A deep learning-based oral x-ray automatic diagnosis system

Abstract

The application discloses a kind of oral cavity X-ray automatic diagnosis system based on deep learning.Belong to intelligent medical technology field, including: equipment correction module, feature alignment module, sample enhancement module, light inference module;The equipment correction module is used to obtain the input oral cavity X-ray image, by matching the preset equipment parameter library to execute the adaptive correction of gray, noise and artifact, output standardization image;The feature alignment module is electrically connected with equipment correction module.The present application effectively overcomes the difference between different imaging devices through equipment specificity correction, guarantees the standardized quality of input image, enhances the generalization ability of model in different medical institutions using field adaptive method, reduces the adaptation cost of cross-institution deployment, for the problem of rare disease sample shortage, expands the disease coverage range of auxiliary diagnosis, while maintaining high diagnostic performance, greatly reduces the dependence on hardware computing power.

Description

Technical Field

[0001] This invention relates to the field of smart medical technology, and in particular to an automated oral X-ray diagnostic system based on deep learning. Background Technology

[0002] With the rapid penetration of deep learning technology into the field of medical imaging, automated dental X-ray diagnostic systems are gradually becoming a core direction for the intelligent upgrading of dental healthcare. These systems can analyze various dental X-ray images, including panoramic films, periapical films, and CBCT scans, enabling the automatic identification of common oral diseases such as dental caries and periodontal disease. They play a positive role in improving diagnostic efficiency and reducing the workload of manual image interpretation. Currently, most related technical solutions are based on general computer vision algorithms and transfer learning methods, relying on large-scale general dental image datasets for model training. They have already been initially applied in standardized treatment scenarios in some large dental institutions.

[0003] However, existing automated dental X-ray diagnostic systems still face several technical limitations in their practical clinical application. Different brands and models of dental X-ray machines generate images with variations in grayscale distribution, noise characteristics, and artifact morphology due to differences in exposure parameters and detector types. Existing generalized normalization processes struggle to achieve accurate image standardization. Furthermore, differences in imaging specifications and image annotation standards among different medical institutions limit the generalization ability of models when deployed across institutions. The limited clinical sample size for rare conditions such as odontogenic jaw cysts and early-stage oral cancer, coupled with the lack of targeted small-sample optimization strategies in current technologies, hinders effective identification of these conditions. Additionally, most diagnostic models are large cloud-based models without lightweight adaptation, failing to meet the deployment needs of low-computing-power devices in primary dental clinics.

[0004] Therefore, there is an urgent need to develop an automated oral X-ray diagnostic system based on deep learning to solve the problems in the existing technology. Summary of the Invention

[0005] The purpose of this invention is to provide an automated oral X-ray diagnostic system based on deep learning, which can solve the problems of cross-device imaging deviation and cross-institutional domain offset, improve the accuracy of identifying rare oral diseases in small samples, and at the same time achieve model lightweighting to adapt to the deployment of low-computing-power equipment in primary medical institutions, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based automated oral X-ray diagnostic system includes: Device calibration module, feature alignment module, sample enhancement module, lightweight inference module; The device calibration module is used to acquire the input oral X-ray image, perform adaptive correction of grayscale, noise and artifacts by matching the preset device parameter library, and output a standardized image. By adopting the above technical solution, the output standardized image specifically refers to: The system acquires input oral X-ray images, extracts their image metadata and imaging features, and matches the corresponding calibration parameter set in the preset equipment parameter library based on the equipment brand, model identifier, and grayscale distribution characteristics. Based on the matched device grayscale offset range, dynamic grayscale normalization is performed on the input influence to map it to the standard grayscale range; Based on the noise model corresponding to the device detector type, a filtering algorithm is selected to suppress device-specific noise in the grayscale normalized image. Based on the contour detection algorithm, the metal artifact regions in the image are identified, and the neighborhood interpolation method is used to fill the artifact regions in a directional manner. Based on the anatomical features of the oral cavity, perspective transformation is performed on the images; The quality of the calibrated image is evaluated. If the evaluation result is lower than the preset threshold, the calibration parameters are adjusted and the calibration process is repeated until a standardized image that meets the quality standard is output.

[0007] As a further aspect of the present invention: the preset device parameter library is constructed as follows: Collect clinical image samples taken by dental X-ray equipment of multiple brands and models, and obtain their corresponding equipment factory parameters and image metadata; Statistical analysis was performed on the collected image samples to extract the inherent imaging features of each type of equipment image, including gray-scale distribution histogram, noise spectrum features, and common artifact morphology and location distribution. Based on the extracted imaging features, the deviation parameters of each type of device are quantitatively defined, including grayscale offset reference value, noise type identifier, noise intensity coefficient, artifact structure template and geometric distortion coefficient. For each type of device, based on its quantization deviation parameters and physical imaging principle, a corresponding correction algorithm or algorithm parameter set is designed, including dynamic grayscale mapping curve, adaptive filter type and kernel parameters, artifact region repair operator and geometric transformation matrix. The equipment model identifier, factory parameters, quantization deviation parameters and associated correction algorithm parameter sets are structured and stored to build a queryable equipment parameter knowledge base; Establish a parameter library update interface. When a new model of device is connected, the sample data and parameters of the new device are imported through this interface, and after analysis through the same process, they are expanded into the parameter library.

[0008] As a further aspect of the present invention: the quality assessment of the rectified image specifically involves: Define a multidimensional quantitative index for image quality assessment, which includes: global sharpness score, local artifact retention index and regional grayscale uniformity coefficient; Calculate the global sharpness score of the image: perform edge detection on the image, calculate the average gradient magnitude of significant edges, and normalize it to a sharpness score within a preset range; Calculate the local artifact retention index of the image: In the expected area where artifact correction has been completed, the statistical difference between its pixel value and the pixel value of the surrounding normal tissue area is statistically analyzed, and the artifact retention index is calculated in combination with the sharpness of the contour of the area. Calculate the regional gray-level uniformity coefficient of the image: Divide the image into multiple non-overlapping regions of interest, calculate the gray-level variance within each region, and calculate the dispersion of the gray-level variance between all regions to evaluate the overall gray-level uniformity. The sharpness score, artifact retention index, and grayscale uniformity coefficient are weighted and fused to generate a comprehensive quality score.

