector of each second training image to obtain a refined representation vector of each second training image.

[0135] In step SII, the detection system performs a deeper feature extraction on the image representation vectors of the second training images obtained in step SI to obtain a refined representation vector for each second training image. This step utilizes the debugged feature extraction branch component that was optimized and improved for a specific task in the previous step.

[0136] In the bone tissue morphology recognition neural network, the feature extraction branch component is responsible for extracting more representative and discriminative features from the input image representation vector. These features are crucial for subsequent bone tissue morphology classification because they can more accurately reflect the differences between different bone tissue morphologies.

[0137] In step SII, the detection system takes the image representation vector of each second training image obtained in step SI as input and passes it to the debugged feature extraction branch component. This component has been debugged and optimized based on the previous training data and verification data, so it has stronger feature extraction capabilities.

[0138] Specifically, the debugged feature extraction branch component performs a series of complex nonlinear transformations on each input image representation vector. These transformations may include multi-layer convolution operations, pooling operations, and fully connected layers. Through these transformations, the component can extract more subtle and complex feature information from the image representation vector and encode this information into a new refined representation vector.

[0139] For a specific example, suppose there is a second training image, which is a medical image showing osteoporosis. In step SI, the image representation vector of this image has been obtained through the image feature extraction component. Now, in step SII, this image representation vector is input into the debugged feature extraction branch component. The component will further process and transform this vector to extract more refined features related to osteoporosis. These features may include the sparseness of trabeculae, the reduction of bone density, etc., which together constitute the refined representation vector of this image.

[0140] Through the processing of step SII, the detection system generates a refined representation vector for each second training image. These vectors not only contain the key information in the image representation vector, but also extract more discriminative features through further processing by the feature extraction branch component. These refined representation vectors will be used to generate target reference representation vectors in subsequent steps, thereby helping the detection system to more accurately identify bone tissue morphology in new medical images.

[0141] Step SIII: performing mean calculation on the refined representation vectors of each second training image to obtain a target reference representation vector of the first bone tissue morphology u.

[0142] In step SIII, the detection system performs mean calculation on the refined representation vectors of each second training image obtained by processing the debugged feature extraction branch component, thereby obtaining the target reference representation vector of the first bone tissue morphology u. The refined representation vectors are extracted from the image representation vectors of the second training image by the debugged feature extraction branch component, and they contain key feature information closely related to the specific bone tissue morphology (first bone tissue morphology u). In step SIII, the detection system first collects the refined representation vectors of all second training images belonging to the first bone tissue morphology u. These vectors represent the characteristic performance of the same type of bone tissue morphology in different samples. Since each sample may have certain differences, the specific values ​​of these vectors will also be different. Then, the detection system performs mean calculation on these refined representation vectors. This calculation process is actually to average the corresponding elements of all vectors, so as to obtain a vector that comprehensively reflects the representative features of the first bone tissue morphology u, that is, the target reference representation vector. The position of this vector in the feature space represents the representative features of the first bone tissue morphology u, which will become an important basis for subsequent bone tissue morphology recognition. For example, suppose there are 100 second training images of the bone tissue morphology "osteoporosis" (i.e., the first bone tissue morphology u). After feature extraction and debugging, the feature extraction branch component processes the 100 images to obtain the refined representation vectors. These vectors may be distributed in a relatively concentrated area in the feature space, but due to differences between samples, their specific positions may also be different. Through mean calculation, an average vector can be obtained, which represents the typical position of the bone tissue morphology "osteoporosis" in the feature space, i.e., the target reference representation vector.

[0143] After obtaining the target reference representation vector, the detection system can use it to compare with the feature vector of the new medical image in the subsequent bone tissue morphology recognition process. By calculating the similarity or distance between the feature vector of the new image and the target reference representation vector, the detection system can determine whether the new image belongs to the bone tissue morphology of "osteoporosis". This recognition method based on the target reference representation vector can more accurately reflect the representative features of bone tissue morphology, thereby improving the accuracy and reliability of recognition.

