Method for identifying rapid progress of gonitis through X-ray image analysis based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence and medical image analysis, in particular to an artificial intelligence-based X-ray image analysis gonitis rapid progress identification method, which comprises the following steps of: acquiring knee joint X-ray image data of a patient to form a time sequence set; preprocessing is carried out, the knee joint bone tissue contour is segmented, and a preprocessed image sequence is generated; performing feature extraction on the image sequence to generate a time sequence feature vector set; constructing a cascade neural network model, inputting the time sequence feature vector set into the cascade neural network model, and outputting a probability value of rapid progress of gonitis; presetting a probability threshold value, and judging and outputting whether the gonitis is in a rapid progress state or not according to a comparison result of the probability value and the preset threshold value. According to the method, the bone structure boundary is enhanced through multi-scale edge detection and adaptive threshold segmentation, multi-scale features are extracted by using the convolutional neural network and spatial pyramid pooling, time sequence modeling is realized by means of the cascaded neural network, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical image analysis technology, and in particular to an artificial intelligence-based method for identifying the rapid progression of knee osteoarthritis using X-ray image analysis. Background Technology

[0002] With the development of medical imaging technology and artificial intelligence, computer-aided diagnosis of knee osteoarthritis has evolved from traditional image processing to deep learning. Early methods relied on manually designed features, extracting morphological features such as joint spaces and osteophytes through algorithms like edge detection and threshold segmentation, and combining these with statistical models for disease assessment. However, these methods are limited by the representational capabilities of manually designed features and struggle to cope with complex image noise and individual anatomical differences. In recent years, deep learning technology has been gradually applied to medical image analysis. Convolutional neural networks have shown advantages in bone structure segmentation and lesion detection in single-frame images, while recurrent neural networks and their variants attempt to analyze disease progression through temporal modeling. However, existing models are insufficient in mining long-range dependencies in multi-time-point images and lack effective fusion of multi-scale and multi-resolution features, resulting in limited accuracy in identifying early, rapidly progressing cases.

[0003] Current technologies for identifying the progression of knee osteoarthritis still suffer from the following core limitations: First, single-time-point image analysis ignores the dynamic temporal characteristics of disease progression and fails to capture subtle changes in joint structure over time. Second, traditional temporal models can only model unidirectional time dependencies and lack the ability to model bidirectional correlations between long-interval time points, easily losing crucial early lesion information. Third, the contrast between bone and soft tissue in knee X-ray images is low, making it difficult for existing preprocessing methods to accurately segment the target region, resulting in background noise during feature extraction. Fourth, most models only use single-dimensional features, lacking joint analysis of multi-scale local textures and multi-resolution global structures, making it difficult to comprehensively depict the complex patterns of disease progression. These technical bottlenecks result in low prediction accuracy for rapid progression of knee osteoarthritis in existing systems, limiting their clinical application value. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned problems and provide an artificial intelligence-based method for identifying the rapid progression of knee osteoarthritis using X-ray image analysis. To achieve the above objective, this invention adopts the following technical solution:

[0005] An AI-based method for identifying rapid progression of knee osteoarthritis using X-ray image analysis includes the following steps:

[0006] Step S1: Obtain X-ray images of the knee joint at different time points for the same patient to form a time series set;

[0007] Step S2: Preprocess the images in the time series set, enhance the contrast of bone structure boundaries through multi-scale edge detection, and segment the contour of knee joint bone tissue based on an adaptive threshold algorithm to generate a preprocessed image sequence;

[0008] Step S3: Extract features from the preprocessed image sequence, fuse multi-scale local texture features and multi-resolution global structural features to generate a set of temporal feature vectors;

[0009] Step S4: Construct a cascaded neural network model, including an image feature extraction network, a temporal feature analysis network, and a progression classification network connected in sequence;

[0010] Step S5: Input the set of temporal feature vectors into the cascaded neural network model. After feature extraction, temporal dependency analysis and classification calculation, output the probability value of rapid progression of knee osteoarthritis.

[0011] Step S6: Set a preset probability threshold. Based on the comparison between the probability value and the preset threshold, determine and output whether the knee osteoarthritis is in a rapidly progressing state.