[0009] The feature alignment module is electrically connected to the device calibration module and is used to receive the standardized image and use an adversarial training method to align its features with the image features of the target medical institution, and output a domain-adapted feature representation. By adopting the above technical solution, the working method of the feature alignment module is as follows: Define a source domain dataset and a target domain dataset. The source domain dataset consists of large-scale standardized oral X-ray images from several institutions, and the target domain dataset consists of standardized oral X-ray images from a specific target medical institution. Configure a domain-adaptive network structure, which includes a shared image feature extraction backbone network and a parallel domain discriminator network; The source domain image samples and the target domain image samples are input into the image feature extraction backbone network to extract the source domain features and the target domain features. The extracted source domain features and target domain features are simultaneously input into the domain discriminator network, which outputs the probability distribution of each feature vector belonging to the source domain or the target domain.

[0010] As a further aspect of the present invention: the adversarial training process of the feature alignment module is specifically as follows: The domain classification loss is calculated based on the output probability of the domain discriminator; the original diagnostic task loss of the source domain image is calculated based on the features extracted by the feature extraction backbone network. The parameters of the domain discriminator network are updated with the goal of minimizing the domain classification loss. At the same time, the parameters of the image feature extraction backbone network are updated with the goal of minimizing the original diagnostic task loss and maximizing the classification error of the domain discriminator. The process of feature extraction, domain discrimination, loss calculation and parameter update is executed iteratively until the classification accuracy of the domain discriminator for source domain features and target domain features is close to the level of random guessing. At this point, it is determined that the feature distribution has been aligned. Once the trained image feature extraction backbone network is fixed, its output feature vector is the domain-adaptive feature representation aligned with the image feature distribution of the target medical institution.

[0011] The sample enhancement module is electrically connected to the feature alignment module. It is used to expand the training sample set based on the aligned feature representation for rare diseases by using generative adversarial network synthesis and oversampling methods, and output enhanced disease sample data. By adopting the above technical solution, the method of expanding the training sample set by generating adversarial network synthesis in the sample enhancement module is as follows: Obtain a small number of real oral X-ray images of the target rare disease as input for training the generative adversarial network; A generative adversarial network is constructed, which includes a generator and a discriminator, wherein the generator is used to receive a random noise vector and synthesize a simulated oral X-ray image, and the discriminator is used to distinguish whether the input image is a real sample or a generated sample. The real samples and random noise vectors are input into the generative adversarial network for adversarial training. The training objective is to generate images that deceive the discriminator and improve the discriminator's ability to distinguish between real and fake images. During adversarial training, spatial prior constraints of oral anatomy are injected into the generator. When the adversarial training reaches a preset equilibrium state, the generator network parameters are fixed, and the generator is used to generate high-fidelity synthetic images that are consistent with the characteristics of the target rare disease in batches. The synthetic images are then added to the original training sample set.

[0012] As a further aspect of the present invention: the method for expanding the training sample set through oversampling in the sample enhancement module is as follows: Feature space mapping is performed on the augmented dataset containing real samples and synthetic images to extract the feature vector of each sample; The SMOTE oversampling algorithm is used to interpolate rare disease samples of a minority class in the feature space to generate a new synthetic sample feature vector between any two minority class samples. The newly generated synthetic sample feature vectors are back-mapped to the image space to obtain the newly added oversampled image samples; During the model training phase, the Focal Loss function is applied to assign loss weights to rare disease samples and common disease samples respectively, requiring that the loss weight of rare disease samples be higher than that of common disease samples. The final output is a dataset of enhanced disease samples containing the original real samples, generative adversarial network synthesized images, and oversampled generated images.

[0013] The lightweight inference module is electrically connected to the sample augmentation module and is used to receive the augmented sample data, perform depthwise separable convolution replacement and attention sparsification compression on the diagnostic model, and output a lightweight diagnostic model.

[0014] By adopting the above technical solution, the method for performing depthwise separable convolutional replacement on the diagnostic model in the lightweight inference module is as follows: Identify all standard convolutional layers in the diagnostic model to be compressed, wherein the standard convolutional layers simultaneously perform spatial feature extraction and channel feature fusion; The standard convolutional layer is replaced with a depthwise separable convolutional structure, which consists of a depthwise convolutional layer and a pointwise convolutional layer connected in series. The depthwise convolutional layer performs spatial convolution independently on each input channel, and the pointwise convolutional layer completes channel feature fusion through 1×1 convolution. During the replacement process, the output feature map size and number of channels of the depth-separable convolutional structure are kept consistent with those of the original standard convolutional layer; The convolutional layer responsible for extracting core lesion features from oral X-ray images in the replaced network is retained without being replaced by lightweighting. The core lesion features include low-density shadow areas, bone destruction areas, and periodontal ligament space morphology.

[0015] As a further aspect of the present invention: the method for attention sparsity compression of the diagnostic model in the lightweight inference module is specifically as follows: Identify all attention-based modules in the diagnostic model to be compressed; The importance of multiple attention heads in each attention module is evaluated, and the feature contribution score of each attention head in the oral X-ray imaging diagnostic task is calculated. Based on the feature contribution score, the attention heads are sorted, and redundant attention heads with contribution scores below a preset threshold are removed. The attention weight matrix inside the retained attention head is approximated by low rank or pruned in a structured manner to further reduce its number of parameters and computational complexity. After completing depthwise separable convolution replacement and attention sparsification compression, the lightweight diagnostic model is fine-tuned by knowledge distillation to restore the diagnostic accuracy lost due to compression, and finally outputs a lightweight diagnostic model that can run on low-computing-power devices.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: Through device-specific correction, it achieves device-specific adaptive correction of grayscale, noise, and artifacts based on a preset device parameter library. Combined with a multi-dimensional quality assessment closed loop of global sharpness, artifact retention index, and grayscale uniformity, it effectively overcomes the differences between different imaging devices and ensures the standardized quality of input images. Employing a domain-adaptive method, it achieves alignment of source and target domain feature distributions through adversarial training of a shared feature extraction backbone network and a domain discriminator, enhancing the model's generalization ability across different medical institutions and reducing the adaptation cost of cross-institutional deployment. Addressing the challenge of insufficient rare disease samples, the system combines generative adversarial network synthesis and oversampling strategies, introducing prior constraints on oral anatomical structure space and Focal... The loss-weighted strategy improves the model's accuracy in identifying difficult diseases without relying on a large number of real cases, and expands the range of diseases covered by auxiliary diagnosis. By designing the model in a lightweight manner, and adopting a combination of strategies such as replacing non-core convolutional layers with depthwise separable convolutions, attention sparsity compression, and knowledge distillation fine-tuning, the system maintains high diagnostic performance while significantly reducing its dependence on hardware computing power, making it suitable for deployment needs of low-computing-power equipment in primary healthcare institutions.

