Specific disease iconography feature extraction method and system based on deep learning

Abstract

The invention discloses a specific disease iconography feature extraction method and system based on deep learning, and relates to the technical field of image analysis, and the method comprises the steps: collecting X-ray image data, stem cell markers, magnetic resonance imaging image data and dynamic ultrasonic image data of an osteoarthritis patient; performing standardization processing on the X-ray image data, the stem cell marker, the magnetic resonance imaging image data and the dynamic ultrasonic image data to obtain a standardized X-ray image data set, a standardized stem cell marker, a standardized magnetic resonance imaging image data set and a standardized dynamic ultrasonic image data set; a standardized X-ray image data set, a standardized magnetic resonance imaging image data set and a standardized dynamic ultrasound image data set are acquired. According to the invention, through multi-modal image data fusion and biomechanics-stem cell characteristic dynamic coupling, early accurate diagnosis and repair potential evaluation of osteoarthritis are realized.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a method and system for extracting specific disease imaging features based on deep learning. Background Technology

[0002] In recent years, the field of medical image analysis has evolved from traditional single-modal diagnosis to multimodal deep learning fusion systems. Existing technologies mainly use three-dimensional convolutional neural networks to process MRI sequence data, combine architectures such as EfficientNet to analyze X-ray images, and use optical flow to extract the temporal features of dynamic ultrasound. In the diagnosis of osteoarthritis, multimodal feature fusion can improve the accuracy of KL grading. With the introduction of the Swin-Transformer architecture, the window self-attention mechanism effectively captures changes in subchondral bone microstructure. Some advanced systems have achieved weighted fusion of stress distribution maps and T2 mapping values.

[0003] However, existing technologies have key shortcomings. In the early diagnosis of osteoarthritis, they fail to establish a mechanical environment, stem cell activity, and tissue degeneration. The stress distribution map only reflects mechanical load and does not incorporate stem cell marker features. Static weighting is used when the features are fused, and the contribution ratio of mechanical parameters and biological features cannot be adaptively adjusted. This results in limited sensitivity for KL0-1 stage cases, and the AUC value for predicting repair progress is low. Furthermore, the microblood flow signals extracted by existing methods in dynamic ultrasound lack spatiotemporal alignment with stem cell migration paths, leading to false positives. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for extracting specific disease imaging features based on deep learning, which solves the core problem of insufficient sensitivity in the diagnosis of early osteoarthritis caused by the inability of existing technologies to adaptively adjust the contribution ratio of biomechanical features and stem cell markers.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for extracting imaging features of specific diseases based on deep learning, which includes collecting X-ray image data, stem cell markers, magnetic resonance imaging data and dynamic ultrasound image data of patients with osteoarthritis, and standardizing the X-ray image data, stem cell markers, magnetic resonance imaging data and dynamic ultrasound image data to obtain a standardized X-ray image data set, a standardized stem cell marker set, a standardized magnetic resonance imaging data set and a standardized dynamic ultrasound image data set.

[0008] The standardized X-ray image data set, standardized magnetic resonance imaging image data set, and standardized dynamic ultrasound image data set are input into the β variational autoencoder network of EfficientNetV2 encoder, 3D SwinTransformer encoder, and ConvLSTM encoder to separate the osteoarthritis disease feature vector.

[0009] By simulating the joint mechanical behavior of osteoarthritis patients under physiological load through finite element analysis, a stress distribution map of the cartilage damage area is generated. The osteoarthritis disease feature vector and the stress distribution map of the cartilage damage area are fused. The fused stress distribution map of the cartilage damage area and stem cell markers are mapped to the biomechanical feature space of the same dimension as the osteoarthritis disease feature vector to obtain the biomechanical enhancement feature vector.

[0010] By inputting biomechanically enhanced feature vectors into a knowledge distillation framework, osteoarthritis grading results, disease progression risk scores, and visual heatmap diagnostic reports are obtained.

[0011] As a preferred embodiment of the deep learning-based method for extracting specific disease imaging features according to the present invention, the method includes: collecting X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data from patients with osteoarthritis; and standardizing the X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data to obtain standardized X-ray image data sets, standardized stem cell markers, standardized magnetic resonance imaging data sets, and standardized dynamic ultrasound image data sets, including the following steps.

[0012] Take anteroposterior and lateral radiographs of the patient's knee joint to obtain DICOM format X-ray image data. Apply a perspective transformation matrix to the X-ray image data to correct the projection angle deviation and obtain corrected X-ray image data. Perform grayscale normalization on the corrected X-ray image data to obtain a standardized X-ray image data set.

[0013] Bone marrow fluid was extracted from the patient's iliac crest, and after separating the mononuclear cell layer by density gradient centrifugation, standardized stem cell markers were obtained by flow cytometry.

[0014] T2 mapping sequences of the knee joint were acquired using a 3T MRI scanner, and magnetic resonance imaging data were output. The N4 bias field correction algorithm was applied to the magnetic resonance imaging data to eliminate brightness non-uniformity, and the corrected magnetic resonance imaging data were output. The corrected magnetic resonance imaging data were processed by B-spline interpolation to output a standardized magnetic resonance imaging data set.

[0015] Dynamic ultrasound video of knee flexion was recorded using a high-frequency linear probe, and dynamic ultrasound image data was output. The Farneback optical flow method was applied to the dynamic ultrasound image data for inter-frame motion compensation, and the registered dynamic ultrasound image data set was output. The femoral condyle and tibia regions were labeled to obtain a standardized dynamic ultrasound image data set.

[0016] As a preferred embodiment of the deep learning-based method for extracting specific disease imaging features according to the present invention, the method includes the following steps: inputting standardized X-ray image data sets, standardized magnetic resonance imaging image data sets, and standardized dynamic ultrasound image data sets into a β-variational autoencoder network consisting of an EfficientNetV2 encoder, a 3D SwinTransformer encoder, and a ConvLSTM encoder to separate the osteoarthritis disease feature vector.

