An Oral Panoramic Film Image Data Enhancement Method and System

Abstract

The present invention discloses an oral panoramic film image data enhancement method and system. The method includes obtaining an oral panoramic image and performing preprocessing; constructing and training an oral panoramic film image data enhancement model, and inputting the preprocessed oral panoramic film image into the trained autoencoder to generate an enhanced image. The present invention effectively suppresses the interference of non-concerned regions, focuses on the alveolar bone region, more accurately extracts the alveolar bone region, avoids over-segmentation and misclassification, effectively strengthens the learning of key features, enhances the response of key regions, effectively eliminates the influence of artifacts, optimizes the reconstruction effect, thereby avoiding non-physiological tooth arrangements after model reconstruction, further enhancing the structural and hierarchical sense of the image, avoiding overfitting, and at the same time optimizing the regularization strength during training.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging, and particularly to a method and system for enhancing oral panoramic image data. Background Art

[0002] As a common imaging examination method, oral panoramic radiographs are widely used in the diagnosis and treatment planning of oral diseases. It can provide clear images of the entire oral region, including structures such as teeth, alveolar bone, and surrounding soft tissues. However, due to the high-dimensional characteristics and complex anatomical structures of oral panoramic images, traditional image processing techniques have great difficulties in extracting key anatomical features, removing noise, and dealing with metal artifacts. In addition, oral panoramic radiographs often contain many interfering elements, such as soft tissues, metal dentures, and other non-concerned areas, which seriously affect the accuracy of image processing and the accuracy of diagnosis.

[0003] The prior art fails to effectively distinguish anatomical structure features from noise, making it difficult to accurately extract features of key regions such as alveolar bone; there are significant differences in the effects when processing high-resolution and low-resolution images, resulting in information loss or distortion; it cannot effectively solve the problem of metal occlusion and distortion of anatomical structures, affecting the accuracy of diagnosis; geometric constraints are not considered during tooth reconstruction, and regularization methods often cannot be dynamically adjusted, resulting in overfitting problems in some cases and affecting the generalization ability of the model, etc. Summary of the Invention

[0004] (I) Technical Problems to be Solved

[0005] To solve the above technical problems, the present invention provides a method and system for enhancing oral panoramic image data.

[0006] (II) Technical Solutions

[0007] To solve the above existing technical problems and achieve the invention purpose, the present invention is realized through the following technical solutions:

[0008] A method for enhancing oral panoramic image data includes the following steps:

[0009] S1. Image data acquisition and preprocessing, including acquiring oral panoramic images using high-resolution digital imaging devices and performing preprocessing.

[0010] S2. Construction and training of an oral panoramic radiograph image data enhancement model, including performing hash function transformation on images of different scales and cross-scale feature fusion, strengthening key feature capture, performing dynamic local calibration mapping, achieving noise suppression and feature enhancement through an activation function with anatomical feature perception, elastic gradient update, manifold-constrained feature fusion, separating mineralized tissue and soft tissue features through a dual-channel gating mechanism, dynamically adjusting the regularization strength according to the sparsity of the feature space, and iteration;

[0011] S3. Enhancement of oral panoramic radiograph image data. Input the preprocessed oral panoramic radiograph image into the trained autoencoder to generate an enhanced image.

[0012] Furthermore, the performing of hash function transformation on images of different scales and cross-scale feature fusion includes:

[0013] Construct a multi-scale feature matrix, separate structural features and noise features. After the input data is geometrically normalized, construct an initial feature space through local sensitive hashing mapping, aggregate similar anatomical structure features according to the hash collision principle, suppress the interference of non-attended regions, decompose the input image into different scales, perform differential hash function transformation on images of different scales, and perform cross-scale feature fusion.

[0014] Furthermore, it also includes performing regional growth operation on the alveolar bone area mask. Starting from the seed point, i.e., the alveolar crest vertex, expand the region according to the gradient constraint condition, expressed as:

[0015]

[0016] , is the adjacent pixel within the grown region; represents the interaction operation; is the pixel 's gradient magnitude, calculated by the Sobel operator, with the unit of HU / mm; is the pixel 's gradient magnitude; is the gradient difference threshold, taking 15 HU / mm, restricting the growth region within the anatomically reasonable range.

[0017] Furthermore, the strengthening of key feature capture includes designing an anisotropic weight initialization strategy based on the density distribution characteristics of dental tissues. Larger initial weights are obtained in high-density regions, and the weights decay in low-density regions.