[0017] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0018] Figure 1 This is a system module diagram of the present invention. Detailed Implementation

[0019] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This system, built upon deep learning technology, achieves fully automated and accurate diagnosis of oral X-ray images through the synergistic interaction of four modules: device calibration, feature alignment, sample enhancement, and lightweight inference. The overall workflow is as follows: First, the device calibration module receives clinical oral X-ray images, matches them to a pre-defined device parameter library to perform grayscale normalization, noise suppression, artifact repair, and fluoroscopic transformation, and outputs standardized images through multi-dimensional quality assessment, providing consistent quality input data for subsequent diagnosis. Second, the feature alignment module employs a domain-adaptive adversarial training architecture to align standardized image features with those of the target medical institution, eliminating domain shift issues caused by differences in cross-institutional imaging standards. Finally, the sample enhancement module addresses the scarcity of samples for rare diseases by combining generative adversarial networks (GANs) with SMOTE (Single Motion Elevation Model). The oversampling algorithm expands the training samples and enhances the model's ability to learn features of difficult diseases through a loss weighting strategy. Finally, the lightweight inference module performs deep separable convolution replacement and attention sparsification compression on the trained diagnostic model, and then fine-tunes it through knowledge distillation to restore diagnostic accuracy. The output is a lightweight model that can be deployed on low-computing-power devices at the grassroots level, ultimately achieving efficient and accurate automatic diagnosis of oral diseases across devices and institutions.

[0021] In this embodiment of the invention, a deep learning-based automated oral X-ray diagnostic system is described, see [link to relevant documentation]. Figure 1 As shown, an automated oral X-ray diagnostic system based on deep learning includes: an equipment calibration module, a feature alignment module, a sample augmentation module, and a lightweight inference module; In this embodiment, the device calibration module is used to acquire the input oral X-ray image, perform adaptive correction of grayscale, noise and artifacts by matching it with a preset device parameter library, and output a standardized image; The output standardized image is specifically: The system acquires the input oral X-ray image, extracts the metadata and imaging features of the oral X-ray image, and matches the calibration parameter set of the corresponding device in the preset device parameter library based on the device brand, model identifier and grayscale distribution features. Based on the matched device grayscale offset range, dynamic grayscale normalization is performed on the input image, mapping it to a standard grayscale range, which is specifically set as follows: This range conforms to the general standard of 8-bit grayscale images, which can balance the preservation of image details and the computational efficiency of subsequent feature extraction. The filtering algorithm is selected based on the noise model corresponding to the detector type. The selection criteria are as follows: when the detector is a CCD (Charge-Coupled Device), its noise model is mainly Gaussian noise, so a Gaussian filtering algorithm is selected; when the detector is a CMOS (Complementary Metal-Oxide-Semiconductor) detector, its noise model contains a mixture of salt-and-pepper noise and Gaussian noise, so a bilateral filtering algorithm is selected; when the detector is an amorphous silicon planar detector, the noise is mainly Poisson noise, so a Wiener filtering algorithm is selected. The kernel parameters are set based on the noise intensity coefficient. The kernel size of the Gaussian filter is set to ,in ( (This indicates rounding operations) The value ranges from 0.5 to 2.0, and is determined by the noise intensity coefficient of the device in the preset device parameter library; The spatial kernel parameters of bilateral filtering are the same as those of Gaussian filtering, while the similarity kernel parameters are based on the noise intensity coefficient. Set as The kernel size of the Wiener filter is set to... The filter coefficients are dynamically adjusted based on the ratio of noise intensity coefficient to signal intensity, with a value range of 0.01 to 0.1. Based on the contour detection algorithm, metallic artifact regions in images are identified. Neighborhood interpolation is used to directionally fill these artifact regions. The contour detection algorithm employs the Canny edge detection operator, with the following parameter configuration: Gaussian smoothing kernel size is set to... Standard deviation The low threshold is set to 50, and the high threshold is set to 150. An 8-neighborhood connection is used for edge connectivity. This operator extracts edge contours with significant gray-level abrupt changes in the image. Combined with the high gray-level and irregular contour features of metal artifacts, the contour boundaries of metal artifact regions are selected. The neighborhood interpolation method uses bilinear interpolation, with the following mathematical form: Let the coordinates of the pixel to be filled within the artifact region be... The coordinates of its four non-artifact neighboring pixels are respectively , , , The corresponding pixel values ​​are respectively , , , ,in , , , ( (This indicates a floor operation), the pixel value to be filled. The formula for calculation is: ; Based on the anatomical features of the oral cavity, perspective transformation was performed on the images. Control points were selected from key landmarks of the oral cavity anatomy, specifically six core control points: the incisal edge of the maxillary central incisor, the cusp of the maxillary first molar, the incisal edge of the mandibular central incisor, the cusp of the mandibular first molar, and the roots of the zygomatic processes of the left and right maxillae. The coordinates of these six control points were accurately located using contour detection and anatomical feature matching algorithms. The perspective transformation matrix was calculated as follows: Let the set of coordinates of the six control points in the original image be... ( (For control point index), the set of coordinates of the corresponding control points in the target standard image is: ( (for control point indexes), perspective transformation matrix for A matrix, in the form of: Satisfy perspective transformation relationship , The perspective transformation matrix is ​​obtained by solving the above system of equations using the least squares method. The element values ​​are calculated using L2 regularization to avoid matrix overfitting. The regularization coefficient is set to 0.001. After the solution is completed, all pixels in the original corrected image are processed through a matrix. Perform perspective transformation to obtain an image that conforms to the standards of oral anatomical structure; The quality of the calibrated image is evaluated. If the evaluation result is lower than the preset threshold, the calibration parameters are adjusted and the calibration process is repeated until a standardized image that meets the quality standard is output.