[0017] The standardized X-ray image data set is input into the EfficientNetV2 encoder to obtain the X-ray feature vector. The standardized magnetic resonance imaging image data set is input into the three-dimensional SwinTransformer encoder, and the magnetic resonance imaging image feature vector is output through the window self-attention mechanism. The dynamic ultrasound image data set is input into the ConvLSTM encoder to obtain the dynamic ultrasound image feature vector.

[0018] The feature vectors of X-ray, magnetic resonance imaging, and dynamic ultrasound are concatenated according to the channel dimension to form a joint feature vector.

[0019] The joint feature vector is input into the encoding network of the β variational autoencoder to obtain the mean and variance of the joint feature vector;

[0020] The mean and variance of the joint feature vector are parameterized to obtain the feature vector of osteoarthritis.

[0021] As a preferred embodiment of the deep learning-based method for extracting specific disease imaging features according to the present invention, the method includes the following steps: simulating the joint biomechanical behavior of osteoarthritis patients under physiological load through finite element analysis to generate a stress distribution map of the cartilage damage area.

[0022] The standardized magnetic resonance imaging data set is input into the nnUNet framework to obtain a binary segmentation mask for cartilage and bone. The MarchingCubes algorithm is applied to the segmentation mask to obtain a three-dimensional triangular mesh model.

[0023] The surface mesh was transformed into a tetrahedral volume mesh using the TetGen algorithm. Tibial fixation constraints were applied to the three-dimensional triangular mesh model. Implicit dynamic solutions were obtained using the ANSYS MechanicalAPDL solver, and stress field data were output.

[0024] Stress data on the cartilage surface is extracted from the stress field data, normalized into a gray value matrix, and then mapped to a stress distribution map of the cartilage damage area.

[0025] As a preferred embodiment of the deep learning-based method for extracting specific disease imaging features according to the present invention, the method involves fusing the feature vectors related to osteoarthritis with the stress distribution map of the cartilage damage area, including the following steps:

[0026] The stress distribution map of the cartilage damage area is input into a 4-layer depthwise separable convolutional network to repair the cartilage damage and output a multi-scale stress coding feature map. Global average pooling is performed on the multi-scale stress coding feature map to obtain a biomechanical compressed feature vector.

[0027] The biomechanical compressed feature vector is expanded by a fully connected layer, and the expanded biomechanical compressed feature vector is concatenated with the osteoarthritis disease feature vector. Attention weights are generated by passing through a linear layer and Sigmoid activation.

[0028] Attention weights were used to weight the biomechanical compression feature vectors to obtain the stress distribution map of the fused cartilage damage area.

[0029] As a preferred embodiment of the deep learning-based method for extracting specific disease imaging features according to the present invention, the method includes mapping the fused stress distribution map and stem cell markers of the cartilage injury region to a biomechanical feature space of the same dimension as the osteoarthritis disease feature vector, comprising the following steps.

[0030] Input the stress distribution map of the cartilage damage area into the dual-branch network and output a multi-scale stress characterization vector.

[0031] The multi-scale stress characterization vector is extended by a fully connected layer to output a preliminary stress feature vector. The preliminary stress feature vector is then normalized by RMSNorm to output a standardized stress feature vector.

[0032] The standardized stress feature vector is concatenated with the osteoarthritis disease feature vector, and then gating weights are generated through a linear layer and Sigmoid activation.

[0033] The standardized stress eigenvector is weighted using gating weights, and the weighted stress eigenvector is output.

[0034] The weighted stress feature vector is added to the osteoarthritis disease feature vector to output the preliminary enhancement feature. The preliminary enhancement feature is then subjected to a nonlinear transformation by a single-layer MLP to output the intermediate enhancement feature. The intermediate enhancement feature is then processed by RMSNorm to output the biomechanical enhancement feature vector.

[0035] As a preferred embodiment of the deep learning-based method for extracting specific disease imaging features according to the present invention, the method includes the following steps: inputting biomechanically enhanced feature vectors into a knowledge distillation framework to obtain osteoarthritis grading results and disease progression risk scores, and then processing these to generate a visual heatmap diagnostic report.

[0036] The biomechanical enhancement feature vector is input into a 3-layer residual block teacher network, which outputs high-dimensional features and osteoarthritis grading results.

[0037] High-dimensional features are input into an LSTM time series model, which outputs a progress risk score.

[0038] Integrate osteoarthritis grading results and progression risk scores to generate a visual heatmap diagnostic report.

[0039] Secondly, the present invention provides a specific disease imaging feature extraction system based on deep learning, including a data acquisition module that collects X-ray image data, stem cell markers, magnetic resonance imaging data and dynamic ultrasound image data of patients with osteoarthritis, and performs standardization processing on the X-ray image data, stem cell markers, magnetic resonance imaging data and dynamic ultrasound image data to obtain a standardized X-ray image data set, a standardized stem cell marker set, a standardized magnetic resonance imaging data set and a standardized dynamic ultrasound image data set;

[0040] The separation module inputs the standardized X-ray image data set, standardized magnetic resonance imaging image data set, and standardized dynamic ultrasound image data set into the β variational autoencoder network of EfficientNetV2 encoder, 3D SwinTransformer encoder, and ConvLSTM encoder to separate the osteoarthritis disease feature vector.

[0041] The feature fusion module simulates the joint mechanical behavior of osteoarthritis patients under physiological load through finite element analysis, generates a stress distribution map of the cartilage damage area, fuses the osteoarthritis disease feature vector with the stress distribution map of the cartilage damage area, and maps the fused stress distribution map of the cartilage damage area and stem cell markers to a biomechanical feature space of the same dimension as the osteoarthritis disease feature vector to obtain a biomechanical enhancement feature vector.