[0144] In some feasible implementation schemes, the feature extraction branch component is debugged by the first label loss under multiple training branches; the graph representation branch component is debugged by the second label loss under the multiple training branches; wherein the first training branch is any training branch among the multiple training branches; the first label loss under the first training branch is the loss between the first bone tissue morphology estimated label and the actual bone tissue morphology label of each verification image in the bone tissue medical image verification set; the second label loss under the first training branch is the loss between the second bone tissue morphology estimated label and the actual bone tissue morphology label of each verification image in the bone tissue medical image verification set.

[0145] In order to further improve the accuracy and robustness of the neural network for bone tissue morphology recognition, the detection system adopts a multi-training branch strategy to debug the feature extraction branch components and the graph representation branch components respectively. The core idea of ​​this strategy is to refine the learning process of network components through different training branches and corresponding loss functions to ensure that they can achieve optimal performance on specific tasks and data sets. In an embodiment of the present application, a training branch refers to an independent training path formed based on different data subsets or different processing flows during the neural network training process. Each training branch generates a specific set of estimated markers (for verification images) and compares them with the actual bone tissue morphology markers to calculate the loss value.

[0146] For the feature extraction branch component, it is responsible for extracting key features from medical images. In this derivative implementation, the feature extraction branch component is debugged by the first label loss under multiple training branches. Specifically, each training branch generates a set of first bone tissue morphology estimated labels for verification images, which are obtained based on a specific data subset or processing flow under the branch. Then, the detection system compares these estimated labels with the actual bone tissue morphology labels of the verification image and calculates the first label loss. This loss value reflects the performance of the feature extraction branch component under the current training branch. By summarizing the first label losses under multiple training branches, the detection system can globally optimize the feature extraction branch component to ensure that it can extract more accurate and representative feature vectors. For a specific example, assume that there are three training branches: branch 1, branch 2, and branch 3. Each branch is debugged based on different data augmentation strategies or different neural network parameter settings. In the verification stage, each branch generates a set of first bone tissue morphology estimated labels for verification images. Then, the detection system compares these estimated labels with the actual labels of the verification image and calculates the first label losses under three branches: loss 1, loss 2, and loss 3. Finally, the detection system will debug the feature extraction branch component according to the three loss values ​​to optimize its performance under different conditions. Similarly, the graph representation branch component (for example, a graph neural network) is responsible for processing the association information between images. In this derivative embodiment, the graph representation branch component is debugged by the second labeling loss under multiple training branches. The second labeling loss is calculated based on the difference between the second bone tissue morphology estimated label and the actual label of the verification image. Unlike the first labeling loss, the second estimated label is generated by the reference estimation component, which includes the debugged feature extraction branch component and the first classification branch component. Therefore, the second labeling loss can more comprehensively reflect the performance of the graph representation branch component when using the association information between images to estimate the bone tissue morphology. By debugging the graph representation branch component to reduce the second labeling loss, the detection system can improve the recognition accuracy of the neural network in complex scenarios. The feature extraction branch component and the graph representation branch component are debugged separately through the strategy of multiple training branches and the corresponding loss function. This strategy helps to ensure that the neural network can achieve optimal performance on specific tasks and data sets, thereby improving the accuracy and robustness of bone tissue morphology recognition.

[0147] In some implementation schemes that can be implemented separately, in step S110, obtaining the first training branch may include: obtaining a debugging image set, the debugging image set including Z categories of bone tissue morphology, each category of bone tissue morphology including multiple debugging images, Z≥1; randomly selecting A categories of bone tissue morphology from the Z categories of bone tissue morphology contained in the debugging image set as the first bone tissue morphology of the A categories; randomly selecting S+T debugging images of each category of first bone tissue morphology from the multiple debugging images of each category of first bone tissue morphology; forming the first training branch based on the A categories of first bone tissue morphology and the S+T debugging images of each category of first bone tissue morphology.

