Orthopedic image auxiliary detection system based on big data

Abstract

The invention relates to the technical field of medical image detection, and discloses an orthopedic image auxiliary detection system based on big data, and the system comprises the steps: carrying out the sharpness processing of a to-be-detected orthopedic image through fuzzy detection, obtaining a corresponding clear orthopedic image, building and training an integrated classification-positioning cascade model based on big data, judging whether the clear orthopedic image to be detected is normal or not, positioning an abnormal part in the abnormal clear orthopedic image, marking the positioned abnormal part as an abnormal image block, constructing and training an orthopedic disease classification model based on big data, identifying the disease of the abnormal image block through the orthopedic disease classification model, and if the disease cannot be identified, judging whether the image is abnormal or not. And if yes, marking the abnormal image block as a difficult image block, triggering difficult identification, constructing a structured difficult orthopaedic image library, if difficult identification is triggered, screening similar images of the difficult image block in the difficult orthopaedic image library as auxiliary detection basis output, and updating the difficult orthopaedic image library.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging detection, and in particular to an orthopedic imaging-assisted detection system based on big data. Background Art

[0002] In the field of medical imaging diagnosis, accurate detection and analysis of orthopedic images are crucial for early detection, precise diagnosis, and treatment plan formulation of diseases. With the rapid development of medical imaging technology, the amount of orthopedic imaging data has exploded. Traditional manual diagnostic methods face huge challenges. They are inefficient and easily affected by factors such as the doctor's subjective experience and fatigue, resulting in frequent misdiagnosis and missed diagnosis.

[0003] The invention patent with announcement number CN118247276B provides an orthopedic image-assisted detection method and system, including: obtaining a patient's bone image; segmenting the bone image according to different grayscale levels in the grayscale histogram of the bone image to obtain an initial bone area; obtaining the bone density of each pixel point in the initial bone area; obtaining the grayscale value sequence and bone density sequence of the local area of ​​each pixel point in the initial bone area; obtaining the possibility that the pixel point in the initial bone area belongs to the bone area based on the correlation coefficient and standard correlation coefficient between the grayscale value sequence and the bone density sequence; and obtaining the bone area of ​​the bone image based on the possibility that each pixel point in the initial bone area belongs to the bone area.

[0004] In the existing technology, orthopedic images are prone to blurring during the acquisition process due to various factors such as equipment performance, patient movement, and improper operation. Blurred images not only reduce the visual quality of the image, but also interfere with subsequent image analysis and feature extraction, seriously affecting the accuracy of diagnosis. Existing technologies often ignore the evaluation and processing of image quality and directly analyze the original images, resulting in unreliable detection results on blurred images and increasing the risk of misdiagnosis. On the other hand, in the existing technology, a single model is used for classification and positioning when judging whether the image is normal and locating abnormal parts. It is difficult to cope with the complex and changeable orthopedic image features and is prone to misjudgment and omission. In addition, the specific positioning of the abnormal part is not accurate enough, and it is impossible to provide doctors with detailed lesion location information, which is not conducive to subsequent precise treatment. Orthopedic diseases are diverse and the symptoms are complex. The imaging features of different diseases may have similarities, which increases the difficulty of disease identification. Existing technologies have weak recognition capabilities for some rare diseases or diseases with atypical imaging features, resulting in an inability to provide practical assistance to patient diagnosis.

[0005] In response to the above problems, the present invention proposes an orthopedic image-assisted detection system based on big data. Summary of the Invention

[0006] The object of the present invention is to provide an orthopedic image-assisted detection system based on big data to solve at least one of the above-mentioned problems in the prior art.

[0007] The present invention provides an orthopedic image-assisted detection system based on big data, which is characterized by comprising the following modules:

[0008] Blur processing module: performs clarification processing on the orthopedic image to be detected to obtain a clear orthopedic image;

[0009] Abnormality analysis module: Based on big data training, the integrated classification-localization cascade model is used to determine abnormalities in clear orthopedic images. If abnormalities are detected, the abnormal area in the clear orthopedic image is located to generate abnormal image blocks.