[0042] The diagnostic decision module inputs biomechanically enhanced feature vectors into a knowledge distillation framework to obtain osteoarthritis grading results, disease progression risk scores, and visual heatmap diagnostic reports.

[0043] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the deep learning-based method for extracting specific disease imaging features as described in the first aspect of the present invention.

[0044] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based method for extracting specific disease imaging features as described in the first aspect of the present invention.

[0045] The beneficial effects of this invention are as follows: By fusing multimodal image data and dynamically coupling biomechanical-stem cell features, early and accurate diagnosis and assessment of repair potential of osteoarthritis are achieved. A β variational autoencoder network is used to decouple X-ray, MRI and ultrasound features. A three-dimensional SwinTransformer is used to capture 3D features of cartilage damage. ConvLSTM is used to track stem cell migration signals. Stress distribution from finite element analysis and stem cell markers from flow cytometry are fused through gated attention. Standardized preprocessing ensures data reliability. A knowledge distillation framework is used to output a visual heatmap, which simultaneously displays mechanically abnormal areas and repair hotspots, thereby improving the sensitivity of early diagnosis and reducing false positives. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a method for extracting specific disease imaging features based on deep learning.

[0048] Figure 2 This is a schematic diagram of a deep learning-based system for extracting imaging features of specific diseases.

[0049] Figure 3 This is a schematic diagram of stress distribution.

[0050] Figure 4 This is a schematic diagram of data acquisition and standardization processes. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0054] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for extracting specific disease imaging features based on deep learning, including the following steps:

[0055] S1. Collect X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data from patients with osteoarthritis, and standardize the X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data to obtain standardized X-ray image data sets, standardized stem cell markers, standardized magnetic resonance imaging data sets, and standardized dynamic ultrasound image data sets.

[0056] S1.1. Take anteroposterior and lateral radiographs of the patient's knee joint to obtain X-ray image data in DICOM format. Apply a perspective transformation matrix to the X-ray image data to correct the projection angle deviation and obtain corrected X-ray image data. Perform grayscale normalization on the corrected X-ray image data to obtain a standardized X-ray image data set.

[0057] Furthermore, the coordinates of the four tibial plateau corner points were used to obtain the perspective transformation matrix. The original X-ray image data was geometrically corrected through matrix operations to eliminate anatomical deformation caused by the projection angle. Then, a linear grayscale transformation with a window width of 2000HU and a window level of 500HU was performed on the corrected X-ray image data to map the pixel values ​​to the standard range, and finally a standardized X-ray image data set of 512×512×1 was generated.

[0058] S1.2 Bone marrow fluid was extracted from the patient's iliac crest, and after separating the mononuclear cell layer by density gradient centrifugation, standardized stem cell markers were obtained by flow cytometry.

[0059] Furthermore, after extracting bone marrow fluid from the patient's iliac crest, the mononuclear cell layer was separated by centrifugation using a Ficoll-Paque PLUS density gradient centrifuge. The obtained mononuclear cells were washed twice with PBS buffer and resuspended in α-MEM medium containing 10% fetal bovine serum. The cells were then seeded in culture flasks and cultured at 37°C and 5% CO2 for 72 hours. Adherent cells were collected and labeled with anti-CD73-FITC, anti-CD90-PE, and anti-CD105-APC fluorescent antibodies. The cells were then detected by BDFACSAriaIII flow cytometry. Standardized stem cell markers were output based on CD73 positivity rate ≥95%, CD90 positivity rate ≥98%, and CD105 positivity rate ≥90%.

[0060] S1.3. Obtain the T2 mapping sequence of the knee joint using a 3TMRI scanner, output magnetic resonance imaging data, apply the N4 bias field correction algorithm to the magnetic resonance imaging data to eliminate brightness non-uniformity, output corrected magnetic resonance imaging data, process the corrected magnetic resonance imaging data using B-spline interpolation, and output a standardized magnetic resonance imaging data set.

[0061] Furthermore, the original magnetic resonance imaging (MRI) image data is first iteratively optimized using the N4 bias field correction algorithm. By minimizing the difference between the bias field and the tissue signal, brightness inhomogeneity is eliminated. The corrected MRI image data is then subjected to isotropic resampling using a cubic B-spline interpolation algorithm, and the voxel size is uniformly adjusted to 1mm×1mm×1mm. Finally, a standardized MRI image data set of 256×256×64 is output.

[0062] S1.4. Record dynamic ultrasound video of knee flexion using a high-frequency linear probe, output dynamic ultrasound image data, apply Farneback optical flow method to the dynamic ultrasound image data for inter-frame motion compensation, output the registered dynamic ultrasound image data set, label the femoral condyle and tibia regions, and obtain the standardized dynamic ultrasound image data set.

[0063] Furthermore, a 12MHz high-frequency linear probe was used to record dynamic ultrasound videos of the knee joint flexed from 30° to 90°. After generating dynamic ultrasound image data, the Farneback optical flow method was first applied to calculate the displacement field between consecutive frames, and temporal registration of the dynamic ultrasound image data was achieved through affine transformation. Subsequently, the anatomical boundaries of the femoral condyle and tibial plateau were manually marked on the registered dynamic ultrasound image data set, and a 320×240 pixel region of interest was cropped and retained. Finally, a standardized dynamic ultrasound image data set of 320×240×15 frames was output.

[0064] S2. Input the standardized X-ray image data set, standardized magnetic resonance imaging image data set, and standardized dynamic ultrasound image data set into the β variational autoencoder network of EfficientNetV2 encoder, 3D SwinTransformer encoder, and ConvLSTM encoder to separate the osteoarthritis disease feature vector.

[0065] S2.1 Input the standardized X-ray image data set into the EfficientNetV2 encoder to obtain the X-ray feature vector. Input the standardized magnetic resonance imaging image data set into the three-dimensional SwinTransformer encoder and output the magnetic resonance imaging image feature vector through the window self-attention mechanism. Input the dynamic ultrasound image data set into the ConvLSTM encoder to obtain the dynamic ultrasound image feature vector.