[0148] In this embodiment, the detection system first obtains a debugging image set. This set contains Z different categories of bone tissue morphology, and each category has multiple debugging images. These images are samples of bone tissue medical images and are used for subsequent training and verification of neural networks. For example, the debugging image set may contain multiple bone tissue morphology categories such as normal bone tissue, osteoporosis, and fracture healing, and each category has dozens to hundreds of debugging images. After obtaining the debugging image set, the detection system selects A categories from these Z categories as the first bone tissue morphology for subsequent training. This screening process can be random or based on a certain strategy (such as based on the importance of the category or the number of samples). For example, if the debugging image set contains 10 bone tissue morphology categories, but the current task only focuses on 5 of them (such as normal bone tissue, mild osteoporosis, moderate osteoporosis, severe osteoporosis, and fracture healing), then the system will randomly or selectively select these 5 categories from these 10 categories as the first bone tissue morphology. After determining the first bone tissue morphology, the detection system selects enough debugging images for each type of morphology. These images will be used for training and validation of the neural network. Specifically, the detection system will randomly select S+T images from multiple debugging images under each category of the first bone tissue morphology. Among them, S images will be used for the training set and T images will be used for the validation set. For example, for the category of "mild osteoporosis", the detection system may select 50 images as the training set (S=50) and 10 images as the validation set (T=10), for a total of 60 images. Finally, the detection system forms the first training branch based on the first bone tissue morphology of the A categories screened out and the S+T debugging images under each category. This training branch contains all the data and label information required for neural network training. Specifically, each category of bone tissue morphology will have a corresponding label (such as "normal bone tissue", "mild osteoporosis", etc.), and each debugging image will be assigned a corresponding category label. These label information will be used to supervise the training process of the neural network to ensure that the network can learn the ability to extract bone tissue morphology features from medical images. Through the above steps, the detection system prepares the first training branch for the training of the neural network for bone tissue morphology recognition. This training branch contains diverse bone tissue morphology categories and sufficient debugging image samples, providing a solid foundation for neural network training.

[0149] Furthermore, in order to increase the generalization ability of the network, the method may also include: obtaining an extended image set, the extended image set including L categories of second bone tissue morphology, each category of second bone tissue morphology including multiple extended debugging images, each extended debugging image including its own annotated bone tissue morphology mark, the annotated bone tissue morphology mark of the extended debugging image representing the annotated second bone tissue morphology of the extended debugging image; the L categories of second bone tissue morphology do not completely match the Z category bone tissue morphology, L≥1; using the bone tissue morphology recognition neural network after debugging, and the annotated bone tissue morphology mark of each extended debugging image, to determine the morphological reference representation vector of each category of second bone tissue morphology; and constructing a classification branch component in the bone tissue morphology recognition neural network after debugging.

[0150] On the basis of step S110, in order to further enhance the generalization ability of the bone tissue morphology recognition neural network, that is, to improve its recognition ability of unknown or unseen bone tissue morphology, additional steps are introduced to expand the training data and update the network structure.

[0151] In the above derivative scheme, the detection system obtains an extended image set, which contains L categories of second bone tissue morphology. These second bone tissue morphology do not completely match the Z categories in the original debugging image set, which means that they may contain some new and unseen bone tissue morphology categories. Each category contains multiple extended debugging images, and each extended debugging image is annotated with bone tissue morphology tags, which indicate the bone tissue morphology category corresponding to the image. For example, assume that the original debugging image set contains three categories: normal bone tissue, osteoporosis, and post-fracture healing. In order to enhance the network's recognition ability for unknown bone tissue morphology, an extended image set is obtained, which contains new bone tissue morphology categories such as bone tumors and bone infections. There are dozens of extended debugging images with annotated tags under each category. After obtaining the extended image set, the detection system uses the already debugged bone tissue morphology recognition neural network (i.e., the network trained and optimized in the previous steps) to process the extended debugging images. Specifically, the detection system inputs each extended debugging image into the neural network, and obtains the image representation vector of each image through the feature extraction and classification components of the network. Then, based on these image representation vectors and the annotated bone tissue morphology marks of the images, the detection system can calculate the morphological reference representation vector of each type of second bone tissue morphology. This vector represents the typical features of this type of bone tissue morphology in the feature space. For example, for the category of "bone tumor" in the extended image set, the detection system will input all the extended debugging images under this category into the neural network to obtain the image representation vector of each image. Then, the detection system will calculate the morphological reference representation vector of the category of "bone tumor" based on these vectors and the annotated marks of the images. This vector will serve as an important reference for the subsequent identification of new bone tumor medical images. After determining the morphological reference representation vector of each type of second bone tissue morphology, the detection system will construct new classification branch components in the already debugged bone tissue morphology recognition neural network. These classification branch components correspond to new bone tissue morphology categories, and they contain new morphological reference representation vectors and related classification logic. By integrating these new classification branch components into the original neural network, the detection system can expand the network's recognition ability for unknown bone tissue morphology. For example, after determining the morphological reference representation vectors for categories such as "bone tumor" and "bone infection", the detection system will build new classification branch components in the original neural network. These branch components contain morphological reference representation vectors and classification logic for new categories. When new medical images are input into the network, the detection system can not only identify the original categories (such as normal bone tissue, osteoporosis, etc.), but also these new bone tissue morphological categories.