[0018] Furthermore, the dynamic local calibration mapping achieves noise suppression and feature enhancement through an activation function with anatomical feature perception, expressed as:

[0019]

[0020] In the formula, is the calibrated feature output matrix, with the same dimension as and represents the output of the decoder of the autoencoder; is the element-wise multiplication operation; is the attention weight; is the dynamic response coefficient; is the total variation norm of the weighted features, calculated as ; Dynamically amplify the response in high-gradient regions; is the scaling factor, which controls the steepness of the non-linear activation; is the result of the linear transformation, representing the output of the encoder of the autoencoder.

[0021] Furthermore, an attention mechanism is embedded in the dynamic local calibration mapping, the feature weights are dynamically allocated through the gradient magnitude, and based on the gradient attention generation, the gradient magnitude map of the feature matrix is calculated.

[0022] Furthermore, the elastic gradient update uses a gradient-sensitive regularization term for elastic gradient update.

[0023] Furthermore, the feature fusion with manifold constraints utilizes the geometric continuity of tooth arrangement to construct an anatomical topological constraint term, forcing the feature space distance between adjacent tooth positions to be less than that between non-adjacent tooth positions, and avoiding non-physiological arrangements after reconstruction.

[0024] The present invention also provides an oral panoramic radiograph image data enhancement system, which includes:

[0025] A data acquisition and preprocessing module, which is used to acquire and preprocess oral panoramic images;

[0026] An oral panoramic radiograph image data enhancement model construction and training module, which includes a hash function transformation sub-module, a reinforced key feature capture sub-module, a dynamic local calibration mapping sub-module, an elastic gradient update sub-module, a feature fusion sub-module with manifold constraints, a mineralized tissue and soft tissue separation sub-module, and an adaptive regularization control sub-module;

[0027] An output module, which is used to input the preprocessed oral panoramic radiograph image into the trained autoencoder to complete the oral panoramic radiograph image data enhancement.

[0028] In addition, to achieve the above object, the present invention also provides a computer-readable storage medium, on which program instructions for the oral panoramic radiograph image data enhancement method are stored, and the program instructions for the oral panoramic radiograph image data enhancement can be executed by one or more processors to implement the steps of the oral panoramic radiograph image data enhancement method as described above.

[0029] (III) Beneficial effects

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] 1. The present invention combines the locality-sensitive hashing technique to aggregate the features of similar anatomical structures, thereby effectively suppressing the interference of non-concerned regions (such as soft tissues) and focusing on the alveolar bone region.

[0032] 2. The present invention adopts the non-local means filtering technique to retain the structural edges of the image and suppress noise. Through dynamic threshold adjustment, the alveolar bone region can be more accurately extracted, avoiding over-segmentation and misclassification.

[0033] 3. The present invention adopts different initialization strategies for different density regions, effectively strengthening the learning of key features. The dynamic local calibration mapping enhances the response of key anatomical structures through an activation function that perceives anatomical features.

[0034] 4. Combining the attention mechanism, the feature weights are dynamically adjusted based on the gradient information of the image to strengthen the response of the key region. A gradient-sensitive regularization term is adopted in the autoencoder loss function, which is specifically corrected for the metal artifact region, effectively eliminating the influence of artifacts and optimizing the reconstruction effect.

[0035] 5. Through the manifold constraint and anatomical topological relationship, it is ensured that the feature space distance between adjacent teeth is less than that between non-adjacent teeth, thus avoiding non-physiological tooth arrangement after model reconstruction.

[0036] 6. The dual-channel gating mechanism is adopted to extract the features of mineralized tissues and soft tissues respectively, further enhancing the structural and hierarchical sense of the image.

[0037] 7. Through adaptive regularization control, the regularization strength is dynamically adjusted according to the sparsity of the feature space, avoiding overfitting and optimizing the regularization strength during the training process. Description of the Drawings

[0038] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0039] Figure 1 is a schematic flowchart of a method for enhancing oral panoramic image data according to an embodiment of the present application;

[0040] Figure 2 is a visualization view of intermediate features according to an embodiment of the present application;

[0041] Figure 3 are the oral panoramic films before and after enhancement according to an embodiment of the present application;

[0042] Figure 4 are the comparative experiment results of the feature extraction accuracy and computational efficiency according to the embodiments of the present application and the prior art;

[0043] Figure 5 are the comparative experiment results of the training loss and generalization ability according to the embodiments of the present application and the prior art. Detailed implementation manners

[0044] The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0045] The following uses specific specific examples to illustrate the implementation manners of the present disclosure. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without making creative efforts belong to the scope of protection of the present disclosure.