[0012] Furthermore, in step S2, the multi-scale edge detection is implemented as follows:

[0013] Step S211: Use a Gaussian kernel with standard deviations of σ1 and σ2 (σ1≠σ2) to perform dual-scale smoothing on the image, generating a first-scale smoothed image and a second-scale smoothed image;

[0014] Step S212: Perform the Canny edge detection algorithm on the smoothed images at each scale. The high and low thresholds are dynamically determined by the image grayscale histogram analysis to generate an edge image set.

[0015] Step S213: The dual-scale edge images are weighted and fused according to preset weight coefficients w1 and w2. These weight coefficients are dynamically adjusted based on the edge quality assessment of the training data. The calculation formula for the dual-scale edge images is as follows:

[0016] I j ′=w1·I e,j,1 +w2·I e,j,2

[0017] Among them, I j ′ represents the enhanced j-th image, I e,j,1 and I e,j,2 These are the edge images at the first and second scales, respectively.

[0018] Furthermore, in step S2, the bone tissue contour segmentation is implemented as follows:

[0019] Step S221: Calculate the grayscale histogram of the enhanced image and use the Otsu algorithm to solve for the optimal segmentation threshold T. j This threshold is dynamically determined by maximizing the inter-class variance between the background class and the target class, using the following formula:

[0020]

[0021] Where ω0(t) is the proportion of background pixels, ω1(t) is the proportion of target pixels, μ0(t) is the average gray value of background pixels, and μ1(t) is the average gray value of target pixels;

[0022] Step S222: Convert grayscale values ​​> T j The pixels are divided into bone tissue regions. Odd-sized structuring elements are used to perform morphological closing operations to remove discrete regions with an area smaller than a preset pixel threshold, thus completing bone contour segmentation.

[0023] Further, in step S3, the multi-scale local texture feature extraction is implemented through a convolutional neural network, which includes multiple convolutional blocks. Each convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer. The convolutional layers use an odd number of convolutional kernels with progressively increasing sizes. The number of convolutional kernels increases by a factor of two, and the stride and padding values ​​are dynamically adjusted according to the kernel size to maintain a constant spatial size. The feature maps output by the convolutional blocks are concatenated through channels to generate local texture feature vectors.

[0024] Furthermore, in step S3, the multi-resolution global structural feature extraction employs spatial pyramid pooling technology, including: dividing the image into multiple levels of grids with different resolutions, with the number of grids increasing layer by layer; performing max pooling on each grid cell to generate a feature vector at the corresponding level; concatenating the feature vectors at each level to generate a global structural feature vector, and fusing it with the local texture feature vector to form a temporal feature vector.

[0025] Further, in step S4, the temporal feature analysis network is a bidirectional long short-term memory network, and its input is a temporal feature vector sequence F = [F1, F2, ..., F...]. m The output is a concatenated vector h of the forward and backward hidden states. t Network status updates are implemented through a gating mechanism, as detailed below:

[0026] The forget gate controls the cell state at the previous time step. t-1 The retention ratio is calculated using the following formula:

[0027] f t =σ(W f ·[h t-1 F t ]+b f );

[0028] Among them, f t For the forget gate, σ is the Sigmoid function, and W f Let b be the forget gate weight matrix. f h is the forget gate bias vector. t-1 F is the hidden state from the previous time step. t This is the input feature vector for the current time step;

[0029] The input gate and candidate cell state together determine the updated content of the current input information, and the calculation formula is as follows:

[0030] i t =σ(W i ·[h t-1 ,F t ]+b i );

[0031]

[0032] Among them, i t The input gate output vector, σ is the sigmoid activation function, and W... i Let b be the input gate weight matrix. i h is the input gate bias vector. t-1 Let F be the hidden state vector from the previous time step. t The input feature vector at the current time step. Let W be the candidate cell state vector, tanh be the hyperbolic tangent function, and W be the tangent of the hyperbolic cell state vector. c Let b be the candidate cell state weight matrix. c This is the candidate cell state bias vector;

[0033] The current cell state is updated jointly by the forget gate and the input gate, and the calculation formula is as follows:

[0034]

[0035] Among them, c t f represents the current cell state. t For the Gate of Oblivion, c t-1 Let i be the cell state vector from the previous time step. t The input gate output vector, This represents the candidate cell state vector;

[0036] The output gate controls the output content of the current cell state, and the calculation formula is as follows:

[0037] o t =σ(W o ·[h t-1 ,F t ]+b o )

[0038] h t =o t ⊙tanh(c t )

[0039] Among them, o t Let W be the output vector of the output gate, σ be the sigmoid activation function, and W be the output vector of the output gate. o Let b be the output gate weight matrix. o h is the output gate bias vector. t-1 Let F be the hidden state vector from the previous time step. t h is the input feature vector at the current time step. t Let o be the hidden state vector at the current time step. t c is the output vector of the output gate. t This is the cell state vector at the current time step.