[0022] The preset device parameter library is constructed as follows: Collect clinical image samples taken by dental X-ray equipment of multiple brands and models, and obtain their corresponding equipment factory parameters and image metadata; Statistical analysis was performed on the collected image samples to extract the inherent imaging features of each type of equipment image, including gray-scale distribution histogram, noise spectrum features, and common artifact morphology and location distribution.

[0023] The method for generating a grayscale distribution histogram is as follows: Count the frequency of grayscale values ​​for all pixels in the image, assuming the total number of pixels in the image is... The grayscale range is For each gray level ( ), statistical grayscale value equals Number of pixels Then the frequency function of the gray-level distribution histogram is: This function describes the grayscale distribution characteristics of the image; the noise spectrum characteristic analysis technique uses Fourier transform to analyze the image. Perform a two-dimensional discrete Fourier transform to obtain the frequency domain image. The transformation formula is: in , , , , These are the width and height of the image, respectively. The imaginary unit is used to calculate the frequency domain image. amplitude spectrum The distribution characteristics of noise in different frequency ranges are analyzed. The low frequency range corresponds to the background information of the image, and the high frequency range corresponds to the noise information. By extracting parameters such as the peak value and variance of the amplitude spectrum, the spectral characteristics of the noise are quantified. The extraction of common artifact morphology and location distribution adopts a method combining morphological processing and statistical analysis. First, the artifact region is extracted by threshold segmentation and edge detection algorithm. Then, morphological dilation and erosion operations are used to remove small noise in the artifact region and extract the contour features of the artifact (such as area, perimeter, and roundness). At the same time, the coordinate distribution of the artifact region in all samples is statistically analyzed to obtain the location distribution pattern of the artifact. Based on the extracted imaging features, deviation parameters for each type of device are quantitatively defined, including grayscale offset reference value, noise type identifier, noise intensity coefficient, artifact structure template, and geometric distortion coefficient. The grayscale offset reference value is derived from image metadata, and the derivation formula is as follows: in The maximum grayscale value in the image metadata. The minimum grayscale value, The standard deviation of grayscale This refers to the grayscale offset baseline value, used to describe the baseline level of the overall grayscale offset of the image; the noise type identifier is determined based on the noise spectrum characteristics and image grayscale distribution, using a binary encoding method: 00 represents Gaussian noise, 01 represents salt-and-pepper noise, 10 represents Poisson noise, and 11 represents mixed noise. The corresponding noise type identifier is determined by matching the noise spectrum characteristics with the feature templates of preset noise types; the noise intensity coefficient is defined as the ratio of noise variance to signal variance, calculated as follows: in The gray-level variance of the noise region. The grayscale variance of the normal tissue region. This is the noise intensity coefficient, which ranges from 0 to 1. The larger the value, the stronger the noise. The process of creating an artifact structure template is as follows: Select the contour features of typical artifact regions in the image of this type of equipment, and normalize them to a uniform size (e.g., ...). (pixels), extract the binary feature matrix of the contour. This matrix is ​​the artifact structure template. In the matrix, 1 represents the artifact region and 0 represents the non-artifact region. Artifacts in images from similar devices can be quickly identified through template matching. The calibration steps for geometric distortion coefficients are as follows: Select a standard calibration board (with a grid pattern of known size), use the device to capture an image of the calibration board, extract the actual coordinates and standard coordinates of the calibration board grid, calculate the coordinate deviation of each grid point, and let the standard coordinates be... The actual coordinates are Then the distortion deviation of a single grid point is , Geometric distortion coefficient The formula for calculation is: in The geometric distortion coefficient of the device is obtained by averaging the distortion coefficients of all grid points, using the standard grid spacing. , used to describe the degree of geometric distortion in an image; For each type of device, based on its quantization deviation parameters and physical imaging principle, a corresponding correction algorithm or algorithm parameter set is designed, including dynamic grayscale mapping curve, adaptive filter type and kernel parameters, artifact region repair operator and geometric transformation matrix. The equipment model identifier, factory parameters, quantization deviation parameters and associated correction algorithm parameter sets are structured and stored to build a queryable equipment parameter knowledge base; Establish a parameter library update interface. When a new model of device is connected, the sample data and parameters of the new device are imported through this interface, and after analysis through the same process, they are expanded into the parameter library.

[0024] The quality assessment of the rectified image specifically involves: Define a multidimensional quantitative index for image quality assessment, which includes: global sharpness score, local artifact retention index and regional grayscale uniformity coefficient; Calculate the global sharpness score of the image: Perform edge detection on the image using the Sobel edge detection operator, and calculate the gradient magnitudes in the x and y directions respectively. Directional gradient , Directional gradient pixel The gradient magnitude is The average gradient magnitude of all significant edges is calculated. The criteria for determining significant edges are as follows: ,in For edge threshold, Let the total number of pixels at the significant edge be . The average gradient magnitude is The resolution score is normalized to a preset range using a linear normalization method. The calculation formula is as follows: in It is the minimum of the mean magnitude of significant marginal gradients across all samples. The maximum value, This is the global sharpness score, which ranges from 0 to 100. The higher the score, the sharper the image. Calculate the local artifact retention index of the image: In the expected area where artifact correction has been completed, statistically analyze the difference between its pixel value and the pixel value of the surrounding normal tissue area, and combine it with the sharpness of the contour of the area to calculate the artifact retention index.