[0010] Specifically, the integrated classification-localization cascade model includes a basic recognition model and an anomaly localization model;

[0011] Based on big data, a large number of orthopedic images labeled as normal are obtained and preprocessed to obtain training normal orthopedic images, and all training normal orthopedic images are integrated into a training dataset;

[0012] Construct a basic recognition model and design its architecture. Input the training dataset into the basic recognition model in batches for training. Divide the clear orthopedic images into blocks with a 50% overlap ratio. Input each block into the trained basic recognition model to obtain the abnormality probability of the block. If the abnormality probability of any block exceeds the probability threshold, the basic recognition model determines that the clear orthopedic image is abnormal.

[0013] If the clear orthopedic image is determined to be abnormal, the abnormality localization model is triggered, the clear orthopedic image is input into the abnormality localization model, the suspicious abnormal area in the clear orthopedic image is marked, and a segmentation mask of the suspicious abnormal area is output;

[0014] Perform contour detection on the obtained segmentation mask to generate a minimum bounding rectangle, and mark the clear orthopedic image within the minimum bounding rectangle as an abnormal image block;

[0015] Symptom identification module: This module trains an orthopedic disease classification model based on big data, screens for difficult diseases during the training process, and uses the orthopedic disease classification model to identify abnormal image blocks. If the abnormal image block cannot be identified, it will be marked as a difficult image block, triggering difficult disease identification.

[0016] Based on big data, a large number of orthopedic images of various conditions are acquired and preprocessed to obtain orthopedic images of training conditions. All orthopedic images of training conditions are integrated into a disease training dataset. An orthopedic disease classification sub-model is constructed. The disease training dataset is input into the orthopedic disease classification sub-model to complete basic training, and the recognition accuracy of each disease on the validation set is obtained.

[0017] If the identification accuracy of any disease is less than or equal to the accuracy threshold, the part of the orthopedic disease classification submodel for identifying the disease is removed, otherwise it is retained, and a trained orthopedic disease classification submodel is obtained, the disease training data set is updated, and a new orthopedic disease classification submodel is trained using the updated disease training data set;

[0018] The training of the new orthopedic disease classification submodel is repeated until the accuracy of all diseases on the test set is greater than the accuracy threshold, or the identified diseases of the orthopedic disease classification submodels of adjacent two times of training are the same, the training of the new orthopedic disease classification submodel is stopped, and if there is still a disease with an accuracy less than the accuracy threshold on the test set after stopping the training, the disease is marked as a difficult disease;

[0019] All trained orthopedic disease classification submodels are combined to obtain a trained orthopedic disease classification model;

[0020] The orthopedic disease classification model is used to identify the abnormal image block, the weighted average method is used to fuse the probability distribution output by each orthopedic disease classification submodel, and the disease category with the maximum probability is selected as the identification result of the abnormal image block according to the fused probability distribution, and if the maximum probability is less than the probability threshold, the abnormal image block is marked as a difficult image block;

[0021] The difficult identification module: if the difficult identification is triggered, the similar images of the abnormal image block are matched in the constructed structured difficult image library as auxiliary diagnosis basis;

[0022] A large number of difficult disease orthopedic images with complete medical records are obtained through big data and preprocessed, the preprocessed difficult disease orthopedic images are classified and stored according to the disease types in the complete medical records, and a structured difficult orthopedic image library is obtained;

[0023] The features of the difficult image block are extracted, and the geometric features, texture features and gray features obtained by extracting the features of the difficult image block are combined into a multi-dimensional feature vector;

[0024] The multi-dimensional feature vector of the difficult image block is extracted, and the Euclidean distance between the multi-dimensional feature vector and the multi-dimensional feature vectors of all difficult disease orthopedic images in the difficult orthopedic image library is calculated, if there is a difficult disease orthopedic image in the difficult orthopedic image library and the Euclidean distance between the multi-dimensional feature vectors of the difficult image block satisfies the determination condition, the difficult disease orthopedic image is output as the similar image of the difficult image block as the auxiliary detection basis for the difficult image block.

[0025] The beneficial effects of the present application are:

[0026] 1. The present invention effectively solves the blurring problem of orthopedic images caused by image acquisition problems through blur detection and clarity processing, greatly improving image quality and laying the foundation for subsequent accurate analysis. The constructed and trained integrated classification-positioning cascade model, with the support of big data and the advantages of integrated learning, can accurately determine whether the image is normal or not, and accurately locate the abnormal part and mark it as an abnormal image block. This not only improves the efficiency of preliminary screening for orthopedic diseases, but also provides strong support for doctors to quickly locate the lesion location, reducing misdiagnosis and missed diagnosis.