[0066] Furthermore, the standardized X-ray image dataset is input into the EfficientNetV2 encoder. The EfficientNetV2 encoder's multi-layer convolutional structure extracts and compresses the spatial features of the standardized X-ray image dataset. During extraction, layer-by-layer convolutional kernel operations combined with nonlinear activation functions encode structural features such as edges and density distributions. After multi-layer downsampling operations, the final output is an X-ray feature vector. Next, the standardized magnetic resonance imaging (MRI) image dataset is input into a 3D SwingTransformer encoder. In the 3D SwingTransformer encoder, the standardized MRI image dataset is first divided into several non-overlapping three-dimensional... The system first divides the data into local blocks and applies a window self-attention mechanism to each block to capture structural features within the region. Simultaneously, it utilizes inter-window shifting operations to achieve cross-regional feature interaction. After multi-layer window division and self-attention calculation, it outputs a magnetic resonance imaging feature vector. Then, it inputs the standardized dynamic ultrasound image data set into a ConvLSTM encoder. The temporal series modeling structure in the ConvLSTM encoder encodes the spatiotemporal information between consecutive frames in the standardized dynamic ultrasound image data set. First, it extracts spatial structural features on each frame through convolution operations. Then, it uses the input gate, forget gate, and output gate in the LSTM structure to update the state and select and retain information for the temporal correlation between frames. Finally, it outputs a dynamic ultrasound image feature vector.

[0067] S2.2. The X-ray feature vector, magnetic resonance imaging feature vector, and dynamic ultrasound image feature vector are concatenated according to the channel dimension to form a joint feature vector.

[0068] Furthermore, after obtaining the X-ray feature vector, magnetic resonance imaging feature vector, and dynamic ultrasound image feature vector, the X-ray feature vector, magnetic resonance imaging feature vector, and dynamic ultrasound image feature vector are stitched together according to the channel dimension. During the stitching process, it is first ensured that the X-ray feature vector, magnetic resonance imaging feature vector, and dynamic ultrasound image feature vector are consistent in spatial dimensions other than the channel dimension. For example, the spatial size of each feature vector is adjusted to be consistent through necessary upsampling or downsampling operations. Then, the three feature vectors are stacked channel by channel according to the channel dimension, so that the channel information of each spatial location is jointly fused to form a joint feature vector.

[0069] S2.3 Input the joint feature vector into the encoding network of the β variational autoencoder to obtain the mean and variance of the joint feature vector.

[0070] Furthermore, the joint feature vector is spatially compressed by using multiple convolutional layers in the encoding network. The convolution operation is used to extract local patterns under different receptive fields. Then, the feature representation is nonlinearly mapped by batch normalization and activation functions to gradually compress the feature dimension and enhance the representation ability. Finally, the mean distribution parameter and variance distribution parameter of the joint feature vector in the latent space are calculated by two parallel fully connected layers, and the mean and variance of the joint feature vector are output.

[0071] S2.4. Parametrically sample the mean and variance of the joint feature vector to obtain the osteoarthritis disease feature vector.

[0072] Furthermore, a random noise tensor with the same dimension as the joint feature vector is sampled from the standard normal distribution as a perturbation term. Then, using the reparameter technique, the random noise tensor is element-wise calculated with the mean and variance of the joint feature vector. The variance of the joint feature vector is square-rooted to obtain the standard deviation tensor, which is then multiplied element-wise with the standard normal noise tensor. Finally, the result is added element-wise with the mean of the joint feature vector to generate a latent vector that conforms to the target latent distribution. The latent vector obtained in this way contains two parts of information: one part is the osteoarthritis disease feature vector that expresses the pathogenesis and structural manifestation of osteoarthritis, and the other part is the modality-independent noise feature vector that does not change with modality in the multimodal image input, which is used to distinguish between the intrinsic disease information and modality-specific differences.

[0073] S3. Simulate the joint mechanical behavior of osteoarthritis patients under physiological load through finite element analysis to generate stress distribution maps of cartilage damage areas.

[0074] S3.1 Input the standardized magnetic resonance imaging data into the nnUNet framework to obtain a binary segmentation mask for cartilage and bone. Apply the MarchingCubes algorithm to the segmentation mask to obtain a three-dimensional triangular mesh model.

[0075] Furthermore, the standardized magnetic resonance imaging (MRI) image dataset is input into the nnUNet framework. The nnUNet framework performs multi-layer feature extraction and context encoding on the voxel information in the standardized MRI image dataset. Through its fully convolutional neural network structure, it achieves segmentation and discrimination of local anatomical structures in the spatial dimension, ultimately outputting voxel-level prediction results consistent with the dimensions of the input standardized MRI image dataset. Voxels identified as cartilage and bone regions are designated as foreground, and the rest as background, thus constructing a binary segmentation mask for cartilage and bone. Subsequently, the Marching Cubes algorithm is applied to this binary segmentation mask for 3D surface reconstruction. The Marching Cubes algorithm extracts the boundary surfaces of foreground voxels in the 3D voxel mesh through local isosurface tracing. By performing triangulation on each local voxel block containing a mixture of foreground and background voxels, a 3D triangular mesh model forming a complete geometric structure is output.

[0076] It should be noted that the nnUNet framework is a general medical image segmentation network structure. Its training process employs supervised learning, with training data consisting of precisely labeled magnetic resonance imaging (MRI) images and corresponding cartilage and bone segmentation labels. The training objective is to minimize the voxel-level difference between the predicted segmentation mask and the ground truth segmentation labels. The loss function can be either Dice loss or cross-entropy loss. The example optimization algorithm is the Adam optimizer, with an example learning rate of 1e-3. Data augmentation techniques are used to improve generalization ability. Finally, the output nnUNet network parameters, after achieving stable convergence on the validation set, are used in the inference phase to generate binary segmentation masks for cartilage and bone corresponding to the standardized MRI image dataset.