[0152] The present application also provides a detection system, which includes one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors. When the one or more programs are executed by the processors, the above bone tissue compressive strength detection data analysis method is implemented. Figure 2 As shown, the detection system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the detection system 100 may further include a transceiver 104. The structure of the detection system 100 does not constitute a limitation on the embodiments of the present application.

Claims

1. A method for analyzing bone tissue compressive strength detection data, characterized in that: include: Obtain a first training branch, wherein the first training branch includes a bone tissue medical image training set and a bone tissue medical image verification set, wherein the bone tissue medical image training set includes A categories of first bone tissue morphologies, each category of the first bone tissue morphology includes S training images, each training image includes its own actual bone tissue morphology label and annotated bone tissue morphology label, the actual bone tissue morphology label of the training image represents the first bone tissue morphology actually corresponding to the training image, and the annotated bone tissue morphology label of the training image represents the annotated first bone tissue morphology of the training image; The bone tissue medical image verification set includes the A categories of first bone tissue morphology, each category of the first bone tissue morphology includes T verification images, each verification image includes its own actual bone tissue morphology mark, and the actual bone tissue morphology mark of the verification image represents the first bone tissue morphology actually corresponding to the verification image; wherein A≥1; T≥1; S≥1; Using a bone tissue morphology recognition neural network, adjusting the annotated bone tissue morphology mark of each training image in the bone tissue medical image training set according to the image association between each training image in the bone tissue medical image training set to obtain an adjustment mark, and generating a morphology reference representation vector of each type of first bone tissue morphology according to the adjustment mark and the image representation vector of each training image in the bone tissue medical image training set; The bone tissue morphology recognition neural network is used to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the image representation vectors of each verification image in the bone tissue medical image verification set and the morphology reference representation vectors of each type of first bone tissue morphology, so as to obtain the first bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set; Debugging the bone tissue morphology recognition neural network according to the actual bone tissue morphology mark of each training image in the bone tissue medical image training set, and the actual bone tissue morphology mark and the first bone tissue morphology estimation mark of each verification image in the bone tissue medical image verification set; The debugged bone tissue morphology recognition neural network is used to identify the bone tissue morphology corresponding to the target bone tissue medical image.

2. The method according to claim 1, characterized in that The bone tissue morphology recognition neural network includes a label regulation network component, and the label regulation network component includes a graph representation branch component and an iterative regulation branch component; The bone tissue morphology recognition neural network is used to adjust the annotated bone tissue morphology mark of each training image in the bone tissue medical image training set according to the image association between each training image in the bone tissue medical image training set to obtain the adjustment mark, including: Using the graph representation branch component, mapping each training image in the bone tissue medical image training set to a bone tissue morphological feature space according to image associations between each training image in the bone tissue medical image training set; The iterative adjustment branch component is used to adjust the annotated bone tissue morphological marks of each training image in the bone tissue medical image training set according to the spatial coordinates of each training image in the bone tissue morphological feature space to obtain the adjustment marks.

3. The method according to claim 2, characterized in that Each training image in the bone tissue medical image training set includes respective morphological representation information in the bone tissue morphological feature space; The iterative adjustment branch component is used to adjust the annotated bone tissue morphology mark of each training image in the bone tissue medical image training set according to the spatial coordinates of each training image in the bone tissue morphology feature space to obtain the adjustment mark, including: Determine the first importance factor of each training image in the bone tissue medical image training set according to the spatial coordinates of each training image in the bone tissue medical image training set and other training images annotated with the same bone tissue morphological markers in the bone tissue morphological feature space, wherein the first importance factor of the training image represents the adaptation support coefficient between the annotated bone tissue morphological markers and the actual bone tissue morphological markers contained in the training image; Generate a basic reference representation vector for each type of the annotated bone tissue morphological markers according to the first importance factor and morphological representation information of all training images containing the same annotated bone tissue morphological markers; Determine the reference bone tissue morphological markers of each training image in the bone tissue medical image training set according to the commonality measurement results between the morphological representation information of each training image in the bone tissue medical image training set and the basic reference representation vector of each type of the annotated bone tissue morphological markers; The reference bone tissue morphology mark of each training image in the bone tissue medical image training set is diffused according to the mark transfer mechanism to obtain the adjustment mark.