[0027] 2. The present invention combines the orthopedic disease classification model with the difficult image processing mechanism to perform detailed classification and identification of diseases. When identification is impossible, difficult identification is triggered, and similar images are screened from the structured difficult orthopedic image library for auxiliary detection, providing doctors with rich references. If there are no similar images in the difficult orthopedic image library, it will be updated in time to achieve dynamic accumulation of difficult image knowledge, help improve the diagnosis level of difficult diseases, and promote the development of orthopedic imaging diagnosis technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1 This is a system module architecture diagram of an orthopedic image-assisted detection system based on big data provided by an embodiment of the present invention;

[0030] Figure 2 This is a flow chart of the steps for generating abnormal image blocks in an orthopedic image-assisted detection system based on big data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0032] Example 1

[0033] like Figure 1 As shown, an orthopedic image-assisted detection system based on big data provided by an embodiment of the present invention specifically includes the following modules:

[0034] Blur processing module: obtains the orthopedic image to be detected, and marks the orthopedic image to be detected as a clear orthopedic image or a fuzzy orthopedic image through blur detection. If it is marked as a fuzzy orthopedic image, the fuzzy orthopedic image is cleared to obtain the corresponding clear orthopedic image;

[0035] Obtain orthopedic images to be tested, which mainly include X-rays and CT images. Perform spatial normalization on these images. For orthopedic images acquired by different devices with large resolution differences, use bilinear interpolation to unify the pixels of the images and resample the slices to equal spacing to balance detail preservation with computational efficiency and eliminate slice thickness differences.

[0036] The quantum noise or electronic noise of the spatially normalized orthopedic images was denoised using the non-local means filtering (NLM) algorithm, and the denoised orthopedic images were grayscale normalized to obtain preprocessed orthopedic images.

[0037] Based on any preprocessed orthopedic image, the Sobel operator is used to calculate the gradient components Gx and Gy of each pixel in the preprocessed orthopedic image in the horizontal and vertical directions respectively, using the formula:

[0038]

[0039] Calculate the gradient amplitude G of each pixel in the preprocessed orthopedic image, sum and average the gradient amplitudes G of all pixels in the preprocessed orthopedic image, and obtain the gradient mean GM of the entire image;

[0040] It should be noted that in clear pre-processed orthopedic images, the bone edges are sharp and the whole-image gradient mean GM is higher. Conversely, the lower the whole-image gradient mean GM is, the more blurred the pre-processed orthopedic image is.

[0041] Based on any pre-processed orthopedic image, the pre-processed orthopedic image is convolved with the Laplace operator to obtain a Laplace response matrix L, and the variance of the Laplace response matrix L is calculated to obtain the Laplace response variance VLP of the pre-processed orthopedic image;

[0042] It should be noted that the larger the Laplace response variance VLP is, the greater the difference in the Laplace response value is. There are a large number of grayscale mutation areas in the preprocessed orthopedic images, such as clear cortical bone edges and dense trabecular bone textures, that is, rich high-frequency components, and the preprocessed orthopedic images have high clarity. The smaller the VLP is, the more consistent the Laplace response values ​​are, the grayscale changes of the preprocessed orthopedic images are gentle, and the high-frequency components are scarce, such as blurred edges and lost textures, and the preprocessed orthopedic images are blurred.

[0043] If the full-image gradient mean GM of the preprocessed orthopedic image is greater than or equal to the gradient threshold, and the Laplace response variance VLP is greater than or equal to the variance threshold, the preprocessed orthopedic image is judged to be clear, and the preprocessed orthopedic image is marked as a clear orthopedic image;

[0044] If the full-image gradient mean GM of the preprocessed orthopedic image is less than the gradient threshold, or the Laplace response variance VLP is less than the variance threshold, the preprocessed orthopedic image is judged to be blurred, and the preprocessed orthopedic image is marked as a blurred orthopedic image;

[0045] For fuzzy orthopedic images, the Lucy-R i chardson iterative deconvolution algorithm is used to make them clear;

[0046] Specifically, based on the blur kernel k-iteration estimation of clear orthopedic images, the iterative update formula is:

[0047]

[0048] Among them, u (n) Indicates the estimated value of the clear orthopedic image at the nth iteration, the initial value u (0) is the blurred orthopedic image v or the preprocessed image, v represents the input blurred orthopedic image, k is the blur kernel, and the blind deblurring algorithm is used to estimate the blurred orthopedic image by performing block PSF estimation. * Represents the conjugate transpose of the blur kernel, which is used for inverse convolution to compensate for the blur effect;

[0049] The input blurred orthopedic image is iterated repeatedly. During the iteration process, total variation regularization is used to penalize unnecessary pixel changes in the image, suppress noise amplification and ringing effects, and forcibly retain the sharp edges of the bone structure to obtain a restored clear orthopedic image.