[0025] The statistical difference measure uses the variance ratio method, assuming the pixel set of the expected region for artifact correction is . The set of pixels in the surrounding normal tissue area is Calculate the set gray variance and set gray variance The statistical difference measure is The sharpness of the region contour is represented by the standard deviation of the contour gradient magnitude. Let the set of contour pixels in the expected region for artifact correction be... Calculate the gradient magnitude of each pixel in the set. ( ), Sharpness of the outline ( (This indicates the calculation of standard deviation). The formula for calculating the local artifact persistence index is: in This is the weighting coefficient, with a value of 0.6. This represents the maximum sharpness of the contour across all samples. The value ranges from 0 to 1, and the closer the value is to 0, the less artifact residue there is; Calculate the regional grayscale uniformity coefficient of the image: Divide the image into multiple non-overlapping regions of interest, and set the size of the regions of interest to be... If the image size is not divisible by 64, zero-padding is performed at the image edges. Let the total number of regions of interest after division be... Calculate the gray-level variance within each region. ( , (For the region of interest index), the dispersion of gray-level variance among all regions is represented by the coefficient of variation, calculated as follows: ,in The standard deviation of the grayscale variance for each region is denoted as . Let be the mean of the gray-level variance of each region. The formula for calculating the regional gray-level uniformity coefficient is: , The value ranges from 0 to 1, and the closer the value is to 1, the better the uniformity of image grayscale. The sharpness score, artifact retention index, and grayscale uniformity coefficient are weighted and fused to generate a comprehensive quality score. The weights in the weighting fusion are determined using the analytic hierarchy process (AHP), taking into account clinical diagnostic needs, and the weights of the global sharpness score are calculated accordingly. Weight of local artifact persistence index Weight of the regional gray-scale uniformity coefficient The weights satisfy The formula for calculating the overall quality score is as follows: , The value range is 0~100, and the preset quality assessment threshold is 80. If the image meets the quality standards, it is determined that the image meets the quality standards; otherwise, the correction parameters need to be adjusted and the image needs to be recalibrated. In this embodiment, the feature alignment module is electrically connected to the device calibration module, and is used to receive the standardized image, and use an adversarial training method to align its features with the image features of the target medical institution, and output a domain-adapted feature representation. The specific working method of the feature alignment module is as follows: Define a source domain dataset and a target domain dataset. The source domain dataset consists of large-scale standardized oral X-ray images from several institutions, and the target domain dataset consists of standardized oral X-ray images from a specific target medical institution. A domain-adaptive network architecture is configured, comprising a shared image feature extraction backbone network and a parallel domain discriminator network. The image feature extraction backbone network adopts a lightweight ResNet (residual network) 18 architecture, with the following convolutional layer configuration: The input layer receives standardized dental X-ray images, with a size of [size missing]. ; The first convolutional layer (Conv1) uses 64... A convolutional kernel of size [size missing], stride of 2, padding type [same], output feature map size [size missing]. ; Then, four residual blocks (ResBlock1-ResBlock4) are set up, each containing two convolutional layers with kernel sizes of [missing value]. The step size is adjusted according to the position of the residual block. The step size of ResBlock1 and ResBlock2 is 1, and the output feature map size remains unchanged. The step size of ResBlock3 and ResBlock4 is 2, and the output feature map size is halved. Each convolutional layer is followed by a batch normalization (BN) layer, using ReLU as the activation function. The activation function expression is as follows: ; Finally, the feature map is transformed using a global average pooling (GAP) layer. The feature vectors are used as the feature representation of the image. The domain discriminator network adopts a fully connected layer design, with the following specific structure: The input layer receives the 512-dimensional feature vector output by the backbone network; The first fully connected layer (FC1) has 256 nodes. The activation function is LeakyReLU, with a leakage coefficient set to 0.2. The activation function expression is: ; The second fully connected layer (FC2) has 128 nodes, and the activation function is also LeakyReLU; The third fully connected layer (FC3) has 1 node, and the activation function is Sigmoid. The activation function expression is: The output is a probability value between 0 and 1, which is used to determine whether the input feature vector belongs to the source domain or the target domain. The source domain image samples and the target domain image samples are input into the image feature extraction backbone network to extract the source domain features and the target domain features. The extracted source domain features and target domain features are simultaneously input into the domain discriminator network, which outputs the probability distribution of each feature vector belonging to the source domain or the target domain.

[0026] The adversarial training process of the feature alignment module is as follows: The domain classification loss is calculated based on the output probability of the domain discriminator; the original diagnostic task loss of the source domain image is calculated based on the features extracted by the feature extraction backbone network. The domain classification loss adopts the cross-entropy loss function, specifically in the form of: in For domain labels, when the input features are source domain features, When the input features are target domain features, , The loss function is used to measure the classification accuracy of the domain discriminator, where the output probability is the domain discriminator. The original diagnostic task loss uses the cross-entropy loss function, which is suitable for multi-class diagnostic tasks on oral X-ray images (such as normal, caries, periodontitis, periapical periodontitis), and its specific form is as follows: ,in For category indexing, For the first One-hot encoding of diagnostic labels, For the model to predict the first The probability of a class, this loss function is used to measure the diagnostic effectiveness of the features extracted by the backbone network; The parameters of the domain discriminator network are updated with the goal of minimizing the domain classification loss. At the same time, the parameters of the image feature extraction backbone network are updated with the goal of minimizing the original diagnostic task loss and maximizing the classification error of the domain discriminator. The process of feature extraction, domain discrimination, loss calculation and parameter update is executed iteratively until the classification accuracy of the domain discriminator for source domain features and target domain features is close to the level of random guessing. At this point, it is determined that the feature distribution has been aligned.

[0027] The stopping conditions for training iterations include two aspects: first, the number of iterations reaches a preset maximum value, which is set to 5000; second, the loss of the original diagnostic task on the validation set no longer decreases after 50 consecutive iterations, and the decrease is less than [a certain value]. The iteration stops when any one of the conditions is met. The specific threshold for the domain discriminator's classification accuracy to approach the level of random guessing is defined as 45%~55%. When the classification accuracy of the domain discriminator is within this threshold range, it indicates that the distribution of source domain features and target domain features is basically consistent, and the domain discriminator cannot effectively distinguish them. At this time, feature alignment is completed. Once the trained image feature extraction backbone network is fixed, its output feature vector is the domain-adaptive feature representation aligned with the image feature distribution of the target medical institution.

[0028] In this embodiment, the sample enhancement module is electrically connected to the feature alignment module, and is used to expand the training sample set based on the aligned feature representation for rare diseases by using generative adversarial network synthesis and oversampling methods, and output enhanced disease sample data. The sample augmentation module uses a generative adversarial network synthesis method to expand the training sample set as follows: A small number of real oral X-ray images of the target rare disease are obtained as input for training the generative adversarial network; A generative adversarial network is constructed, comprising a generator and a discriminator, wherein the generator is used to receive a random noise vector and synthesize a simulated oral X-ray image, and the discriminator is used to distinguish whether the input image is a real sample or a generated sample.

[0029] The real samples and random noise vectors are input into the generative adversarial network for adversarial training. The training objective is to generate images that deceive the discriminator and improve the discriminator's ability to distinguish between real and fake images.