[0046] It should also be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present disclosure in a schematic manner. The drawings only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and ratio of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0047] See Figure 1 , an oral panoramic radiograph image data enhancement method includes the following steps:

[0048] S1. Image data acquisition and preprocessing

[0049] Collect oral panoramic radiograph images, collected from a medical imaging system or device. Specifically, obtain oral panoramic images using a high-resolution digital imaging device (such as an oral panoramic X-ray machine or a CBCT device).

[0050] Furthermore, for the originally collected oral panoramic images, due to factors such as equipment noise, uneven illumination, and patient movement during the collection process, the image quality may be degraded. Therefore, the present invention first performs denoising processing on the images and uses a median filtering algorithm to remove the noise.

[0051] Further, for the complex background of oral panoramic images, the present invention automatically identifies the oral region through image segmentation technology and performs cropping. The cropped image only retains the oral-related region, removes unnecessary background parts, and reduces the complexity of subsequent processing.

[0052] Further, since the images taken by different hospitals and devices may have differences in aspects such as size, resolution, and perspective, it is necessary to standardize the collected images, unify the resolution of each oral panoramic image to 512×512 pixels, and ensure the consistency of data in subsequent algorithms.

[0053] S2. Construction and training of the oral panoramic film image data enhancement model

[0054] S201. Perform hash function transformation on images of different scales and conduct cross-scale feature fusion; for the high-dimensional characteristics of oral panoramic films, construct a multi-scale feature matrix, separate structural features and noise features. After the input data is geometrically standardized, an initial feature space is constructed through local sensitive hash mapping. According to the hash collision principle, similar anatomical structure features are aggregated, and the interference of non-concerned regions (such as soft tissues) is suppressed. The input image is decomposed into ( represents the total scale, such as , with scales of original, 1 / 2, and 1 / 4 resolutions). Different hash function transformations are adopted for images of different scales, and cross-scale feature fusion is performed, expressed as:

[0055]

[0056]

[0057] In the formula, is the feature matrix of the th sample, representing the key anatomical features of the input image; is the local sensitive hash function, with the input being the oral panoramic film image , and the output being the binary hash coding matrix, aggregating similar image block features through hash collision; is the th scale of the oral panoramic film image; is the alveolar bone region mask, focusing on the key anatomical region, with the same dimension as , and the element value being 0 (non-concerned region) or 1 (alveolar bone region); is the discrete cosine transform function, enhancing the feature expression of the alveolar bone microstructure (such as trabeculae) using multi-scale discrete cosine transform; is the weight matrix of the th scale, which is a training parameter and is obtained through gradient descent training to determine the contribution degree of each scale; Indicates the element-wise product, which is used to confine the hash features within the anatomical key regions; is the panoramic oral radiograph image.

[0058] Furthermore, during the mask generation process, first, non-local means filtering is performed on the original panoramic oral radiograph. By comparing the gray-scale distribution differences in the local neighborhoods, noise is suppressed while preserving the edges of the anatomical structures, expressed as:

[0059]

[0060] In the formula, is the panoramic oral radiograph after non-local means filtering; is the normalization constant, and its calculation method is , ensuring that the sum of the weights is 1; , representing the neighborhood similarity between the th pixel and the th one, represents the neighborhood of pixel ; is the filtering parameter, and its value range is from to , which is used to adjust the similarity sensitivity, represents the maximum pixel value in the original panoramic oral radiograph; is the coordinate of the target pixel, that is, the position of the pixel to be denoised currently; is the preset search window; is the pixel coordinate within the search window , representing the positions of other pixels in the neighborhood around pixel ; is the gray-scale value of pixel .

[0061] Furthermore, based on the alveolar bone CT value distribution (usually 800 - 3000 HU), an adaptive thresholding method is adopted for threshold segmentation, expressed as:

[0062]

[0063] In the formula, is the dynamic threshold, which is used to distinguish the alveolar bone (high density) from the non-bone regions (low density), with the unit of HU; is the mean value within the local window (30×30 pixels), which reflects the regional average density; is the standard deviation within the local window (30×30 pixels), which quantifies the regional density fluctuation; is the empirical coefficient, which is used to adjust the sensitivity of the threshold to the density fluctuation. Preferably, is set to 1.5.

[0064] Further, an image segmentation operation is performed. Non-local means filtering is carried out for all pixel values and compared with a dynamic threshold to achieve pixel binarization, which is expressed as:

[0065]

[0066] is the binarization value of the pixel , 1 represents the alveolar bone candidate region, 0 represents the background; is the non-local means filtering pixel value of the pixel .