[0050] It should be noted that the function of this module is to identify fuzzy orthopedic images and realize standardized processing from fuzzy orthopedic images to clear orthopedic images, so as to prevent problems such as unclear bone cortical edges and poor display of joint spaces caused by fuzzy orthopedic images from affecting the accuracy of subsequent detection;

[0051] Abnormality analysis module: Based on big data, a large number of orthopedic images labeled as normal are acquired, and an integrated classification-localization cascade model is constructed and trained. This model is used to determine whether the clear orthopedic images to be tested are normal, locate abnormal parts in abnormal clear orthopedic images, and mark the located abnormal parts as abnormal image blocks.

[0052] like Figure 2 As shown, the specific steps of generating the abnormal image block are as follows:

[0053] Based on big data, a large number of orthopedic images labeled as normal are obtained, and a model is built and trained based on these images to determine whether the clear orthopedic images to be tested are normal.

[0054] Specifically, we integrated public databases and hospital PACS systems, screened and labeled normal orthopedic images, and used federated learning technology to desensitize hospital data, remove patient IDs, and blur areas outside the lesion to ensure compliance with H IPAA / GDPR privacy protection standards;

[0055] The acquired normal orthopedic images were spatially normalized, their pixels were unified using bilinear interpolation, and they were resampled to equal spacing using layer thickness. Quantum noise or electronic noise in the spatially normalized normal orthopedic images was denoised using a non-local mean filtering algorithm. The denoised normal orthopedic images were then grayscale normalized to obtain training normal orthopedic images. All training normal orthopedic images were integrated into a training dataset, which was then divided into a training set, a validation set, and a test set according to an 8:1:1 ratio.

[0056] Build an integrated classification-localization cascade model, including a basic recognition model and an anomaly localization model;

[0057] Specifically, a basic recognition model was constructed and its architecture was designed. The backbone network used Vision Transformer, which segmented the input orthopedic images into 16×16 pixel patches. Each patch was converted into a vector and then input into the Transformer. A 12-layer multi-head self-attention mechanism was used to capture the long-range dependencies of the skeletal structure, including the arrangement continuity of the vertebrae of the spine and the symmetry of the limb bones. Two fully connected layers were connected after the global features output by the Transformer. The first fully connected layer reduced the dimensionality and introduced the ReLU activation function to enhance nonlinear expression. The second fully connected layer outputted binary classification probabilities, with dimension 0 representing normal probability and dimension 1 representing abnormal probability.

[0058] The training set is input into the basic recognition model in batches. The backbone network extracts patch-level features and generates global features through the self-attention mechanism. The classification head outputs normal or abnormal probabilities, calculates the loss function and performs backpropagation, and updates the model parameters. DropPath is introduced in the Transformer layer to randomly close some network paths to enhance model generalization. Dropout is added to the classification head to randomly block neuron connections to reduce parameter co-adaptation. After each training round, the accuracy is calculated on the validation set and the parameters are adjusted. If the accuracy on the validation set does not improve for 10 consecutive training rounds, early stopping is triggered. When the training process stops according to the early stopping mechanism, the trained basic recognition model is finally evaluated using the test set to obtain the trained basic recognition model.

[0059] It should be noted that compared with traditional CNN, Transformer can better capture the global correlation of bone structure. For example, determining whether the tibia and fibula are aligned requires analyzing the positional relationship of the two bones at the same time, reducing the missed diagnosis rate;

[0060] The clear orthopedic image to be tested is divided into 512×512 pixel blocks with a block overlap rate of 50% to avoid missing edge lesions. Each block is input into the trained basic recognition model to obtain the abnormality probability of the block;

[0061] If the abnormal probabilities of all blocks are less than or equal to the probability threshold, the basic recognition model determines that the clear orthopedic image is normal and marks the clear orthopedic image as a detected healthy orthopedic image;

[0062] If the abnormal probability of any block is greater than the probability threshold, the basic recognition model determines that the clear orthopedic image is abnormal and triggers the abnormality localization model;