[0077] S3.2. Use the TetGen algorithm to convert the surface mesh into a tetrahedral body mesh, apply tibial fixation constraints to the three-dimensional triangular mesh model, and perform implicit dynamic solution through the ANSYS Mechanical APDL solver to output stress field data.

[0078] Furthermore, the TetGen algorithm is used to perform volume meshing on the 3D triangular mesh model, refining the original surface mesh structure composed of triangular facets into a tetrahedral volume mesh structure filled with internal components. When processing a 3D triangular mesh model, the TetGen algorithm spatially partitions the surface mesh and introduces auxiliary nodes to construct structurally sound and mass-balanced tetrahedral elements within the geometric boundaries, forming a complete and closed tetrahedral volume mesh for subsequent finite element mechanical simulation input. After generating the tetrahedral volume mesh, a fixed constraint is applied to the tibia region corresponding to the 3D triangular mesh model. The fixed constraint restricts the displacement degrees of freedom of the tibia region in all directions, ensuring its existence as a mechanical boundary condition during the stress simulation. Subsequently, the constructed tetrahedral mesh structure and the set tibia fixed constraint are input into the ANSYS Mechanical APDL solver. The ANSYS Mechanical APDL solver performs implicit dynamics solving, which uses a global stability calculation method throughout the loading process to solve the stress distribution and deformation response of the main mesh structure under external force loading conditions. Finally, the stress field data corresponding to each element node is output.

[0079] S3.3 Extract cartilage surface stress data from the stress field data, normalize it into a gray value matrix, and map the gray value matrix into a stress distribution map of the cartilage damage area.

[0080] Specifically, the expression is,

[0081]

[0082] Where I(x,y) is the gray value matrix of the image at coordinates (x,y), σ max (x,y) represents the maximum principal stress value on the cartilage surface at coordinates (x,y), σ max σ represents the global maximum stress value on the cartilage surface. min This represents the global minimum stress value on the cartilage surface.

[0083] Furthermore, the maximum principal stress value of the cartilage surface is extracted from the stress field data according to the grid node index corresponding to the cartilage region, thus obtaining the cartilage surface stress data. Subsequently, the global maximum and minimum principal stress values ​​in the cartilage surface stress data are statistically analyzed, and based on the linear normalization method, the principal stress value of each cartilage surface node is converted into a normalized value according to its relative position between the minimum and maximum values. After normalization, a two-dimensional grayscale matrix consistent with the cartilage surface topology is constructed. The normalized grayscale values ​​are then mapped to the cartilage surface locations according to spatial correspondence, generating a stress distribution map that can be used for two-dimensional visualization of the cartilage damage area.

[0084] S4. Perform feature fusion between the osteoarthritis disease-related feature vectors and the stress distribution map of the cartilage damage area.

[0085] S4.1 Input the stress distribution map of the cartilage damage area into a 4-layer depthwise separable convolutional network to repair the cartilage damage and output a multi-scale stress coding feature map. Perform global average pooling on the multi-scale stress coding feature map to obtain a biomechanical compressed feature vector.

[0086] Furthermore, the four-layer deep separable convolutional network sequentially performs channel-wise spatial convolution and point-wise channel convolution operations in each layer to extract spatial texture features and inter-channel correlation features within the local region of the stress distribution map of the cartilage injury area. In the four consecutive deep separable convolutional layers, by setting different receptive fields and strides, multi-scale stress encoding feature maps extracted from regions of different scales are obtained. After completing the four-layer deep separable convolutional operations, the cartilage injury is repaired, and global average pooling is performed on the obtained multi-scale stress encoding feature maps. The average value of all spatial locations is calculated in each channel dimension to obtain the global response level of each channel, ultimately forming a biomechanically compressed feature vector with channel dimensions.

[0087] It should be noted that the training of the 4-layer deep separable convolutional network is conducted in a supervised or self-supervised manner. The training data consists of stress distribution maps of cartilage injury areas and their labels under a specific task objective (e.g., osteoarthritis grading or mechanical category). The loss function can be either cross-entropy loss or mean squared error loss. The network uses the Adam optimizer for parameter updates with a learning rate of 1e-4. Data augmentation methods can be used during training to improve generalization ability. Once the network converges and stabilizes on the validation set, the output parameters are used to generate multi-scale stress-encoded feature maps and extract biomechanical compressed feature vectors.

[0088] S4.2. The biomechanical compressed feature vector is expanded through a fully connected layer. The expanded biomechanical compressed feature vector is concatenated with the osteoarthritis disease feature vector and then generated into attention weights through a linear layer and Sigmoid activation.

[0089] Furthermore, the biomechanical compressed feature vector is dimensionally expanded using a fully connected layer, mapping the values ​​of each original channel dimension to a higher-dimensional space to enhance its representational power. The expanded biomechanical compressed feature vector is then dimensionally aligned with the osteoarthritis disease feature vector obtained through parameterized sampling via a β-variable autoencoder network in the previous step, and concatenated along the channel dimensions to form a fused joint feature representation. This concatenated joint feature representation is input into a linear layer, where a linear transformation is performed to obtain an intermediate vector. A Sigmoid activation function is applied to this intermediate vector to normalize the output value to the [0, 1] interval, generating attention weights.

[0090] S4.3. Use attention weights to weight the biomechanical compression feature vector to obtain the stress distribution map of the fused cartilage damage area.

[0091] Furthermore, the attention weights are multiplied element-wise with the biomechanical compressed feature vector to achieve a weighted adjustment of the features of each channel in the biomechanical compressed feature vector. The weight value corresponding to each channel in the attention weights is multiplied by the feature value of the same channel in the biomechanical compressed feature vector. This results in the enhanced features of channels with higher weights and the suppressed features of channels with lower weights in the weighted result. After the weighting process, the resulting feature representation is the stress distribution map of the fused cartilage injury area.