[0067] Further, morphological optimization is performed. First, small holes in the panoramic dental radiograph are eliminated through closing operation, which is expressed as:

[0068]

[0069] In the formula, is a 3×3 circular structuring element; represents the dilation operation, defined as , which is used to fill holes, is a circular structuring element centered on the pixel , is an empty set; represents the erosion operation, defined as , which is used to smooth the edges; is the alveolar bone region mask; is the binarized panoramic dental radiograph image.

[0070] Further, a region growing operation is performed on the alveolar bone region mask. Starting from the seed point (the alveolar crest vertex), the region is expanded according to the gradient constraint condition, which is expressed as:

[0071]

[0072] , is an adjacent pixel within the grown region; represents the interactive operation; is the gradient magnitude of the pixel , calculated by the Sobel operator, with the unit of HU / mm; is the gradient magnitude of the pixel ; is the gradient difference threshold, taking 15 HU / mm, which restricts the growth region within a reasonable anatomical range.

[0073] The interactive operation symbol Indicates the linkage between conditional judgment and region expansion. When a condition is triggered (by comparing the gradient difference between a candidate pixel and the neighboring pixels of the grown region to be less than the gradient difference threshold), it dynamically determines whether to expand the region, and is added to , otherwise it is skipped;

[0074] That is, the positive condition is that if the gradient difference with the adjacent pixel is less than the threshold , then is incorporated into the growing region ;

[0075] The reverse condition is that if is incorporated into , then it must satisfy .

[0076] S202. Enhance key feature capture; perform weight initialization of the autoencoder. Based on the density distribution characteristics of dental tissues, design an anisotropic weight initialization strategy. The high-density region (enamel) obtains a larger initial weight, and the weight of the low-density region (pulp cavity) decays to enhance key feature capture, expressed as:

[0077]

[0078]

[0079] In the formula, is the initial weight matrix of the autoencoder; is a normal distribution with a mean of 0 and a variance of ; is the normalized total variation norm, quantifying the local structure complexity; ensures the comparability of weights in different anatomical regions; is the local feature calibration intensity coefficient, with a value range of , controlling the decay rate of the weight; is the exponential function; is the gradient amplitude of the th feature matrix, calculated by the Sobel operator; is the median of all feature gradient amplitudes, used for normalization; is a very small constant to prevent the denominator from being zero; is the number of samples input to the autoencoder. Preferably, is set to 0.01, is set to 0.5, is set to .

[0080] S203. Perform dynamic local calibration mapping to achieve noise suppression and feature enhancement through an activation function based on anatomical feature perception, expressed as:

[0081]

[0082] In the formula, is the calibrated feature output matrix, with the same dimension as and represents the output of the decoder of the autoencoder; is the element-wise multiplication operation; is the attention weight; is the dynamic response coefficient, which is adaptively adjusted according to , and the calculation method is . According to the dynamic adjustment mechanism, when , grows adaptively according to to strengthen the anatomical edge features; is the total variation norm of the weighted features, and the calculation method is ; dynamically amplifies the response in the high-gradient region (root edge); is the scaling factor, which controls the steepness of the non-linear activation, and the value range is ; is the result of the linear transformation, representing the output of the encoder of the autoencoder, , where are the weights of the autoencoder, is the bias of the autoencoder, is the feature matrix of the sample. Preferably, is set to 1.5.

[0083] Furthermore, an attention mechanism is embedded in the dynamic local calibration mapping to dynamically allocate feature weights based on the gradient magnitude, and a gradient magnitude map of the feature matrix is generated based on gradient attention, expressed as:

[0084]

[0085] In the formula, is the result of the linear transformation; is the gradient magnitude map of the feature matrix; is the gradient of the result of the linear transformation; is the L2 norm.

[0086] Furthermore, the attention weight is calculated according to the gradient magnitude map of the feature matrix to enhance the response intensity in key regions such as the root edge, expressed as:

[0087]

[0088] In the formula, is the parameter matrix of the attention weight, which is a training parameter and is obtained by training with the gradient descent method; is the Sigmoid activation function; is the 1×1 convolution function; is the fused feature after element-wise addition of the linear transformation result and the gradient magnitude map of the feature matrix.

[0089] S204. Elastic gradient update: For the common metal artifact problem in panoramic dental radiographs, the loss function of the autoencoder uses a gradient-sensitive regularization term for elastic gradient update. The calculation method of the loss function of the autoencoder is expressed as:

[0090]

[0091] In the formula, is the loss function of the autoencoder, which includes the reconstruction loss part and the elastic regularization term part ; the elastic regularization term part penalizes the abnormal gradients in the metal artifact regions (high but low ), and the total variation norm in the denominator realizes region-adaptive regularization intensity; is the output feature of the autoencoder model for the th sample, that is, the sample feature reconstructed by the th autoencoder; is the target feature of the th sample; is the regularization coefficient of the autoencoder at the th iteration; is the L2 norm, and its calculation method is the same as that of the Euclidean distance, which represents the reconstruction error; is the gradient vector of the weights of the th layer of the autoencoder; is the L1 norm, which is used to penalize abnormal gradients; is the feature output of the sample at the th layer of the autoencoder; is the total variation norm of the feature output of the sample at the th layer of the autoencoder. Preferably, is set to 0.02.