[0030] During adversarial training, spatial prior constraints on oral anatomy are injected into the generator. The injection mechanism is as follows: a standard oral anatomy atlas is pre-acquired and converted into a grayscale image with the same size as the training samples. Contour features of the anatomical structures are extracted using a contour extraction algorithm to obtain an anatomical structure mask matrix. In the mask matrix, 1 represents the anatomical structure region and 0 represents the non-anatomical structure region. This mask matrix is ​​used as an additional input to the generator, concatenated with a random noise vector, and then input into the generator. Simultaneously, a contour matching loss is added to the generator's loss function. The contour matching loss is calculated using mean squared error (MSE), and the formula is: in To generate the contour mask matrix of the image, The contour mask matrix of the standard anatomical atlas, The loss function is used to guide the generator to produce images that conform to the oral anatomy, thus avoiding anatomical disorganization in the generated images. Once the adversarial training reaches a preset equilibrium state, the criteria for determining the preset equilibrium state are that both the generator loss and the discriminator loss tend to stabilize, and the fluctuation amplitude is less than [a certain value]. Meanwhile, the structural similarity index (SSIM) of the generated images is greater than 0.85 compared to the average value of the real samples. The generator network parameters are fixed, and the generator is used to generate high-fidelity synthetic images that are consistent with the characteristics of the target rare disease in batches. The synthetic images are then added to the original training sample set. In the sample augmentation module, the method for expanding the training sample set through oversampling is as follows: Feature space mapping is performed on the augmented dataset containing real samples and synthetic images to extract the feature vector of each sample; The SMOTE oversampling algorithm is used to interpolate rare disease samples of a minority class in the feature space, generating a new synthetic sample feature vector between any two minority class samples.

[0031] The SMOTE oversampling algorithm employs a k-nearest neighbor (k-NN) algorithm for selecting neighbor samples during interpolation in the feature space. The value is set to 5 for each minority class sample. In the feature space, find the five nearest minority class samples to form a neighbor sample set. ; The distance is calculated using Euclidean distance, and the formula is: in For feature dimension indexing, Samples, In the The values ​​are taken on the dimensional features; the interpolation ratio is set based on the ratio of the number of minority class samples to the number of majority class samples, let the number of minority class samples be . The number of samples in the majority class is Interpolation ratio This ensures that the number of minority class samples is consistent with the number of majority class samples after oversampling; the specific formula for the interpolation process is as follows: ,in for Neighbor samples, The formula generates a new sample feature vector using random numbers between 0 and 1. ; The newly generated synthetic sample feature vectors are back-mapped to the image space to obtain the newly added oversampled image samples; During the model training phase, the Focal Loss function is applied to assign loss weights to rare disease samples and common disease samples respectively, requiring that the loss weight of rare disease samples be higher than that of common disease samples.

[0032] The specific calculation rule for weight allocation in the Focal Loss function is as follows: Let the total number of samples in the dataset be... The number of rare disease samples is The number of samples for common diseases is , First, calculate the weighting factors for each type of sample, and the weighting factors for rare disease samples. Weighting factors for common disease samples The weighting factors satisfy This ensures that rare disease samples receive higher loss weights; the specific form of the Focal Loss function is: in The weighting factor corresponding to the sample ( or ), To predict the probability of this sample class for the model, To focus the parameters, we set them to 2, which reduces the loss weight of easily classified samples and further increases the loss ratio of rare disease samples (difficult-to-classify samples). The final output is a dataset of enhanced disease samples containing the original real samples, generative adversarial network synthesized images, and oversampled generated images.

[0033] In this embodiment, the lightweight inference module is electrically connected to the sample augmentation module and is used to receive the augmented sample data, perform depthwise separable convolution replacement and attention sparsification compression on the diagnostic model, and output a lightweight diagnostic model.

[0034] The method for performing depthwise separable convolutional replacement on the diagnostic model in the lightweight inference module is as follows: Identify all standard convolutional layers in the diagnostic model to be compressed, wherein the standard convolutional layers simultaneously perform spatial feature extraction and channel feature fusion; The standard convolutional layer is replaced with a depthwise separable convolutional structure, which consists of a depthwise convolutional layer and a pointwise convolutional layer connected in series. The depthwise convolutional layer performs spatial convolution independently on each input channel, and the pointwise convolutional layer... Convolution completes the fusion of channel features; During the replacement process, the output feature map size and number of channels of the depth-separable convolutional structure are kept consistent with those of the original standard convolutional layer; The convolutional layer responsible for extracting core lesion features from oral X-ray images in the replaced network is retained without being replaced by lightweighting. The core lesion features include low-density shadow areas, bone destruction areas, and periodontal ligament space morphology.