[0040] Further, in step S4, the progress classification network is a fully connected neural network, comprising:

[0041] Input layer: The number of neurons is the same as the output dimension of the Bi-LSTM, and it receives the hidden state vector h output by the temporal feature analysis network. t ;

[0042] Hidden layer: consists of multiple neurons, employing Dropout regularization and ReLU activation function. The calculation formula is as follows:

[0043] a = Dropout(max(0, W)) h ·h t +b h ))

[0044] Where a is the output vector, W h Let b be the hidden layer weight matrix. h Here, is the hidden layer bias vector, and Dropout is the random deactivation function;

[0045] Output layer: A two-dimensional probability vector is generated using the Softmax function. The formula for calculating the fast progression probability is:

[0046]

[0047] Where z0 is the input value of the output layer for non-rapid progression, z1 is the input value of the output layer for rapid progression; exp is the exponential function that converts the original score into a non-negative value; and P is the probability value of rapid progression of knee osteoarthritis.

[0048] Further, in step S4, the training process of the cascaded neural network model includes:

[0049] Step S41: Construct a multi-timepoint image dataset containing a sufficient number of patients, and label cases with joint space change rate > clinical standard threshold as rapid progression;

[0050] Step S42: Train using an adaptive optimizer and a binary cross-entropy loss function. The initial learning rate and batch size are set according to training requirements. The loss function calculation formula is:

[0051]

[0052] in, y is the cross-entropy loss value for binary classification, where N is the batch size and y is the number of samples. i p is the true label of the i-th sample. i The probability of rapid progress predicted by the model;

[0053] Step S43: When the change in the validation set loss over a preset number of consecutive rounds is less than a preset convergence threshold, stop training and save the model parameters. The advantages of this invention are:

[0054] 1. This invention performs multi-scale edge detection and adaptive threshold segmentation on knee joint X-ray image sequences. It uses Gaussian kernels with different standard deviations to smooth the images and dynamically determines the Canny edge detection threshold. Combined with the Otsu algorithm to maximize the inter-class variance, it achieves accurate segmentation of bone tissue contours. This effectively enhances the boundary clarity of bone structures in low-contrast images, reduces segmentation errors caused by noise and individual anatomical differences, provides a high-quality preprocessed data foundation for subsequent feature extraction, and improves the reliability of image analysis.

[0055] 2. This invention extracts multi-scale local texture features by constructing a convolutional neural network containing multiple convolutional blocks, generates multi-resolution global structural features by combining spatial pyramid pooling technology, and forms temporal feature vectors by channel splicing and dimension fusion. It can capture multi-level information from local texture to global structure from images, breaking through the limitations of traditional single-scale feature extraction. This enables the model to more comprehensively capture the subtle changes in joint structure during the progression of knee osteoarthritis, significantly improving the richness of feature expression and sensitivity to early lesions.

[0056] 3. This invention designs a cascaded neural network model, utilizes a bidirectional long short-term memory network to analyze the bidirectional dependencies of time-series feature vector sequences, combines a fully connected neural network for classification calculation, and dynamically adjusts model parameters through an adaptive optimizer and a binary cross-entropy loss function. This effectively solves the problem of insufficient modeling of long-interval time point dependencies in traditional time-series models, realizes dynamic time-series modeling of the rapid progression of knee osteoarthritis, significantly improves prediction accuracy, and provides efficient and reliable technical support for early clinical intervention decisions. Attached Figure Description

[0057] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0058] In the attached diagram:

[0059] Figure 1 This is a flowchart of an artificial intelligence-based method for identifying the rapid progression of knee osteoarthritis using X-ray image analysis. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0061] The present invention will now be described in detail and specifically through specific embodiments to enable a better understanding of the invention. However, the following embodiments do not limit the scope of protection of the present invention.