[0092] S5. Map the stress distribution map and stem cell markers of the fused cartilage damage area to a biomechanical feature space with the same dimension as the osteoarthritis disease feature vector.

[0093] S5.1 Input the stress distribution map of the cartilage damage area into the dual-branch network and output the multi-scale stress characterization vector.

[0094] Furthermore, the stress distribution map of the cartilage injury area is used as input and fed into two parallel processing channels of the dual-branch network. Each channel performs convolution and pooling operations on features at different scales to extract the representation information of joint stress at different spatial resolutions. The two branches capture fine-grained and coarse-grained stress features through multiple convolutional layers, and the output features of the two branches are fused to form a stress representation vector containing multi-scale information.

[0095] S5.2. The multi-scale stress characterization vector is extended through a fully connected layer to output a preliminary stress feature vector. The preliminary stress feature vector is normalized by applying RMSNorm to output a standardized stress feature vector.

[0096] Furthermore, the multi-scale stress characterization vector is input into the fully connected layer, and its dimension is expanded through weighted summation and linear transformation to output a preliminary stress feature vector with a higher dimension. Then, the RMSNorm normalization operation is applied to the preliminary stress feature vector. By taking the root mean square of each element of the vector and scaling and shifting, the dimensional differences and distribution offsets between features are eliminated, and finally, a standardized stress feature vector is obtained.

[0097] S5.3. The standardized stress feature vector is concatenated with the osteoarthritis disease feature vector, and then gating weights are generated through a linear layer and Sigmoid activation.

[0098] Specifically, the expression is,

[0099] g=σ(W g [f stress ;z d ]+b g );

[0100] Where g is the gating weight vector, W g Let b be the weight matrix. g For the bias term, f stress For the standardized stress eigenvector, z d This represents the feature vector of osteoarthritis.

[0101] Furthermore, the standardized stress feature vector and the osteoarthritis disease feature vector are concatenated along the feature dimension to form a unified feature representation. The concatenated features are then input into a linear layer. By performing weighted summation and bias superposition on the features, a linear mapping of the features is achieved, and the original values ​​of the gating weights are output. The Sigmoid activation function is then applied to the linear mapping result to map the output to between 0 and 1, generating gating weights that can be used to regulate the information flow.

[0102] S5.4 uses gated weights to weight the standardized stress eigenvectors and outputs a weighted stress eigenvector.

[0103] Specifically, the expression is,

[0104] f weighted =g⊙(W e f stress );

[0105] Among them, f weighted W is the weighted stress eigenvector. e It is a learnable transformation matrix.

[0106] Furthermore, the standardized stress feature vector is multiplied element-wise with the gated weight vector. The gated weight vector is constrained to the range of 0 to 1 by the Sigmoid activation function, achieving adaptive adjustment of the feature dimensions, and outputting a weighted stress feature vector. This operation applies weights through the Hadamard product. The gated weight vector is generated by concatenating the standardized stress feature vector and the osteoarthritis disease feature vector and then performing a linear transformation. The final output weighted stress feature vector retains the key disease-related dimensions of the biomechanical features.

[0107] S5.5. Add the weighted stress feature vector to the osteoarthritis disease feature vector to output the preliminary enhancement feature. Perform a nonlinear transformation on the preliminary enhancement feature through a single-layer MLP to output the intermediate enhancement feature. Process the intermediate enhancement feature with RMSNorm to output the biomechanical enhancement feature vector.

[0108] Furthermore, the weighted stress feature vector and the osteoarthritis disease feature vector are added element-wise to output preliminary enhancement features. These preliminary enhancement features are then input into a single-layer multilayer perceptron and subjected to a nonlinear transformation using the GeLU activation function to output 256-dimensional intermediate enhancement features. The intermediate enhancement features are then normalized using RMSNorm to output a biomechanical enhancement feature vector.

[0109] S6. Input the biomechanically enhanced feature vector into the knowledge distillation framework to obtain osteoarthritis grading results and disease progression risk scores, and then organize them to generate a visual heatmap diagnostic report.

[0110] S6.1 Input the biomechanical enhancement feature vector into the 3-layer residual block teacher network and output high-dimensional features and osteoarthritis grading results.

[0111] Furthermore, the biomechanical enhancement feature vector is input into a 3-layer residual block teacher network. Each residual block contains two fully connected layers and a skip connection. The first layer expands the input dimension to 512 dimensions, the second layer maintains 512 dimensions, and the third layer compresses it back to 256 dimensions. Each layer is followed by a ReLU activation function. The original feature information is preserved through the residual connections, and the final output is a 256-dimensional high-dimensional feature. The high-dimensional feature is input into a Softmax classifier, which outputs a 5-dimensional osteoarthritis grading probability distribution, corresponding to grades 0 to 4 of the Kellgren-Lawrence grading standard. The 3-layer residual block teacher network achieves deep analysis of the biomechanical enhancement feature vector through hierarchical feature transformation and residual learning. The skip connection alleviates the gradient vanishing problem, and the final grading result reflects the degree of joint degeneration.

[0112] S6.2 Input the high-dimensional features into the LSTM time series model and output the progress risk score.

[0113] Furthermore, high-dimensional features are input into the LSTM time series model, and the information flow is controlled by forget gate, input gate and output gate to perform time series modeling on the input feature sequence. The hidden state of the last time step of the LSTM time series model is input into the fully connected layer, and the output is constrained to the interval between 0 and 1 using the Sigmoid activation function to generate an osteoarthritis progression risk score.

[0114] S6.3 Integrate osteoarthritis grading results and progression risk scores to generate a visual heatmap diagnostic report.