4. The method according to claim 3, characterized in that The bone tissue medical image training set includes an annotated bone tissue morphology mark a, and the bone tissue medical image training set has a total of B training images corresponding to the annotated bone tissue morphology mark a, and any training image in the B training images is represented as a training image b; wherein a is not greater than A, and b is not greater than B; The first importance factor of each training image in the bone tissue medical image training set is determined based on the spatial coordinates of each training image in the bone tissue medical image training set and other training images annotated with the same bone tissue morphology mark in the bone tissue morphology feature space, including: Determine the spatial difference between the morphological representation information of the training image b and the morphological representation information of other B-1 training images, wherein the spatial difference is used to indicate the distribution of the two training images in the bone tissue morphological feature space; The smallest spatial difference among the determined spatial differences is used as the second importance factor of the training image b, and the second importance factor is inversely correlated with the smallest spatial difference; The second importance factor of the training image b is normalized by the second importance factors of each of the B training images to obtain the first importance factor of the training image b; wherein the basic reference representation vector of the annotated bone tissue morphological marker a is the result of weighted adjustment of the morphological representation information of the B training images according to the first importance factors of the B training images; The label transfer mechanism includes a label diffusion number table and a morphological label number table; the label transfer mechanism is used to diffuse the reference bone tissue morphological labels of each training image in the bone tissue medical image training set to obtain the adjustment label, including: According to the morphological representation information of each training image in the bone tissue medical image training set, the labeled diffusion number table is optimized; the number table element at the position (x, y) in the optimized labeled diffusion number table is determined by the spatial difference between the morphological representation information of the training image x and the morphological representation information of the training image y; the training image x represents a random training image in the bone tissue medical image training set, the training image y represents a random training image in the bone tissue medical image training set, and x and y are both greater than or equal to 1; According to the reference bone tissue morphological markers of each training image in the bone tissue medical image training set, the morphological marker number table is optimized; the xth row in the optimized morphological marker number table is a one-bit effective code of A dimensions constructed according to the reference bone tissue morphological markers of the training image x; The optimized label diffusion number table and the optimized morphological label number table are multiplied, and the temporary number table obtained after the multiplication is summed with the morphological label number table to obtain a target number table; wherein the number table element at position (x, a) in the target number table represents that the actual bone tissue morphological label of the training image x is the support coefficient of the first bone tissue morphology u among the first bone tissue morphologies of the A categories, and u is not greater than A; An adjustment mark is obtained based on the target number table, and the adjustment mark includes the adjustment bone tissue morphology mark of each training image in the bone tissue medical image training set, and the adjustment bone tissue morphology mark of the training image x is the first bone tissue morphology corresponding to the maximum number table element in the xth row of the target number table.

5. The method according to claim 1, characterized in that The adjustment mark includes the adjustment bone tissue morphology mark of each training image in the bone tissue medical image training set; the first bone tissue morphology u is any one of the first bone tissue morphologies of the A categories, and u is not greater than A; the bone tissue morphology recognition neural network includes an image feature extraction component and a bone tissue morphology classification component, and the bone tissue morphology classification component includes a feature extraction branch component; The step of generating a morphological reference representation vector of each type of first bone tissue morphology based on the adjustment mark and the image representation vector of each training image in the bone tissue medical image training set comprises: The image feature extraction component is used to extract features from one or more first training images to obtain image representation vectors of each first training image; the first training image is a training image in which the adjusted bone tissue morphology marker is the first bone tissue morphology u in the bone tissue medical image training set; Using the feature extraction branch component to perform feature extraction on the image representation vector of each first training image to obtain a refined representation vector of each first training image; The refined representation vectors of the first training images are averaged to obtain a morphological reference representation vector of the first bone tissue morphology u.