[0062] Example 1

[0063] An AI-based method for identifying rapid progression of knee osteoarthritis using X-ray image analysis includes the following steps:

[0064] Step S1: Obtain X-ray images of the knee joint at different time points for the same patient to form a time series set;

[0065] Step S2: Preprocess the images in the time series set, enhance the contrast of bone structure boundaries through multi-scale edge detection, and segment the contour of knee joint bone tissue based on an adaptive threshold algorithm to generate a preprocessed image sequence;

[0066] Step S3: Extract features from the preprocessed image sequence, fuse multi-scale local texture features and multi-resolution global structural features to generate a set of temporal feature vectors;

[0067] Step S4: Construct a cascaded neural network model, including an image feature extraction network, a temporal feature analysis network, and a progression classification network connected in sequence;

[0068] Step S5: Input the set of temporal feature vectors into the cascaded neural network model. After feature extraction, temporal dependency analysis and classification calculation, output the probability value of rapid progression of knee osteoarthritis.

[0069] Step S6: Set a preset probability threshold. Based on the comparison between the probability value and the preset threshold, determine and output whether the knee osteoarthritis is in a rapidly progressing state.

[0070] In a specific embodiment,

[0071] 1. Obtain anteroposterior X-ray images of the knee joint of the same patient at baseline and at 12 and 24 months of follow-up, forming an image sequence set containing three time points with a resolution of 512×512 pixels and a grayscale value range of 0-255.

[0072] 2. The image sequence is preprocessed. First, multi-scale edge detection is used to enhance the contrast of bone structure boundaries, which is achieved by dual-scale Gaussian smoothing, Canny edge detection and weighted fusion. Then, the contour of knee joint bone tissue is segmented based on an adaptive threshold algorithm, which is achieved by calculating the segmentation threshold using the Otsu algorithm and combining it with morphological closing operation to generate the preprocessed image sequence.

[0073] 3. Feature extraction is performed on the preprocessed image sequence. First, multi-scale local texture features are extracted using a convolutional neural network. Then, spatial pyramid pooling technology is used to extract multi-resolution global structural features. The two types of feature vectors are concatenated by channel to generate a temporal feature vector set.

[0074] 4. Construct a cascaded neural network model, connecting the image feature extraction network, the temporal feature analysis network, and the progression classification network in sequence. The parameters of each network are determined through training.

[0075] 5. Input the set of temporal feature vectors into the cascaded neural network model. The image feature extraction network extracts spatial features, the temporal feature analysis network analyzes the time series dependencies, and the progression classification network performs classification calculations to output the probability value P of rapid progression of knee osteoarthritis.

[0076] 6. A preset probability threshold of 0.6 can be set. If P≥0.6, it is determined to be a rapid progress state; otherwise, it is determined to be a non-rapid progress state, and the determination result is output.

[0077] The threshold setting was based on model training optimization using multi-timepoint image datasets: by collecting knee X-ray images of the same patient at different time points, cases with joint space change rates exceeding clinical standards were labeled as rapidly progressing samples, and an adaptive optimizer was used to train the cascaded neural network model. During training, binary cross-entropy was used as the loss function, and training stopped when the validation set loss changed by less than 0.001 for five consecutive rounds. At this point, the probability threshold of 0.6 output by the model was the optimal split point for balancing classification performance.

[0078] The rationality of the judgment results was verified through the following logic: First, using the joint space change rate as the clinical gold standard, the model prediction results were compared with the actual annotations, and the true positive rate and true negative rate were calculated on the test set; second, a bidirectional long short-term memory network was used to capture the bidirectional temporal dependencies of the image sequence, and its ability to capture the correlation information of long-interval time points was verified by comparing it with a single time point analysis model; third, multi-scale local texture features were extracted by a convolutional neural network, and multi-resolution global structural features were generated by combining spatial pyramid pooling technology, and ablation experiments were conducted to prove that the fusion of the two types of features can improve the classification accuracy; finally, the model's generalization ability was tested on an independent external dataset to ensure the stability of the threshold segmentation effect, so that the judgment results are consistent with the clinical diagnostic logic.

[0079] Furthermore, in step S2, the multi-scale edge detection is implemented as follows:

[0080] Step S211: Use a Gaussian kernel with standard deviations of σ1 and σ2 (σ1≠σ2) to perform dual-scale smoothing on the image, generating a first-scale smoothed image and a second-scale smoothed image;

[0081] Step S212: Perform the Canny edge detection algorithm on the smoothed images at each scale. The high and low thresholds are dynamically determined by the image grayscale histogram analysis to generate an edge image set.