[0063] If the abnormality localization model is triggered, a clear orthopedic image is input into the abnormality localization model. The abnormality localization model generates a heat map using the Grad-CAM++ algorithm, marking the image area that the basic recognition model focuses on when judging the clear orthopedic image abnormality as a suspicious abnormal area. U-Net++ is used, and the encoder in U-Net++ is a pre-trained ResNet-34. The suspicious abnormal area is input into U-Net++, and a segmentation mask of the same size as the input is output. At the same time, Dice Loss (a measure of the overlap of the segmented areas) and BCELoss (binary cross entropy) are optimized to ensure that the abnormality localization model accurately distinguishes the boundaries of the suspicious abnormal area.

[0064] Perform contour detection on the obtained segmentation mask to generate a minimum bounding rectangle, and mark the clear orthopedic image within the minimum bounding rectangle as an abnormal image block;

[0065] It should be noted that the role of the abnormality localization model is to first narrow the scope of investigation by performing a global analysis of clear orthopedic images, and then use the segmentation network to refine the localization, avoiding the high computational cost of directly performing pixel-level detection on the entire image and improving detection efficiency.

[0066] Obtain all abnormal image blocks in clear orthopedic images, compare the distance between any two abnormal image blocks, and if the distance is less than the merging distance, determine that the two abnormal image blocks belong to the same lesion. Merge the positioning results of the two abnormal image blocks to avoid repeated labeling of the same lesion.

[0067] It should be noted that the role of this module is to learn global features based on Transformer, efficiently distinguish normal from abnormal images, reduce missed diagnoses, and narrow the scope of investigation from coarse screening of the entire image to fine screening of local areas, reducing computing costs, and output abnormal image blocks for subsequent identification. It combines the basic recognition model with the abnormality localization model, taking into account both global correlation analysis and pixel-level precise positioning.

[0068] Symptom Identification Module: Based on big data, a large number of orthopedic images with symptoms are acquired, and an orthopedic symptom classification model is constructed and trained. Symptoms that cannot be identified by the orthopedic symptom classification model are marked as difficult symptoms. The orthopedic symptom classification model is used to identify symptoms in abnormal image blocks. If the abnormal image blocks cannot be identified, they are marked as difficult image blocks, triggering difficult identification.

[0069] Based on big data, a large number of orthopedic images with complete medical records for various diseases are obtained. An orthopedic disease classification model is constructed and trained based on these images to identify abnormal image blocks and determine the disease of the abnormal image blocks.

[0070] Specifically, we integrated public databases and hospital PACS systems, screened orthopedic images of various conditions with complete medical records, marked them as condition-specific orthopedic images, and used federated learning technology to desensitize hospital data, remove patient IDs, and blur areas outside the lesion to ensure compliance with HIPAA / GDPR privacy protection standards;

[0071] The acquired orthopedic disease images are spatially standardized, and the pixels of the orthopedic disease images are unified through bilinear interpolation. The images are then resampled to equal spacing through layer thickness. The quantum noise or electronic noise in the spatially standardized orthopedic disease images is denoised using a non-local mean filtering algorithm. The denoised orthopedic disease images are then grayscale normalized to obtain training orthopedic disease images. All training orthopedic disease images are then integrated into a disease training dataset.

[0072] Based on the obtained training disease orthopedic images, an orthopedic disease classification model is constructed and trained, wherein the orthopedic disease classification model includes several orthopedic disease classification sub-models;

[0073] Specifically, the architecture of the orthopedic disease classification sub-model was designed. The backbone network selected ResNet-50, and the input layer was replaced with a three-channel convolution. The four-level residual block design of ResNet-50 was adopted. Each level consists of several bottleneck residual blocks. The SE attention module was connected after the fourth residual block. Channel-level attention was used to strengthen the features of key bone regions and suppress irrelevant background such as muscle and fat. The SE attention module outputs a feature map, which is connected to a fully connected layer with an output dimension equal to the number of disease categories. Softmax was used to generate and output a probability distribution. Focal loss was used as the loss function, and L2 regularization was set to suppress overfitting. A Dropout layer was added before the classification head to randomly block neuronal connections to improve generalization.