[0035] The specific criteria for retaining convolutional layers during depthwise separable convolutional replacement are as follows: Based on the importance assessment results of lesion features, a feature contribution scoring method is used to calculate the contribution of each convolutional layer's output feature to the lesion diagnosis result. The calculation process for the contribution score is as follows: input the convolutional layer's output feature into the classifier of the diagnostic model, calculate the class activation map (CAM) corresponding to the feature, and extract the mean of the class activation maps as the feature contribution score. The contribution threshold is set at 0.7. When the convolutional layer is determined to be responsible for extracting core lesion features, it is preserved and not replaced; when At that time, it is determined that the convolutional layer is responsible for extracting non-core features, and depthwise separable convolutional replacement is performed; the mean of the class activation map is calculated as follows: ,in Activation mapping for class at pixel The value, The size of the activation map for the class; The method for attention sparsity compression of the diagnostic model in the lightweight inference module is as follows: Identify all attention-based modules in the diagnostic model to be compressed; The importance of multiple attention heads within each attention module is evaluated, and the feature contribution score of each attention head in the oral X-ray imaging diagnostic task is calculated. The specific method for importance evaluation in attention sparsity compression involves the following calculation process for the attention head feature contribution score: For each attention head, its output feature vector is extracted, and correlation analysis is performed between this feature vector and the diagnostic label to calculate the Pearson correlation coefficient. Simultaneously calculate the variance of the output features of the attention head. The formula for calculating the feature contribution score is as follows: ,in This represents the maximum value of the variance of the output features of all attention heads in this attention module. The value ranges from 0 to 1 and is used to measure the importance of attention. Based on the feature contribution score, the attention heads are sorted, and redundant attention heads with contribution scores lower than a preset threshold are removed. The preset threshold is set to 0.3. The attention weight matrix within the retained attention head is further reduced by performing low-rank approximation or structured pruning to decrease the number of parameters and computational complexity. The specific steps for low-rank approximation are as follows: For the attention weight matrix... ( (for the feature dimension of the attention head), for the matrix Perform singular value decomposition (SVD), the decomposition formula is as follows: ,in It is a left singular matrix. It is a singular value diagonal matrix. It is a right singular matrix. The number of singular values ​​to be retained. The value is That is, retaining the first 50% of the singular values, through Reconstruct the attention weight matrix, where , , They are respectively , , The former Columns / rows, the dimensions of the reconstructed matrix are The number of parameters is reduced by 50%; the specific implementation steps of structured pruning are as follows: for the attention weight matrix Set pruning threshold ,in The weight matrix is ​​the mean of the absolute values ​​of all elements. Elements with absolute values ​​less than [a certain value] are then grouped into the matrix. The elements are set to 0 to form a sparse matrix. At the same time, rows and columns with all zeros are removed from the weight matrix to further compress the number of parameters. After pruning, the sparsity of the weight matrix is ​​retained to be 50%~60%. After completing the depthwise separable convolution replacement and attention sparsity compression, the lightweight diagnostic model is fine-tuned using knowledge distillation to restore the diagnostic accuracy lost due to compression. The final output is a lightweight diagnostic model that can run on low-computing devices. In the knowledge distillation fine-tuning, the teacher model is derived from the original diagnostic model without any compression. This model has been fully trained and possesses high diagnostic accuracy. The student model is initialized as follows: the lightweight model after depthwise separable convolution replacement and attention sparsity compression is used as the student model, and its initial parameters are inherited from the parameters after compression. For the retained convolutional layers and attention heads, the parameters remain unchanged. For the replaced depthwise separable convolutional layers, the parameters are randomly initialized and then fine-tuned. The distillation loss function is a weighted sum of the basic loss and the distillation loss, specifically in the following form: in The cross-entropy loss is the same as the loss for the original diagnostic task in the feature alignment module. The basic loss weight is set to 0.4; For distillation losses, KL divergence is used for calculation, and the formula is: in The softening probability distribution output by the teacher model. The softening probability distribution output by the student model. The weighting for distillation loss is set to 0.6; temperature parameter The value is set to 10, and the formula for calculating the softening probability distribution is: ,in The logits value output by the model is determined by the temperature parameter. The probability distribution is softened to better transfer knowledge from the teacher model to the student model; the fine-tuning process uses the Adam optimizer with a learning rate set to [value missing]. The number of iterations was set to 1000, and the batch size was set to 32, until the diagnostic accuracy of the student model reached more than 95% of the accuracy of the teacher model, and the fine-tuning was completed.

[0036] This invention provides an automated oral X-ray diagnostic system based on deep learning, which can solve the problems of cross-device imaging bias and cross-institutional domain offset, improve the accuracy of identifying rare oral diseases in small samples, and at the same time achieve model lightweighting to adapt to the deployment of low-computing-power equipment in primary medical institutions.

[0037] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0038] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A deep learning-based automated oral X-ray diagnostic system, characterized in that, include: Device calibration module, feature alignment module, sample enhancement module, lightweight inference module; The device calibration module is used to acquire the input oral X-ray image, perform adaptive correction of grayscale, noise and artifacts by matching the preset device parameter library, and output a standardized image. The feature alignment module is electrically connected to the device calibration module and is used to receive the standardized image and use an adversarial training method to align its features with the image features of the target medical institution, and output a domain-adapted feature representation. The sample enhancement module is electrically connected to the feature alignment module. It is used to expand the training sample set based on the aligned feature representation for rare diseases by using generative adversarial network synthesis and oversampling methods, and output enhanced disease sample data. The lightweight inference module is electrically connected to the sample augmentation module and is used to receive the augmented sample data, perform depthwise separable convolution replacement and attention sparsification compression on the diagnostic model, and output a lightweight diagnostic model.

2. The deep learning-based automated oral X-ray diagnostic system according to claim 1, characterized in that, The output standardized image is specifically: The system acquires input oral X-ray images, extracts their image metadata and imaging features, and matches the corresponding calibration parameter set in the preset equipment parameter library based on the equipment brand, model identifier, and grayscale distribution characteristics. Based on the matched device grayscale offset range, dynamic grayscale normalization is performed on the input influence to map it to the standard grayscale range; Based on the noise model corresponding to the device detector type, a filtering algorithm is selected to suppress device-specific noise in the grayscale normalized image. Based on the contour detection algorithm, the metal artifact regions in the image are identified, and the neighborhood interpolation method is used to fill the artifact regions in a directional manner. Based on the anatomical features of the oral cavity, perspective transformation is performed on the images; The quality of the calibrated image is evaluated. If the evaluation result is lower than the preset threshold, the calibration parameters are adjusted and the calibration process is repeated until a standardized image that meets the quality standard is output.

3. The deep learning-based automated oral X-ray diagnostic system according to claim 2, characterized in that, The preset device parameter library is constructed as follows: Collect clinical image samples taken by dental X-ray equipment of multiple brands and models, and obtain their corresponding equipment factory parameters and image metadata; Statistical analysis was performed on the collected image samples to extract the inherent imaging features of each type of equipment image, including gray-scale distribution histogram, noise spectrum features, and common artifact morphology and location distribution. Based on the extracted imaging features, the deviation parameters of each type of device are quantitatively defined, including grayscale offset reference value, noise type identifier, noise intensity coefficient, artifact structure template and geometric distortion coefficient. For each type of device, based on its quantization deviation parameters and physical imaging principle, a corresponding correction algorithm or algorithm parameter set is designed, including dynamic grayscale mapping curve, adaptive filter type and kernel parameters, artifact region repair operator and geometric transformation matrix. The equipment model identifier, factory parameters, quantization deviation parameters and associated correction algorithm parameter sets are structured and stored to build a queryable equipment parameter knowledge base; Establish a parameter library update interface. When a new model of device is connected, the sample data and parameters of the new device are imported through this interface, and after analysis through the same process, they are expanded into the parameter library.