[0082] Step S213: The dual-scale edge images are weighted and fused according to preset weight coefficients w1 and w2. These weight coefficients are dynamically adjusted based on the edge quality assessment of the training data. The calculation formula for the dual-scale edge images is as follows:

[0083] I j ′=w1·I e,j,1 +w2·I e,j,2

[0084] Among them, I j ′ represents the enhanced j-th image, I e,j,1 and I e,j,2 These are the edge images at the first and second scales, respectively.

[0085] In a specific embodiment,

[0086] 1. The image is smoothed using a Gaussian kernel with standard deviations of σ1 = 1.0 and σ2 = 2.0 to generate smoothed images at the first and second scales.

[0087] 2. The Canny edge detection algorithm is applied to smoothed images at various scales. The high and low thresholds are dynamically determined by analyzing the image grayscale histogram. The high threshold is set to the 75th percentile of the grayscale value distribution, and the low threshold is 0.4 times the high threshold, generating an edge image set.

[0088] 3. The dual-scale edge images are weighted and fused using preset initial weight coefficients w1 = 0.6 and w2 = 0.4. The weight coefficients are dynamically adjusted based on the edge quality assessment of the training data. The fusion formula is: I j ′=w1·I e,j,1 +w2·I e,j,2 .

[0089] Furthermore, in step S2, the bone tissue contour segmentation is implemented as follows:

[0090] Step S221: Calculate the grayscale histogram of the enhanced image and use the Otsu algorithm to solve for the optimal segmentation threshold T. j This threshold is dynamically determined by maximizing the inter-class variance between the background class and the target class, using the following formula:

[0091]

[0092] Where ω0(t) is the proportion of background pixels, ω1(t) is the proportion of target pixels, μ0(t) is the average gray value of background pixels, and μ1(t) is the average gray value of target pixels;

[0093] Step S222: Convert grayscale values ​​> T j The pixels are divided into bone tissue regions. Odd-sized structuring elements are used to perform morphological closing operations to remove discrete regions with an area smaller than a preset pixel threshold, thus completing bone contour segmentation.

[0094] In a specific embodiment, an odd number of structuring elements with a size of 3×3 can be used to perform morphological closing operations to remove discrete regions with an area of ​​less than 50 pixels, thereby completing bone contour segmentation.

[0095] Further, in step S3, the multi-scale local texture feature extraction is implemented through a convolutional neural network, which includes multiple convolutional blocks. Each convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer. The convolutional layers use an odd number of convolutional kernels with progressively increasing sizes. The number of convolutional kernels increases by a factor of two, and the stride and padding values ​​are dynamically adjusted according to the kernel size to maintain a constant spatial size. The feature maps output by the convolutional blocks are concatenated through channels to generate local texture feature vectors.

[0096] In a specific embodiment, feature extraction is achieved through a convolutional neural network, which contains three convolutional blocks:

[0097] First convolutional block: The convolutional layer uses a 3×3 convolutional kernel with 32 kernels, a stride of 1, padding of 1, followed by a batch normalization layer and a ReLU activation function layer;

[0098] Second convolutional block: The convolutional layer uses a 5×5 convolutional kernel, with a quantity of 64, a stride of 1, padding of 2, followed by a batch normalization layer and a ReLU activation function layer;

[0099] The third convolutional block: The convolutional layer uses a 7×7 convolutional kernel with 128 kernels, a stride of 1, padding of 3, followed by a batch normalization layer and a ReLU activation function layer;

[0100] The feature maps output by each convolutional block are concatenated through channels to generate local texture feature vectors.

[0101] Furthermore, in step S3, the multi-resolution global structural feature extraction employs spatial pyramid pooling technology, including: dividing the image into multiple levels of grids with different resolutions, with the number of grids increasing layer by layer; performing max pooling on each grid cell to generate a feature vector at the corresponding level; concatenating the feature vectors at each level to generate a global structural feature vector, and fusing it with the local texture feature vector to form a temporal feature vector.

[0102] In a specific embodiment, spatial pyramid pooling technology can be used to divide the image into three levels of grids with different resolutions:

[0103] Level 1: 1×1 grid, performs max pooling on the entire image to generate a 1-dimensional feature vector;

[0104] The second level uses a 2×2 grid to divide the image into four regions. Max pooling is performed on each region to generate a 4-dimensional feature vector.