[0074] The obtained disease training dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The orthopedic disease classification sub-model was initialized and hyperparameters were set. A batch of data and corresponding labels were extracted from the training set and fed into the model for forward propagation. The focal loss with L2 regularization was calculated. The gradient of the focal loss with respect to the model parameters was calculated using the automatic differentiation framework. After each training round, the accuracy of each disease on the validation set was calculated and the parameters were adjusted. If the accuracy of each disease on the validation set did not improve after 10 consecutive training rounds, early stopping was triggered. When the training process was stopped according to the early stopping mechanism, the orthopedic disease classification sub-model completed basic training.

[0075] For the orthopedic disease classification sub-model that has completed basic training, calculate the recognition accuracy of each disease on the validation set. For any disease type, if the recognition accuracy of the disease is greater than the accuracy threshold, the disease is determined to be accurately recognized and marked as an accurately recognized disease. If the recognition accuracy of the disease is less than or equal to the accuracy threshold, the disease is marked as an inaccurately recognized disease.

[0076] Analyze all symptoms and divide them into accurately recognized symptoms and inaccurately recognized symptoms. Keep the part of the orthopedic disease classification sub-model that recognizes accurately recognized symptoms and remove the part of the orthopedic disease classification sub-model that recognizes inaccurately recognized symptoms, thus obtaining a trained orthopedic disease classification sub-model.

[0077] In the disease training dataset, orthopedic images that accurately identify diseases are removed, orthopedic images that inaccurately identify diseases are retrieved from the medical big data platform, the disease training dataset is supplemented to obtain an updated disease training dataset, the labels of inaccurately identified diseases are removed, and the updated disease training dataset is used to train a new trained orthopedic disease classification sub-model;

[0078] Repeat the training of different orthopedic disease classification sub-models until the accuracy of all diseases on the test set is greater than the accuracy threshold, or the identification of the same disease by the orthopedic disease classification sub-models trained twice are stopped. If the accuracy of any disease on the test set is still less than the accuracy threshold after stopping training, the disease will be marked as difficult to identify.

[0079] Combining all trained orthopedic disease classification sub-models to obtain a trained orthopedic disease classification model;

[0080] The trained orthopedic disease classification model is used to identify the abnormal image blocks obtained. The probability distribution output by each orthopedic disease classification sub-model is fused using the weighted average method. Based on the fused probability distribution, the disease category with the highest probability is selected as the recognition result output of the abnormal image block. If the maximum probability is less than the probability threshold, the abnormal image block is judged to be unrecognizable and marked as a difficult image block, triggering difficult recognition;

[0081] It should be noted that this module targets multiple orthopedic conditions, strengthens the features of key skeletal regions, improves classification accuracy, dynamically updates training data, focuses on identifying inaccurate conditions, gradually improves the model's generalization ability for rare symptoms, retains high-accuracy orthopedic condition classification sub-models, removes low-accuracy ones, avoids redundant parameters, improves inference efficiency, and supplements data for conditions that do not meet the standards, achieving accurate iteration of orthopedic condition classification models.

[0082] Problem identification module: Builds a structured library of difficult orthopedic images. If problem identification is triggered, similar images of the difficult image block in the library are screened and output as auxiliary detection basis. If similar images of the difficult image block do not exist in the library, the library is updated.

[0083] Based on labeled difficult diseases, we use big data to obtain a large number of difficult orthopedic images with complete medical records and build a structured difficult orthopedic image library;

[0084] Specifically, we integrated public databases and hospital PACS systems, screened and annotated orthopedic images of difficult diseases with complete medical records, marked them as difficult orthopedic images, and used federated learning technology to desensitize the hospital data, remove patient IDs, and blur areas outside the lesion to ensure compliance with HIPAA / GDPR privacy protection standards;

[0085] The acquired orthopedic images of difficult orthopedic conditions are spatially standardized. The pixels of the images are unified through bilinear interpolation and resampled to equal spacing through layer thickness. The quantum noise or electronic noise in the spatially standardized orthopedic images is denoised using a non-local mean filtering algorithm. The denoised orthopedic images are then grayscale normalized and stored according to the disease type in the annotated complete medical records to obtain a structured library of difficult orthopedic images.

[0086] Feature extraction is performed on each difficult orthopedic image in the difficult orthopedic image database. The extracted features include geometric features, texture features, and grayscale features. The geometric features, texture features, and grayscale features extracted from each difficult orthopedic image are combined into a multidimensional feature vector.