[0092] Furthermore, backpropagation correction is performed. When is detected, gradient truncation is initiated for the weight channel, that is, random inactivation is performed on the neurons of the th layer of the autoencoder.

[0093] S205. Perform feature fusion with manifold constraints. Utilize the geometric continuity of tooth arrangement to construct an anatomical topological constraint term, which forces the feature space distance between adjacent tooth positions to be less than that between non-adjacent tooth positions, thus avoiding non-physiological arrangements after reconstruction, expressed as:

[0094]

[0095] In the formula, is the fused feature matrix; is the local feature weight coefficient, and its value range is ; is the target feature of the sample; represents the anatomical adjacency relationship, is the th and the th tooth position spatial distance, calculated by the distance between the central pixel points of the teeth; is the output feature of the autoencoder model for the th sample, is the output feature of the autoencoder model for the th sample; is the tooth position distance hyperparameter; represents the set of neighborhood feature points of the kth tooth, including the indexes of the central pixel points of the adjacent 3 teeth; is the regularization coefficient of the autoencoder. Preferably, is set to 0.5, is set to 0.1.

[0096] S206. Separate the mineralized tissue and soft tissue features through a dual-channel gating mechanism, expressed as:

[0097]

[0098] In the formula, is the element-wise multiplication operation; is the separated feature; is the ReLU activation function, characterizes the mineralized tissue channel; is the Sigmoid activation function; is the mineralized tissue feature extraction weight matrix, which is a training parameter and is updated by gradient descent; is the gating weight matrix, which controls the activation threshold of the mineralized tissue feature and is a training parameter and is updated by gradient descent; is the soft tissue feature extraction weight matrix, which is a training parameter and is updated by gradient descent; is the Softplus function, characterizes the soft tissue channel and is used to retain the low-frequency information of the soft tissue.

[0099] S207. Perform adaptive regularization control, dynamically adjust the regularization strength according to the sparsity of the feature space, which is expressed as:

[0100]

[0101] In the formula, is the regularization coefficient of the autoencoder for the th iteration; KL is the KL divergence function, which measures the degree of data distribution matching; is the distribution of the original data; is the distribution of the reconstructed data of the autoencoder model; is a parameter that controls the overfitting suppression strength. Preferably, is set to 0.2.

[0102] S208. Repeat the above steps iteratively until the preset iteration stop condition is satisfied, which means the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0103] S3. Oral panoramic radiograph image data augmentation

[0104] Input the preprocessed oral panoramic radiograph image into the trained autoencoder. The structure of the autoencoder usually includes:

[0105] Encoder part: Compress the high-dimensional features of the input image into a low-dimensional latent space and extract the main features of the image;

[0106] Decoder part: Remap the low-dimensional features in the latent space back to the space of the original image to reconstruct the image.

[0107] At this time, the decoder of the autoencoder will generate the enhanced image according to the weights learned during the training process.

[0108] As Figure 2 shows the visualization views of each intermediate feature, where:

[0109] (1) Original image

[0110] represents the original oral panoramic image without any processing. The original image usually contains a lot of noise and its contrast may be low. In this image, the contour of the teeth is not clear enough, possibly due to soft tissue interference, uneven lighting, or other noise during the imaging process.

[0111] (2) Image with enhanced contrast

[0112] It represents the image after enhancing the contrast through histogram equalization. After the contrast enhancement, the details and structures of the image become more obvious, especially the contours of the teeth and the alveolar bone area. Histogram equalization makes the brightness distribution of the image more uniform, making the dark areas brighter and the bright areas more prominent, reducing the influence of noise and enhancing the structural features of the image.

[0113] (3)Scale 1

[0114] It represents the image processed at the original resolution of the enhanced image. Without scaling, the original image retains the finest details and is suitable for detecting high-resolution structural features, such as the details of teeth and the skeleton. However, it may be too sensitive to noise and details and may require further noise reduction or feature extraction.

[0115] (4)Scale 0.5

[0116] It represents the image with a 1 / 2 resolution of the enhanced image. By reducing the resolution, it can better emphasize the large structures, reduce the interference of fine noise, and is suitable for identifying anatomical structure features in a larger range, such as the general contour of the alveolar bone. The lower resolution is effective in noise reduction and extracting structural features.