[0105] Level 3: 4×4 grid, dividing the image into 16 regions, performing max pooling on each region to generate a 16-dimensional feature vector;

[0106] The feature vectors of each level are concatenated (1+4+16=21 dimensions) to generate a global structural feature vector, which is then fused with the local texture feature vector to form a temporal feature vector.

[0107] Further, in step S4, the temporal feature analysis network is a bidirectional long short-term memory network, and its input is a temporal feature vector sequence F = [F1, F2, ..., F...]. m The output is a concatenated vector h of the forward and backward hidden states. t Network status updates are implemented through a gating mechanism, as detailed below:

[0108] The forget gate controls the cell state at the previous time step. t-1 The retention ratio is calculated using the following formula:

[0109] f t =σ(W f ·[h t-1 F t ]+b f );

[0110] Among them, f t For the forget gate, σ is the Sigmoid function, and Wf Let b be the forget gate weight matrix. f h is the forget gate bias vector. t-1 F is the hidden state from the previous time step. t This is the input feature vector for the current time step;

[0111] The input gate and candidate cell state together determine the updated content of the current input information, and the calculation formula is as follows:

[0112] i t =σ(W i ·[h t-1 ,F t ]+b i );

[0113]

[0114] Among them, i t The input gate output vector, σ is the sigmoid activation function, and W... i Let b be the input gate weight matrix. i h is the input gate bias vector. t-1 Let F be the hidden state vector from the previous time step. t The input feature vector at the current time step. Let W be the candidate cell state vector, tanh be the hyperbolic tangent function, and W be the tangent of the hyperbolic cell state vector. c Let b be the candidate cell state weight matrix. c This is the candidate cell state bias vector;

[0115] The current cell state is updated jointly by the forget gate and the input gate, and the calculation formula is as follows:

[0116]

[0117] Among them, c t f represents the current cell state. t For the Gate of Oblivion, c t-1 Let i be the cell state vector from the previous time step. t The input gate output vector, This represents the candidate cell state vector;

[0118] The output gate controls the output content of the current cell state, and the calculation formula is as follows:

[0119] o t =σ(W o ·[h t-1 ,F t ]+b o )

[0120] h t =o t⊙tanh(c t )

[0121] Among them, o t Let W be the output vector of the output gate, σ be the sigmoid activation function, and W be the output vector of the output gate. o Let b be the output gate weight matrix. o h is the output gate bias vector. t-1 Let F be the hidden state vector from the previous time step. t h is the input feature vector at the current time step. t Let o be the hidden state vector at the current time step. t c is the output vector of the output gate. t This is the cell state vector at the current time step.

[0122] In a specific embodiment, the temporal feature analysis network can be set to a time step of m=3, corresponding to an image sequence at 3 time points, and input feature vector F. t The network has a dimension of 200. It contains one bidirectional LSTM layer, with both the forward and backward hidden layers having a dimension of 128. After the forward and backward LSTMs compute the hidden states separately, they are concatenated into a 256-dimensional vector as the output. Therefore, the output hidden state concatenation vector h is... t The dimension is 256 (128×2).

[0123] Further, in step S4, the progress classification network is a fully connected neural network, comprising:

[0124] Input layer: The number of neurons is the same as the output dimension of the Bi-LSTM, and it receives the hidden state vector h output by the temporal feature analysis network. t ;

[0125] Hidden layer: consists of multiple neurons, employing Dropout regularization and ReLU activation function. The calculation formula is as follows:

[0126] a = Dropout(max(0, W)) h ·h t +b h ))

[0127] Where a is the output vector, W h Let b be the hidden layer weight matrix. h Here, is the hidden layer bias vector, and Dropout is the random deactivation function;

[0128] Output layer: A two-dimensional probability vector is generated using the Softmax function. The formula for calculating the fast progression probability is:

[0129]

[0130] Where z0 is the input value of the output layer for non-rapid progression, z1 is the input value of the output layer for rapid progression; exp is the exponential function that converts the original score into a non-negative value; and P is the probability value of rapid progression of knee osteoarthritis.

[0131] In a specific embodiment,

[0132] Input layer: 256 neurons, corresponding to the hidden state dimension of the Bi-LSTM output.

[0133] Hidden layer: Contains a two-layer structure:

[0134] The first layer consists of 256 neurons, using Dropout regularization and the ReLU activation function.

[0135] The second layer consists of 128 neurons, using Dropout regularization and the ReLU activation function.

[0136] Output layer: 2 neurons, generating a two-dimensional probability vector through the Softmax function.