[0087] If difficult recognition is triggered, based on any difficult image block obtained, feature extraction is performed on the difficult image block, and the geometric features, texture features and grayscale features obtained by feature extraction of the difficult image block are combined into a multi-dimensional feature vector;

[0088] Calculate the Euclidean distance d(x,y) between the multidimensional feature vector of the difficult image block and the multidimensional feature vector of any difficult orthopedic image. The formula is:

[0089]

[0090] Among them, x represents the multidimensional feature vector of the difficult image block, y represents the multidimensional feature vector of the difficult orthopedic image, N represents the dimension of the feature vector, x i The i-th element in the multidimensional feature vector representing the difficult image block, y i The i-th element in the multidimensional feature vector representing difficult orthopedic imaging;

[0091] According to the calculated Euclidean distance, it is determined whether there is an orthopedic image of a difficult disease in the difficult orthopedic image database and the Euclidean distance between the multidimensional feature vector of the difficult image block is less than the distance threshold;

[0092] If so, the orthopedic image of the difficult disease is output as a similar image of the difficult image block, which serves as an auxiliary detection basis for the difficult image block;

[0093] If it does not exist, it is determined that there is no orthopedic image of a difficult disease similar to the difficult image block in the difficult orthopedic image library, and a new entry is created in the difficult orthopedic image library. The orthopedic surgeon further analyzes and diagnoses the difficult image block, marks the disease type of the difficult image block, and includes the difficult image block in the newly created entry in the difficult orthopedic image library;

[0094] It should be noted that the role of this module is to provide similar case references for unidentified diseases, assist doctors in diagnosis, and gradually expand the difficult disease database through manual annotation and data iteration to form a sustainable and evolving knowledge base of orthopedic diseases. It combines geometric features, texture features, and grayscale features to construct a multidimensional feature vector, which is more comprehensive than single feature matching and improves the ability to distinguish difficult diseases. The combination of automatic matching and manual annotation not only utilizes the efficiency of AI, but also ensures the clinical accuracy of rare disease annotation, avoiding the limitations of pure data-driven methods.

[0095] The technical solution of the embodiment of the present invention is as follows: obtaining an orthopedic image to be detected, marking the orthopedic image to be detected as a clear orthopedic image or a fuzzy orthopedic image through fuzzy detection, if it is marked as a fuzzy orthopedic image, performing a clear processing on the fuzzy orthopedic image to obtain a corresponding clear orthopedic image, obtaining a large number of orthopedic images marked as normal based on big data, constructing and training an integrated classification-positioning cascade model, using the integrated classification-positioning cascade model to determine whether the clear orthopedic image to be detected is normal, and locating the abnormal part in the abnormal clear orthopedic image, marking the located abnormal part as abnormal For common image blocks, a large number of orthopedic images of diseases are obtained based on big data, and an orthopedic disease classification model is constructed and trained. Diseases that cannot be identified by the orthopedic disease classification model are marked as difficult diseases. The orthopedic disease classification model is used to identify the diseases of abnormal image blocks. If they cannot be identified, the abnormal image blocks are marked as difficult image blocks, triggering difficult recognition and building a structured difficult orthopedic image library. If difficult recognition is triggered, similar images of difficult image blocks in the difficult orthopedic image library are screened as auxiliary detection basis output. If similar images of difficult image blocks do not exist in the difficult orthopedic image library, the difficult orthopedic image library is updated.

[0096] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An orthopedic image-assisted detection system based on big data, characterized in that: Includes the following modules: Blur processing module: performs clarification processing on the orthopedic image to be detected to obtain a clear orthopedic image; Abnormality analysis module: Based on big data training, the integrated classification-localization cascade model is used to determine abnormalities in clear orthopedic images. If abnormalities are detected, the abnormal area in the clear orthopedic image is located to generate abnormal image blocks. Symptom identification module: This module trains an orthopedic disease classification model based on big data, screens for difficult diseases during the training process, and uses the orthopedic disease classification model to identify abnormal image blocks. If the abnormal image block cannot be identified, it will be marked as a difficult image block, triggering difficult disease identification. Problem identification module: If problem identification is triggered, similar images are matched for abnormal image blocks in the constructed structured problem image library as auxiliary diagnosis basis.

2. The orthopedic image-assisted detection system based on big data according to claim 1, characterized in that: The integrated classification-localization cascade model includes a basic recognition model and an anomaly localization model.