[0117] (5)Scale 0.25

[0118] It represents the image with a 1 / 4 resolution of the enhanced image. At a lower resolution, it further simplifies the image details and emphasizes the large structures of the image, such as the overall shape of the alveolar bone, and is suitable for global feature analysis, but a large amount of microscopic details will be lost.

[0119] (6)Image after strong denoising

[0120] It represents the image after denoising using non-local means filtering. The denoised image looks smoother, the background noise is reduced, and the edges of the teeth and the skeleton become clearer. It can remove some unwanted noise while retaining the structural features of the image, and is particularly suitable for medical image processing and image processing in high-noise environments.

[0121] (7)Canny edge detection

[0122] It represents the edges of the image detected by the Canny algorithm. Canny edge detection emphasizes the edges and contours in the image, especially the contours of the teeth and the skeleton. By detecting the strong edges in the image, it highlights the teeth, alveolar bone and their boundaries.

[0123] (8)Adaptive threshold segmentation

[0124] It represents the result of segmentation using the adaptive threshold method. The adaptive threshold segmentation method dynamically determines the segmentation threshold according to local gray characteristics, which can effectively distinguish the alveolar bone (high-density area) from other low-density areas, showing the alveolar bone area extracted from the overall image by this method, and is suitable for subsequent region growing and further feature analysis.

[0125] (9)Morphological optimization

[0126] It represents the image after applying morphological operations. Morphological operations (such as closing and erosion) further remove small holes in the image and smooth the edges, which can better highlight the key areas of the alveolar bone, remove irrelevant noise and small artifacts. The closing operation effectively fills the gaps, while the erosion operation makes the edges smoother and enhances the coherence of the structure.

[0127] Furthermore, by comparing the panoramic oral radiographs before and after enhancement using the autoencoder, it can be found that through the improvement of the autoencoder and multi-scale feature fusion, the reconstructed panoramic oral radiograph is more superior than the one before reconstruction, achieving the effect of data enhancement.

[0128] As Figure 3 shown are the panoramic oral radiographs before and after enhancement.

[0129] Through analysis, it is known that the reason for achieving this image enhancement effect is the adoption of a multi-scale feature matrix to separate structural features and noise features. By performing geometric normalization, hash mapping, and cross-scale feature fusion on the image, anatomical structures in the image (such as alveolar bone) are better expressed, while non-interest regions (such as soft tissues) are suppressed. Moreover, non-local mean filtering is used to remove noise in the image, suppressing the surrounding noise while maintaining the edges of anatomical structures. Moreover, adaptive threshold segmentation based on the density distribution of the alveolar bone region is adopted, enabling the region of the alveolar bone to be more accurately identified and extracted. Moreover, through dynamic local calibration mapping and an attention mechanism, the response intensity of key anatomical regions (such as the alveolar bone edge, root edge, etc.) is emphasized, and feature weights are dynamically allocated based on the gradient magnitude map and the gradient magnitude map of the feature matrix, thereby enhancing the attention to key regions. Moreover, the loss function of the autoencoder includes elastic gradient update and regularization terms, which are specifically used to penalize metal artifact regions. By penalizing the abnormal gradients of metal artifacts, the artifact problem caused by metal objects in the image is effectively avoided. Moreover, through manifold-constrained feature fusion, using the geometric continuity of tooth arrangement, an anatomical topology constraint term is constructed, forcing the feature space distance between adjacent tooth positions to be less than that between non-adjacent tooth positions, avoiding non-physiological arrangements after reconstruction. Moreover, through a dual-channel gating mechanism, the features of mineralized tissues and soft tissues are separated, and the ReLU and Sigmoid functions are used to extract the features of mineralized tissues and soft tissues respectively, thereby enhancing the expression of mineralized tissue features and suppressing the low-frequency information of soft tissues.

[0130] To further quantify the experimental effect, experimental analysis was carried out as Figures 4 - 5 :

[0131] To verify the effectiveness of the multi-scale feature fusion method in extracting key anatomical structure features in a noise interference environment, a comparison was made with conventional single-scale processing techniques. The experiment simulated medical image data with different noise levels and compared the performance of the two methods in terms of feature matching accuracy and computational efficiency. Figure 4 The experimental results show that as the noise level increases, the feature matching accuracy of the conventional single-scale method decreases significantly, while this technology effectively separates noise and structural features through the multi-scale feature fusion mechanism and still maintains a high recognition accuracy under the same noise conditions. Although the processing time of this technology increases slightly due to multi-scale calculation, its accuracy advantage has higher clinical value in the medical image analysis scenario, indicating that the multi-scale design can enhance the robustness of the model to complex noise while retaining key anatomical details such as the micro-structure of the alveolar bone.