4. The deep learning-based automated oral X-ray diagnostic system according to claim 3, characterized in that, The quality assessment of the rectified image specifically involves: Define a multidimensional quantitative index for image quality assessment, which includes: global sharpness score, local artifact retention index and regional grayscale uniformity coefficient; Calculate the global sharpness score of the image: perform edge detection on the image, calculate the average gradient magnitude of significant edges, and normalize it to a sharpness score within a preset range; Calculate the local artifact retention index of the image: In the expected area where artifact correction has been completed, the statistical difference between its pixel value and the pixel value of the surrounding normal tissue area is statistically analyzed, and the artifact retention index is calculated in combination with the sharpness of the contour of the area. Calculate the regional gray-level uniformity coefficient of the image: Divide the image into multiple non-overlapping regions of interest, calculate the gray-level variance within each region, and calculate the dispersion of the gray-level variance between all regions to evaluate the overall gray-level uniformity. The sharpness score, artifact retention index, and grayscale uniformity coefficient are weighted and fused to generate a comprehensive quality score.

5. The deep learning-based automated oral X-ray diagnostic system according to claim 4, characterized in that, The specific working method of the feature alignment module is as follows: Define a source domain dataset and a target domain dataset. The source domain dataset consists of large-scale standardized oral X-ray images from several institutions, and the target domain dataset consists of standardized oral X-ray images from a specific target medical institution. Configure a domain-adaptive network structure, which includes a shared image feature extraction backbone network and a parallel domain discriminator network; The source domain image samples and the target domain image samples are input into the image feature extraction backbone network to extract the source domain features and the target domain features. The extracted source domain features and target domain features are simultaneously input into the domain discriminator network, which outputs the probability distribution of each feature vector belonging to the source domain or the target domain.

6. The deep learning-based automated oral X-ray diagnostic system according to claim 5, characterized in that, The adversarial training process of the feature alignment module is as follows: The domain classification loss is calculated based on the output probability of the domain discriminator; the original diagnostic task loss of the source domain image is calculated based on the features extracted by the feature extraction backbone network. The parameters of the domain discriminator network are updated with the goal of minimizing the domain classification loss. At the same time, the parameters of the image feature extraction backbone network are updated with the goal of minimizing the original diagnostic task loss and maximizing the classification error of the domain discriminator. The process of feature extraction, domain discrimination, loss calculation and parameter update is executed iteratively until the classification accuracy of the domain discriminator for source domain features and target domain features is close to the level of random guessing. At this point, it is determined that the feature distribution has been aligned. Once the trained image feature extraction backbone network is fixed, its output feature vector is the domain-adaptive feature representation aligned with the image feature distribution of the target medical institution.

7. The deep learning-based automated oral X-ray diagnostic system according to claim 6, characterized in that, The sample augmentation module uses a generative adversarial network synthesis method to expand the training sample set as follows: Obtain a small number of real oral X-ray images of the target rare disease as input for training the generative adversarial network; A generative adversarial network is constructed, which includes a generator and a discriminator, wherein the generator is used to receive a random noise vector and synthesize a simulated oral X-ray image, and the discriminator is used to distinguish whether the input image is a real sample or a generated sample. The real samples and random noise vectors are input into the generative adversarial network for adversarial training. The training objective is to generate images that deceive the discriminator and improve the discriminator's ability to distinguish between real and fake images. During adversarial training, spatial prior constraints of oral anatomy are injected into the generator. When the adversarial training reaches a preset equilibrium state, the generator network parameters are fixed, and the generator is used to generate high-fidelity synthetic images that are consistent with the characteristics of the target rare disease in batches. The synthetic images are then added to the original training sample set.

8. The deep learning-based automated oral X-ray diagnostic system according to claim 7, characterized in that, In the sample augmentation module, the method for expanding the training sample set through oversampling is as follows: Feature space mapping is performed on the augmented dataset containing real samples and synthetic images to extract the feature vector of each sample; The SMOTE oversampling algorithm is used to interpolate rare disease samples of a minority class in the feature space to generate a new synthetic sample feature vector between any two minority class samples. The newly generated synthetic sample feature vectors are back-mapped to the image space to obtain the newly added oversampled image samples; During the model training phase, the Focal Loss function is applied to assign loss weights to rare disease samples and common disease samples respectively, requiring that the loss weight of rare disease samples be higher than that of common disease samples. The final output is a dataset of enhanced disease samples containing the original real samples, generative adversarial network synthesized images, and oversampled generated images.

9. The deep learning-based automated oral X-ray diagnostic system according to claim 8, characterized in that, The method for performing depthwise separable convolutional replacement on the diagnostic model in the lightweight inference module is as follows: Identify all standard convolutional layers in the diagnostic model to be compressed, wherein the standard convolutional layers simultaneously perform spatial feature extraction and channel feature fusion; The standard convolutional layer is replaced with a depthwise separable convolutional structure, which consists of a depthwise convolutional layer and a pointwise convolutional layer connected in series. The depthwise convolutional layer performs spatial convolution independently on each input channel, and the pointwise convolutional layer completes channel feature fusion through 1×1 convolution. During the replacement process, the output feature map size and number of channels of the depth-separable convolutional structure are kept consistent with those of the original standard convolutional layer; The convolutional layer responsible for extracting core lesion features from oral X-ray images in the replaced network is retained without being replaced by lightweighting. The core lesion features include low-density shadow areas, bone destruction areas, and periodontal ligament space morphology.

10. The deep learning-based automated oral X-ray diagnostic system according to claim 9, characterized in that, The method for attention sparsity compression of the diagnostic model in the lightweight inference module is as follows: Identify all attention-based modules in the diagnostic model to be compressed; The importance of multiple attention heads in each attention module is evaluated, and the feature contribution score of each attention head in the oral X-ray imaging diagnostic task is calculated. Based on the feature contribution score, the attention heads are sorted, and redundant attention heads with contribution scores below a preset threshold are removed. The attention weight matrix inside the retained attention head is approximated by low rank or pruned in a structured manner to further reduce its number of parameters and computational complexity. After completing depthwise separable convolution replacement and attention sparsification compression, the lightweight diagnostic model is fine-tuned by knowledge distillation to restore the diagnostic accuracy lost due to compression, and finally outputs a lightweight diagnostic model that can run on low-computing-power devices.

Images