[0115] Furthermore, the osteoarthritis grading results and progression risk scores are structurally integrated. The osteoarthritis grading results are labeled as levels 0 to 4 according to the Kellgren-Lawrence grading standard, and the progression risk scores are displayed as percentages. A Grad-CAM heatmap is calculated based on the feature map of the last convolutional layer of the teacher network, and high activation regions are marked. At the same time, a stress distribution map generated by finite element analysis is superimposed, and different color levels are used to distinguish mechanical abnormality regions. The grading results, risk scores, and fused heatmaps are formatted according to the standard medical report format to generate a visualized heatmap diagnostic report that includes textual diagnostic conclusions and visualized heatmaps.

[0116] This embodiment also provides a deep learning-based system for extracting specific disease imaging features, including:

[0117] The data acquisition module collects X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data from patients with osteoarthritis. It also standardizes the X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data to obtain standardized X-ray image data sets, standardized stem cell markers, standardized magnetic resonance imaging data sets, and standardized dynamic ultrasound image data sets.

[0118] The separation module inputs the standardized X-ray image data set, standardized magnetic resonance imaging image data set, and standardized dynamic ultrasound image data set into the β variational autoencoder network of EfficientNetV2 encoder, 3D SwinTransformer encoder, and ConvLSTM encoder to separate the osteoarthritis disease feature vector.

[0119] The feature fusion module simulates the joint mechanical behavior of osteoarthritis patients under physiological load through finite element analysis, generates a stress distribution map of the cartilage damage area, fuses the osteoarthritis disease feature vector with the stress distribution map of the cartilage damage area, and maps the fused stress distribution map of the cartilage damage area and stem cell markers to a biomechanical feature space of the same dimension as the osteoarthritis disease feature vector to obtain a biomechanical enhancement feature vector.

[0120] The diagnostic decision module inputs biomechanically enhanced feature vectors into a knowledge distillation framework to obtain osteoarthritis grading results, disease progression risk scores, and visual heatmap diagnostic reports.

[0121] This embodiment also provides a computer device applicable to the case of a deep learning-based method for extracting specific disease imaging features, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based method for extracting specific disease imaging features as proposed in the above embodiment.

[0122] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0123] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for extracting specific disease imaging features based on deep learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0124] In summary, this invention achieves early and accurate diagnosis and assessment of repair potential in osteoarthritis through multimodal image data fusion and dynamic coupling of biomechanics and stem cell features. It employs a β-variable autoencoder network to decouple X-ray, MRI, and ultrasound features, captures 3D features of cartilage damage using a three-dimensional SwinTransformer, tracks stem cell migration signals using ConvLSTM, fuses stress distribution from finite element analysis with stem cell markers from flow cytometry sorting using gated attention, ensures data reliability through standardized preprocessing, and outputs a visualized heatmap through a knowledge distillation framework, simultaneously displaying areas of mechanical abnormality and repair hotspots, thereby improving the sensitivity of early diagnosis and reducing false positives.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting specific disease imaging features based on deep learning, characterized in that: include, X-ray imaging data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound imaging data of patients with osteoarthritis were collected. The X-ray imaging data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound imaging data were standardized to obtain standardized X-ray imaging data sets, standardized stem cell markers, standardized magnetic resonance imaging data sets, and standardized dynamic ultrasound imaging data sets. The standardized X-ray image data set, standardized magnetic resonance imaging image data set, and standardized dynamic ultrasound image data set are input into the β variational autoencoder network of EfficientNetV2 encoder, 3D SwinTransformer encoder, and ConvLSTM encoder to separate the osteoarthritis disease feature vector. By simulating the joint mechanical behavior of osteoarthritis patients under physiological load through finite element analysis, a stress distribution map of the cartilage damage area is generated. The osteoarthritis disease feature vector and the stress distribution map of the cartilage damage area are fused. The fused stress distribution map of the cartilage damage area and stem cell markers are mapped to the biomechanical feature space of the same dimension as the osteoarthritis disease feature vector to obtain the biomechanical enhancement feature vector. By inputting biomechanically enhanced feature vectors into a knowledge distillation framework, osteoarthritis grading results, disease progression risk scores, and visual heatmap diagnostic reports are obtained.

2. The method for extracting specific disease imaging features based on deep learning as described in claim 1, characterized in that: X-ray imaging data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound imaging data of patients with osteoarthritis were collected. The X-ray imaging data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound imaging data were then standardized to obtain standardized X-ray imaging data sets, standardized stem cell marker data sets, standardized magnetic resonance imaging data sets, and standardized dynamic ultrasound imaging data sets. This process included the following steps: Take anteroposterior and lateral radiographs of the patient's knee joint to obtain DICOM format X-ray image data. Apply a perspective transformation matrix to the X-ray image data to correct the projection angle deviation and obtain corrected X-ray image data. Perform grayscale normalization on the corrected X-ray image data to obtain a standardized X-ray image data set. Bone marrow fluid was extracted from the patient's iliac crest, and after separating the mononuclear cell layer by density gradient centrifugation, standardized stem cell markers were obtained by flow cytometry. T2 mapping sequences of the knee joint were acquired using a 3T MRI scanner, and magnetic resonance imaging data were output. The N4 bias field correction algorithm was applied to the magnetic resonance imaging data to eliminate brightness non-uniformity, and the corrected magnetic resonance imaging data were output. The corrected magnetic resonance imaging data were processed by B-spline interpolation to output a standardized magnetic resonance imaging data set. Dynamic ultrasound video of knee flexion was recorded using a high-frequency linear probe, and dynamic ultrasound image data was output. The Farneback optical flow method was applied to the dynamic ultrasound image data for inter-frame motion compensation, and the registered dynamic ultrasound image data set was output. The femoral condyle and tibia regions were labeled to obtain a standardized dynamic ultrasound image data set.