6. The method according to claim 5, characterized in that The bone tissue morphology classification component further includes a classification branch component, wherein the classification branch component includes a morphology reference characterization vector of each type of first bone tissue morphology; The bone tissue morphology recognition neural network is used to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the image representation vectors of each verification image in the bone tissue medical image verification set and the morphology reference representation vectors of each type of first bone tissue morphology, to obtain the first bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set, including: Using the image feature extraction component to extract features from each verification image in the bone tissue medical image verification set, and obtaining an image representation vector of each verification image in the bone tissue medical image verification set; Using the feature extraction branch component to perform feature extraction on the image representation vector of each verification image in the bone tissue medical image verification set, to obtain a refined representation vector of each verification image in the bone tissue medical image verification set; The classification branch component is used to determine the first bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set based on the commonality measurement results between the refined representation vectors of each verification image in the bone tissue medical image verification set and the morphological reference representation vectors of each type of first bone tissue morphology.

7. The method according to claim 2, characterized in that The bone tissue morphology recognition neural network also includes an image feature extraction component; the graph representation branch component is used to map each training image in the bone tissue medical image training set to a bone tissue morphology feature space according to the image association between each training image in the bone tissue medical image training set, including: Generate an undirected complete graph, wherein the undirected complete graph includes a plurality of vertices, the plurality of vertices are fully connected, and a training image in the bone tissue medical image training set corresponds to a vertex in the undirected complete graph; Performing feature extraction on each vertex in the undirected complete graph by the image feature extraction component to obtain an image representation vector of each vertex; The graph is used to represent the branch component, and each vertex is mapped to morphological representation information in the bone tissue morphological feature space according to the image representation vector of each vertex, and one vertex corresponds to one morphological representation information.

8. The method according to claim 1, characterized in that The bone tissue morphology recognition neural network includes a label adjustment network component and a bone tissue morphology classification component, the bone tissue morphology classification component includes a feature extraction branch component, and the label adjustment network component includes a graph representation branch component; The step of debugging the bone tissue morphology recognition neural network according to the actual bone tissue morphology mark of each training image in the bone tissue medical image training set, and the actual bone tissue morphology mark of each verification image in the bone tissue medical image verification set and the first bone tissue morphology estimation mark comprises: According to the loss between the actual bone tissue morphology mark of each verification image in the bone tissue medical image verification set and the first bone tissue morphology estimated mark, debugging the feature extraction branch component in the bone tissue morphology classification component to obtain a debugged feature extraction branch component; A reference estimation component is used to estimate the bone tissue morphology of each verification image in the bone tissue medical image verification set, and a second bone tissue morphology estimation label of each verification image in the bone tissue medical image verification set is obtained; the reference estimation component includes a first classification branch component and a debugged feature extraction branch component, the first classification branch component includes a target reference representation vector of each type of first bone tissue morphology, and the target reference representation vector is generated based on an image representation vector of a training image in the bone tissue medical image training set that includes the same actual bone tissue morphology label; According to the loss between the actual bone tissue morphology mark of each verification image in the bone tissue medical image verification set and the second bone tissue morphology estimated mark, the graph representation branch component is debugged to obtain the debugged graph representation branch component.

9. The method according to claim 8, characterized in that The first bone tissue morphology u is any one of the A categories of first bone tissue morphologies, and u is not greater than A; the bone tissue morphology recognition neural network further includes an image feature extraction component; the method further includes: The image feature extraction component is used to extract features from one or more second training images to obtain image representation vectors of each second training image; the second training image is a training image in the bone tissue medical image training set whose actual bone tissue morphology label is the first bone tissue morphology u; Using the debugged feature extraction branch component to perform feature extraction on the image representation vector of each second training image to obtain a refined representation vector of each second training image; Performing mean calculation on the refined representation vectors of each second training image to obtain a target reference representation vector of the first bone tissue morphology u; The feature extraction branch component is debugged by the first label loss under multiple training branches; the graph representation branch component is debugged by the second label loss under the multiple training branches; wherein the first training branch is any training branch among the multiple training branches; the first label loss under the first training branch is the loss between the first bone tissue morphology estimated label and the actual bone tissue morphology label of each verification image in the bone tissue medical image verification set; the second label loss under the first training branch is the loss between the second bone tissue morphology estimated label and the actual bone tissue morphology label of each verification image in the bone tissue medical image verification set.

10. A detection system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images