[0132] To verify the optimization effect of the dynamic gradient-sensitive regularization strategy on the model training process, it is compared with the fixed-intensity regularization method. By analyzing the change of the training loss decline trend and the generalization ability of the test set, the advantages of this technology in suppressing metal artifact interference and preventing overfitting are verified. As Figure 5 , experiments show that the fixed regularization method has fluctuations in the loss value in the later stage of training, and the growth of the test accuracy tends to stagnate. While this technology can flexibly adjust the regularization intensity, enabling the model to maintain stable generalization performance while converging quickly. The dynamic regularization mechanism can adaptively suppress abnormal gradients according to the sparsity of the feature space, effectively reducing the interference of the metal artifact area on weight update, and finally showing better anatomical structure reconstruction quality on the test set, proving its targeted adaptation ability to special noise patterns in complex medical images.

[0133] In this embodiment, aiming at the high-dimensional characteristics of panoramic dental radiographs, a multi-scale feature matrix construction method is designed, and the features of similar anatomical structures are aggregated by combining the local sensitive hashing technology, so as to effectively suppress the interference of non-concerned areas (such as soft tissues) and focus on the alveolar bone area.

[0134] The non-local mean filtering technology is used to retain the structural edges of the image and suppress noise, so as to obtain clearer boundaries in the alveolar bone area. Based on the adaptive threshold segmentation method of the alveolar bone CT value, through dynamic threshold adjustment, the alveolar bone area can be more accurately extracted, avoiding over-segmentation and misclassification.

[0135] When initializing the weights of the autoencoder, the density distribution characteristics of dental tissues are considered, and different initialization strategies are adopted for different density regions (such as enamel, dental pulp cavity), effectively strengthening the learning of key features. The dynamic local calibration mapping enhances the response of key anatomical structures (such as the root edge) through an activation function that perceives anatomical features.

[0136] Combined with the attention mechanism, the feature weights are dynamically adjusted based on the gradient information of the image to strengthen the response of key regions (such as the root edge). A gradient-sensitive regularization term is adopted in the autoencoder loss function, which is specifically corrected for the metal artifact area, effectively eliminating the influence of artifacts and optimizing the reconstruction effect.

[0137] Through manifold constraints and anatomical topological relationships, it is ensured that the feature space distance between adjacent teeth is less than that between non-adjacent teeth, thus avoiding non-physiological tooth arrangements after model reconstruction.

[0138] A dual-channel gating mechanism is adopted to extract mineralized tissue and soft tissue features respectively, and the ReLU and Sigmoid activation functions are used to activate the mineralized tissue channel and the soft tissue channel respectively, further enhancing the structural and hierarchical sense of the image.

[0139] Through adaptive regularization control, the regularization strength is dynamically adjusted according to the sparsity of the feature space to avoid overfitting, and at the same time, the regularization strength during training is optimized.

[0140] An embodiment of the present invention also proposes an oral panoramic radiograph image data enhancement system, including:

[0141] A data acquisition and preprocessing module, which is used to acquire and preprocess oral panoramic images;

[0142] An oral panoramic radiograph image data enhancement model construction and training module, which includes a hash function transformation sub-module, a key feature capture enhancement sub-module, a dynamic local calibration mapping sub-module, an elastic gradient update sub-module, a feature fusion sub-module with manifold constraints, a mineralized tissue and soft tissue separation sub-module, and an adaptive regularization control sub-module.

[0143] The hash function transformation sub-module is used to perform differential hash function transformation on images of different scales and perform cross-scale feature fusion; the key feature capture enhancement sub-module is used to initialize the weights of the autoencoder and design an anisotropic weight initialization strategy based on the density distribution characteristics of dental tissues; the dynamic local calibration mapping sub-module is used to implement noise suppression and feature enhancement through an activation function that perceives anatomical features; the elastic gradient update sub-module is used to perform elastic gradient update on the loss function of the autoencoder using a gradient-sensitive regularization term; the feature fusion sub-module with manifold constraints is used to construct an anatomical topology constraint term; the mineralized tissue and soft tissue separation sub-module is used to separate mineralized tissue and soft tissue features through a dual-channel gating mechanism; the adaptive regularization control sub-module is used to dynamically adjust the regularization strength according to the sparsity of the feature space.

[0144] An output module, which is used to input the preprocessed oral panoramic radiograph image into the trained autoencoder to complete the enhancement of the oral panoramic radiograph image data.