3. The orthopedic image-assisted detection system based on big data according to claim 1, characterized in that: The abnormal image block is generated as follows: If the clear orthopedic image is determined to be abnormal, the abnormality localization model is triggered, the clear orthopedic image is input into the abnormality localization model, the suspicious abnormal area in the clear orthopedic image is marked, and a segmentation mask of the suspicious abnormal area is output; Contour detection is performed on the obtained segmentation mask to generate a minimum bounding rectangle, and the clear orthopedic image within the minimum bounding rectangle is marked as an abnormal image block.

4. The orthopedic image-assisted detection system based on big data according to claim 3, characterized in that: The method for determining the abnormality of the clear orthopedic image is as follows: Based on big data, a large number of orthopedic images labeled as normal are obtained and preprocessed to obtain training normal orthopedic images, and all training normal orthopedic images are integrated into a training dataset; Construct a basic recognition model, design the architecture of the basic recognition model, input the training data set into the basic recognition model in batches for training, divide the clear orthopedic images into blocks with a block overlap rate of 50%, input each block into the trained basic recognition model, and obtain the abnormal probability of the block. If the abnormal probability of any block is greater than the probability threshold, the basic recognition model determines that the clear orthopedic image is abnormal.

5. The orthopedic image-assisted detection system based on big data according to claim 1, characterized in that: The method for obtaining the difficult image block is: Obtain all trained orthopedic disease classification sub-models and combine them to obtain a trained orthopedic disease classification model; An orthopedic disease classification model is used to identify abnormal image blocks. The probability distribution output by each orthopedic disease classification sub-model is fused using the weighted average method. Based on the fused probability distribution, the disease category with the highest probability is selected as the recognition result output of the abnormal image block. If the maximum probability is less than the probability threshold, the abnormal image block is marked as a difficult image block.

6. The orthopedic image-assisted detection system based on big data according to claim 5, characterized in that: The orthopedic disease classification sub-model is obtained as follows: Based on big data, a large number of orthopedic images of various conditions are acquired and preprocessed to obtain orthopedic images of training conditions. All orthopedic images of training conditions are integrated into a disease training dataset. An orthopedic disease classification sub-model is constructed. The disease training dataset is input into the orthopedic disease classification sub-model to complete basic training, and the recognition accuracy of each disease on the validation set is obtained. If the recognition accuracy of any disease is less than or equal to the accuracy threshold, the part of the orthopedic disease classification sub-model that identifies the disease will be cleared, otherwise it will be retained to obtain the trained orthopedic disease classification sub-model, and the disease training dataset will be updated. The updated disease training dataset will be used to train a new orthopedic disease classification sub-model.

7. The orthopedic image-assisted detection system based on big data according to claim 1, characterized in that: The method of screening difficult diseases is: Repeatedly train the new orthopedic disease classification sub-model until the accuracy of all diseases on the test set is greater than the accuracy threshold, or the identification diseases of the orthopedic disease classification sub-model trained twice are the same, then stop training the new orthopedic disease classification sub-model. If there are still diseases with an accuracy lower than the accuracy threshold on the test set after stopping training, the disease will be marked as a difficult disease.

8. The orthopedic image-assisted detection system based on big data according to claim 1, characterized in that: The method of matching similar images is: The multidimensional feature vector of the difficult image block is extracted, and the Euclidean distance between the multidimensional feature vector and the multidimensional feature vectors of all orthopedic images of difficult diseases in the difficult orthopedic image library is calculated. If the Euclidean distance between the multidimensional feature vector of the orthopedic image of a difficult disease in the difficult orthopedic image library and the difficult image block meets the judgment condition, the orthopedic image of the difficult disease is output as a similar image of the difficult image block as an auxiliary detection basis for the difficult image block.

9. The orthopedic image-assisted detection system based on big data according to claim 8, characterized in that: The multi-dimensional feature vector of the difficult image block is extracted as follows: Feature extraction is performed on the difficult image blocks, and the geometric features, texture features and grayscale features obtained by feature extraction of the difficult image blocks are combined into a multi-dimensional feature vector.

10. The orthopedic image-assisted detection system based on big data according to claim 8, characterized in that: The method for obtaining the difficult orthopedic imaging library is as follows: A large number of orthopedic images of difficult diseases with complete medical records are obtained through big data and preprocessed. The preprocessed orthopedic images of difficult diseases are classified and stored according to the disease types in the complete medical records to obtain a structured difficult orthopedic image library.

Images