[0137] Further, in step S4, the training process of the cascaded neural network model includes:

[0138] Step S41: Construct a multi-timepoint image dataset containing a sufficient number of patients, and label cases with joint space change rate > clinical standard threshold as rapid progression;

[0139] Step S42: Train using an adaptive optimizer and a binary cross-entropy loss function. The initial learning rate and batch size are set according to training requirements. The loss function calculation formula is:

[0140]

[0141] in, y is the cross-entropy loss value for binary classification, where N is the batch size and y is the number of samples. i p is the true label of the i-th sample. i The probability of rapid progress predicted by the model;

[0142] Step S43: When the change in the validation set loss over a preset number of consecutive rounds is less than a preset convergence threshold, stop training and save the model parameters.

[0143] In a specific embodiment,

[0144] 1. Construct a multi-timepoint image dataset containing 1000 patients, label cases with joint space change rate >0.5mm / year as rapid progression, and the rest as negative samples, and divide the dataset into training set, validation set and test set in a ratio of 8:1:1.

[0145] 2. The Adam adaptive optimizer is used with an initial learning rate of 0.001 and a batch size of 32.

[0146] 3. When the validation set loss changes by less than 0.001 for 5 consecutive rounds, stop training and save the model parameters.

[0147] The specific embodiments of the present invention have been described in detail above, but they are merely examples, and the present invention is not equivalent to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent transformations and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis using artificial intelligence, characterized in that, Includes the following steps: Step S1: Obtain X-ray images of the knee joint at different time points for the same patient to form a time series set; Step S2: Preprocess the images in the time series set, enhance the contrast of bone structure boundaries through multi-scale edge detection, and segment the contour of knee joint bone tissue based on an adaptive threshold algorithm to generate a preprocessed image sequence; Step S3: Extract features from the preprocessed image sequence, fuse multi-scale local texture features and multi-resolution global structural features to generate a set of temporal feature vectors; Step S4: Construct a cascaded neural network model, including an image feature extraction network, a temporal feature analysis network, and a progression classification network connected in sequence; Step S5: Input the set of temporal feature vectors into the cascaded neural network model. After feature extraction, temporal dependency analysis and classification calculation, output the probability value of rapid progression of knee osteoarthritis. Step S6: Set a preset probability threshold. Based on the comparison between the probability value and the preset threshold, determine and output whether the knee osteoarthritis is in a rapidly progressing state.

2. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 1, characterized in that: In step S2, the multi-scale edge detection is implemented as follows: Step S211: Use a Gaussian kernel with standard deviations of σ1 and σ2 (σ1≠σ2) to perform dual-scale smoothing on the image, generating a first-scale smoothed image and a second-scale smoothed image; Step S212: Perform the Canny edge detection algorithm on the smoothed images at each scale. The high and low thresholds are dynamically determined by the image grayscale histogram analysis to generate an edge image set. Step S213: The dual-scale edge images are weighted and fused according to preset weight coefficients w1 and w2. These weight coefficients are dynamically adjusted based on the edge quality assessment of the training data. The calculation formula for the dual-scale edge images is as follows: I j ′=w1·I e,j,1 +w2·I e,j,2 Among them, I′ j For the enhanced j-th image, I e,j,1 and I e,j,2 These are the edge images at the first and second scales, respectively.

3. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 2, characterized in that: In step S2, the bone tissue contour segmentation is implemented as follows: Step S221: Calculate the grayscale histogram of the enhanced image and use the Otsu algorithm to solve for the optimal segmentation threshold T. j This threshold is dynamically determined by maximizing the inter-class variance between the background class and the target class, using the following formula: Where ω0(t) is the proportion of background pixels, ω1(t) is the proportion of target pixels, μ0(t) is the average gray value of background pixels, and μ1(t) is the average gray value of target pixels; Step S222: Convert grayscale values ​​> T j The pixels are divided into bone tissue regions. Odd-sized structuring elements are used to perform morphological closing operations to remove discrete regions with an area smaller than a preset pixel threshold, thus completing bone contour segmentation.

4. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 3, characterized in that: In step S3, the multi-scale local texture feature extraction is implemented through a convolutional neural network, which includes multiple convolutional blocks. Each convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer. The convolutional layers use an odd number of convolutional kernels with progressively increasing sizes. The number of convolutional kernels increases by a factor of two, and the stride and padding values ​​are dynamically adjusted according to the kernel size to maintain a constant spatial size. The feature maps output by the convolutional blocks are concatenated through channels to generate local texture feature vectors.

5. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 4, characterized in that: In step S3, the multi-resolution global structural feature extraction adopts spatial pyramid pooling technology, including: dividing the image into multiple levels of grids with different resolutions, with the number of grids increasing layer by layer; performing max pooling operation on each grid cell to generate feature vectors of the corresponding level; concatenating the feature vectors of each level to generate a global structural feature vector, and fusing it with local texture feature vectors to form a temporal feature vector.

6. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 5, characterized in that: In step S4, the temporal feature analysis network is a bidirectional long short-term memory network, and its input is a temporal feature vector sequence F = [F1, F2, ..., F...]. m The output is a concatenated vector h of the forward and backward hidden states. t Network status updates are implemented through a gating mechanism, as detailed below: The forget gate controls the cell state at the previous time step. t-1 The retention ratio is calculated using the following formula: f t =σ(W f ·[h t-1 ,F t ]+b f ); Among them, f t For the forget gate, σ is the Sigmoid function, and W f Let b be the forget gate weight matrix. f h is the forget gate bias vector. t-1 F is the hidden state from the previous time step. t This is the input feature vector for the current time step; The input gate and candidate cell state together determine the updated content of the current input information, and the calculation formula is as follows: i t =σ(W i ·[h t-1 ,F t ]+b i ); Among them, i t The input gate output vector, σ is the sigmoid activation function, and W... i Let b be the input gate weight matrix. i h is the input gate bias vector. t-1 Let F be the hidden state vector from the previous time step. t The input feature vector at the current time step. Let W be the candidate cell state vector, tanh be the hyperbolic tangent function, and W be the tangent of the hyperbolic cell state vector. c Let b be the candidate cell state weight matrix. c This is the candidate cell state bias vector; The current cell state is updated jointly by the forget gate and the input gate, and the calculation formula is as follows: Among them, c t f represents the current cell state. t For the Gate of Oblivion, c t-1 Let i be the cell state vector from the previous time step. t The input gate output vector, This represents the candidate cell state vector; The output gate controls the output content of the current cell state, and the calculation formula is as follows: the t =σ(W o ·[h t-1 ,F t ]+b o ) h t =o t ⊙tanh(c t ) Among them, o t Let W be the output vector of the output gate, σ be the sigmoid activation function, and W be the output vector of the output gate. o Let b be the output gate weight matrix. o h is the output gate bias vector. t-1 Let F be the hidden state vector from the previous time step. t h is the input feature vector at the current time step. t Let o be the hidden state vector at the current time step. t c is the output vector of the output gate. t This is the cell state vector at the current time step.

7. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 6, characterized in that: In step S4, the progress classification network is a fully connected neural network, comprising: Input layer: The number of neurons is the same as the output dimension of the Bi-LSTM, and it receives the hidden state vector h output by the temporal feature analysis network. t ; Hidden layer: consists of multiple neurons, employing Dropout regularization and ReLU activation function. The calculation formula is as follows: a=Dropout(max(0,W h ·h t +b h )) Where a is the output vector, W h Let b be the hidden layer weight matrix. h Here, is the hidden layer bias vector, and Dropout is the random deactivation function; Output layer: A two-dimensional probability vector is generated using the Softmax function. The formula for calculating the fast progression probability is: Where z0 is the input value of the output layer for non-rapid progression, z1 is the input value of the output layer for rapid progression; exp is the exponential function that converts the original score into a non-negative value; and P is the probability value of rapid progression of knee osteoarthritis.

8. The method for identifying rapid progression of knee osteoarthritis based on X-ray image analysis according to claim 7, characterized in that: In step S4, the training process of the cascaded neural network model includes: Step S41: Construct a multi-timepoint image dataset containing a sufficient number of patients, and label cases with joint space change rate > clinical standard threshold as rapid progression; Step S42: Train using an adaptive optimizer and a binary cross-entropy loss function. The initial learning rate and batch size are set according to training requirements. The loss function calculation formula is: in, y is the cross-entropy loss value for binary classification, where N is the batch size and y is the number of samples. i p is the true label of the i-th sample. i The probability of rapid progress predicted by the model; Step S43: When the change in the validation set loss over a preset number of consecutive rounds is less than a preset convergence threshold, stop training and save the model parameters.

Images