[0145] In addition, an embodiment of the present invention also proposes a computer-readable storage medium, on which program instructions for the oral panoramic radiograph image data enhancement method are stored. The oral panoramic radiograph image data enhancement program instructions can be executed by one or more processors to implement the steps of the oral panoramic radiograph image data enhancement method as described above.

[0146] The above embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An oral panoramic radiograph image data enhancement method, characterized in that, The steps include: S1. Image data acquisition and preprocessing, including obtaining oral panoramic images using high-resolution digital imaging equipment and preprocessing; S2. Construction and training of oral panoramic radiograph data enhancement model, including hash function transformation of images of different scales and cross-scale feature fusion, strengthening key feature capture, dynamic local calibration mapping, noise suppression and feature enhancement through anatomical feature-aware activation functions, elastic gradient update, manifold-constrained feature fusion, separation of mineralized tissue and soft tissue features through dual-channel gating mechanism, dynamic adjustment of regularization strength and iteration according to feature space sparsity; S3, oral panoramic film image data enhancement, inputting the preprocessed oral panoramic film image into the trained autoencoder to generate an enhanced image; The enhanced key feature capture includes designing an anisotropic weight initialization strategy based on the density distribution characteristics of tooth tissue, so that high-density areas obtain larger initial weights and low-density areas have weight attenuation; The dynamic local calibration map achieves noise suppression and feature enhancement through an anatomical feature-aware activation function, expressed as: ， In the formula, is the calibrated feature output matrix, with the same dimension as , representing the output of the decoder of the autoencoder; is the element-wise multiplication operation; is the attention weight; is the dynamic response coefficient; is the total variation norm of the weighted feature, calculated as ; dynamically amplifies the response in the high-gradient region; is the scaling factor, controlling the steepness of the non-linear activation; is the result of the linear transformation, representing the output of the encoder of the autoencoder.

2. The method for enhancing oral panoramic radiograph image data according to claim 1, wherein The performing hash function transformation on images of different scales and performing cross-scale feature fusion comprises: A multi-scale feature matrix is constructed to separate structural features from noise features. After the input data is geometrically normalized, the initial feature space is constructed through local sensitive hash mapping. Similar anatomical structure features are aggregated according to the hash collision principle to suppress interference from non-interest areas. The input image is decomposed into different scales, differentiated hash function transformations are used for images of different scales, and cross-scale feature fusion is performed.

3. The method for enhancing oral panoramic radiograph image data according to claim 2, wherein It also includes a regional growth operation on the alveolar bone region mask, starting from the seed point, i.e., the alveolar ridge vertex, and expanding the region according to the gradient constraint, which is expressed as: ， , are adjacent pixels within the grown area; represents an interaction operation; is the pixel gradient magnitude, calculated by the Sobel operator, with the unit of HU / mm; is the pixel gradient magnitude; is the gradient difference threshold, taking 15 HU / mm, which restricts the growth area within a reasonable anatomical range.

4. The method for enhancing oral panoramic radiograph image data according to claim 1, wherein An attention mechanism is embedded in the dynamic local calibration map to dynamically assign feature weights by gradient magnitude, and the gradient magnitude map of the feature matrix is calculated based on gradient attention generation.

5. The method for enhancing oral panoramic radiograph image data according to claim 1, wherein The elastic gradient update uses a gradient-sensitive regularization term to perform elastic gradient update.

6. The method for enhancing oral panoramic radiograph image data according to claim 1, wherein The manifold-constrained feature fusion utilizes the geometric continuity of tooth arrangement to construct anatomical topological constraints, forcing the feature space distance of adjacent teeth to be smaller than that of non-adjacent teeth, thereby avoiding non-physiological arrangement after reconstruction.

7. A system using the oral panoramic radiograph image data enhancement method according to any one of claims 1 to 6, comprising: A data acquisition and preprocessing module, which is used to acquire and preprocess oral panoramic images; Oral panoramic film image data enhancement model construction and training module, which includes hash function transformation submodule, enhanced key feature capture submodule, dynamic local calibration mapping submodule, elastic gradient update submodule, manifold constrained feature fusion submodule, mineralized tissue and soft tissue separation submodule, and adaptive regularization control submodule; The output module is used to input the preprocessed oral panoramic film image into the trained autoencoder to complete the oral panoramic film image data enhancement.

8. A computer-readable storage medium, characterized in that, Program instructions for an oral panoramic radiograph image data enhancement method are stored on the computer-readable storage medium, and the program instructions for the oral panoramic radiograph image data enhancement can be executed by one or more processors to implement the steps of the oral panoramic radiograph image data enhancement method as described in any one of claims 1-6.

Images