3. The method for extracting specific disease imaging features based on deep learning as described in claim 2, characterized in that: Standardized X-ray image data sets, standardized magnetic resonance imaging image data sets, and standardized dynamic ultrasound image data sets are input into a β variational autoencoder network consisting of an EfficientNetV2 encoder, a 3D SwinTransformer encoder, and a ConvLSTM encoder to extract osteoarthritis disease feature vectors. Includes the following steps, The standardized X-ray image data set is input into the EfficientNetV2 encoder to obtain the X-ray feature vector. The standardized magnetic resonance imaging image data set is input into the three-dimensional SwinTransformer encoder, and the magnetic resonance imaging image feature vector is output through the window self-attention mechanism. The dynamic ultrasound image data set is input into the ConvLSTM encoder to obtain the dynamic ultrasound image feature vector. The feature vectors of X-ray, magnetic resonance imaging, and dynamic ultrasound are concatenated according to the channel dimension to form a joint feature vector. The joint feature vector is input into the encoding network of the β variational autoencoder to obtain the mean and variance of the joint feature vector; The mean and variance of the joint feature vector are parameterized to obtain the feature vector of osteoarthritis.

4. The method for extracting specific disease imaging features based on deep learning as described in claim 3, characterized in that: The finite element method (FEM) is used to simulate the joint biomechanical behavior of osteoarthritis patients under physiological load, generating stress distribution maps of cartilage damage areas. This includes the following steps: The standardized magnetic resonance imaging data set is input into the nnUNet framework to obtain a binary segmentation mask for cartilage and bone. The MarchingCubes algorithm is applied to the segmentation mask to obtain a three-dimensional triangular mesh model. The surface mesh was transformed into a tetrahedral volume mesh using the TetGen algorithm. Tibial fixation constraints were applied to the three-dimensional triangular mesh model. Implicit dynamic solutions were obtained using the ANSYS Mechanical APDL solver, and stress field data were output. Stress data on the cartilage surface is extracted from the stress field data, normalized into a gray value matrix, and then mapped to a stress distribution map of the cartilage damage area.

5. The method for extracting specific disease imaging features based on deep learning as described in claim 4, characterized in that: Feature fusion was performed between the disease-related feature vectors of osteoarthritis and the stress distribution map of the cartilage damage area. Includes the following steps, The stress distribution map of the cartilage damage area is input into a 4-layer depthwise separable convolutional network to repair the cartilage damage and output a multi-scale stress coding feature map. Global average pooling is performed on the multi-scale stress coding feature map to obtain a biomechanical compressed feature vector. The biomechanical compressed feature vector is expanded by a fully connected layer, and the expanded biomechanical compressed feature vector is concatenated with the osteoarthritis disease feature vector. Attention weights are generated by passing through a linear layer and Sigmoid activation. Attention weights were used to weight the biomechanical compression feature vectors to obtain the stress distribution map of the fused cartilage damage area.

6. The method for extracting specific disease imaging features based on deep learning as described in claim 5, characterized in that: The stress distribution map and stem cell markers of the fused cartilage injury area were mapped to a biomechanical feature space with the same dimension as the osteoarthritis disease feature vector. Includes the following steps, Input the stress distribution map of the cartilage damage area into the dual-branch network and output a multi-scale stress characterization vector. The multi-scale stress characterization vector is extended by a fully connected layer to output a preliminary stress feature vector. The preliminary stress feature vector is then normalized by RMSNorm to output a standardized stress feature vector. The standardized stress feature vector is concatenated with the osteoarthritis disease feature vector, and then gating weights are generated through a linear layer and Sigmoid activation. The standardized stress eigenvector is weighted using gating weights, and the weighted stress eigenvector is output. The weighted stress feature vector is added to the osteoarthritis disease feature vector to output the preliminary enhancement feature. The preliminary enhancement feature is then subjected to a nonlinear transformation by a single-layer MLP to output the intermediate enhancement feature. The intermediate enhancement feature is then processed by RMSNorm to output the biomechanical enhancement feature vector.

7. The method for extracting specific disease imaging features based on deep learning as described in claim 6, characterized in that: The biomechanically enhanced feature vectors are input into a knowledge distillation framework to obtain osteoarthritis grading results and disease progression risk scores. These are then processed to generate a visual heatmap diagnostic report, including the following steps: The biomechanical enhancement feature vector is input into a 3-layer residual block teacher network, which outputs high-dimensional features and osteoarthritis grading results. High-dimensional features are input into an LSTM time series model, which outputs a progress risk score. Integrate osteoarthritis grading results and progression risk scores to generate a visual heatmap diagnostic report.

8. A disease-specific imaging feature extraction system based on deep learning, based on the disease-specific imaging feature extraction method based on deep learning according to any one of claims 1 to 7, characterized in that: include, The data acquisition module collects X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data from patients with osteoarthritis. It also standardizes the X-ray image data, stem cell markers, magnetic resonance imaging data, and dynamic ultrasound image data to obtain standardized X-ray image data sets, standardized stem cell markers, standardized magnetic resonance imaging data sets, and standardized dynamic ultrasound image data sets. The separation module inputs the standardized X-ray image data set, standardized magnetic resonance imaging image data set, and standardized dynamic ultrasound image data set into the β variational autoencoder network of EfficientNetV2 encoder, 3D SwinTransformer encoder, and ConvLSTM encoder to separate the osteoarthritis disease feature vector. The feature fusion module simulates the joint mechanical behavior of osteoarthritis patients under physiological load through finite element analysis, generates a stress distribution map of the cartilage damage area, fuses the osteoarthritis disease feature vector with the stress distribution map of the cartilage damage area, and maps the fused stress distribution map of the cartilage damage area and stem cell markers to a biomechanical feature space of the same dimension as the osteoarthritis disease feature vector to obtain a biomechanical enhancement feature vector. The diagnostic decision module inputs biomechanically enhanced feature vectors into a knowledge distillation framework to obtain osteoarthritis grading results, disease progression risk scores, and visual heatmap diagnostic reports.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for extracting specific disease imaging features based on deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based method for extracting specific disease imaging features as described in any one of claims 1 to 7.

Images