Oral disease diagnosis method based on multi-path feature extraction and frequency enhancement

Abstract

The invention provides an oral disease diagnosis method based on multi-path feature extraction and frequency enhancement, and relates to the technical field of oral medical treatment, and the method employs a space channel to cooperate with attention enhancement feature extraction, improves the representation of a network on image features, employs a full-dimensional dynamic convolution kernel, employs a three-branch parallel structure to combine with a frequency enhancement technology, and achieves the diagnosis of oral diseases. The collaborative modeling of global and local features of the network is improved, the perceptual ability of edge information is enhanced, difficult sample learning is enhanced through an adaptive threshold focus loss function, the problem of sample imbalance is improved, rare disease focuses with limited sample sizes and low-resolution images are more sensitive, and the detection result is more accurate.

Description

Technical Field

[0001] This invention relates to the field of oral medical technology, specifically to a method for diagnosing oral diseases based on multi-path feature extraction and frequency enhancement. Background Technology

[0002] Oral diseases have become a prevalent health problem worldwide. According to the World Health Organization, approximately 3.5 billion people globally suffer from oral diseases, with common conditions such as dental caries, periodontitis, and periapical periodontitis affecting over 90% of cases. These diseases not only severely impact patients' chewing function, speech ability, and aesthetic appearance, reducing their quality of life, but also have a clear link to major health problems such as systemic infections and cardiovascular diseases, posing a significant threat to human health. Panoramic dental X-ray imaging, due to its wide imaging range, low radiation dose, and low cost, has become the preferred tool for oral disease screening, diagnostic staging, and treatment planning. Accurate detection and classification of oral lesions are crucial for clinical decision-making, directly determining the effectiveness of treatment plans and the reliability of prognostic assessments.

[0003] Traditional oral disease diagnosis relies on doctors' visual observation, manual exploration, and two-dimensional imaging examinations. The diagnosis of caries requires the use of a mouth mirror and probes, which can cause severe discomfort to patients. Furthermore, the diagnostic process depends on the doctor's subjective experience, and there are core pain points such as low diagnostic efficiency and limited diagnostic accuracy, especially when identifying subtle and variable abnormal targets.

[0004] Existing deep learning-based lesion detection methods are mostly designed for single disease types, exhibiting poor generalization performance when extended to multi-label classification tasks. Firstly, due to the small grayscale differences and blurred boundaries between different tissues in two-dimensional oral images, traditional detection networks struggle to effectively capture the texture and morphological features of lesions, resulting in structural information loss in the feature maps output by shallow networks. Secondly, the scale differences between lesions and the high degree of similarity between different categories make it difficult for single-scale receptive field network structures to accommodate the detection needs of targets of different sizes, easily leading to inter-class feature confusion. Finally, due to the uneven distribution of category data, models tend to overfit high-frequency disease features during training, while lacking sufficient representation ability for rare diseases with limited sample sizes, ultimately leading to systematic biases and misjudgments in rare disease identification.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing oral diseases based on multi-path feature extraction and frequency enhancement, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A diagnostic method for oral diseases based on multi-path feature extraction and frequency enhancement, comprising the following steps: S1: Collect historical panoramic dental X-ray image datasets, and automatically filter and remove low-quality images with motion artifacts, overexposure, and severe metal artifacts based on the preprocessing unit, so as to keep the filtered images; S2: Establish 13 categories of oral abnormality detection standards covering periapical lesions, caries, wisdom teeth, missing teeth, metal fillings, root canal fillings, implants, metal crowns, metal bridges, metal brackets, ceramic brackets, malocclusion, and supernumerary teeth, in order to determine the multi-category lesion labels of each screened image; S3: Image enhancement is performed on the selected images based on generative adversarial networks to obtain enhanced images with clear boundary morphology and prominent lesion features; S4: Construct a multi-label detection network including a spatial channel collaborative attention module, a full-dimensional dynamic convolution module, and an adaptive threshold focus loss to ensure the multi-label detection network's ability to perceive low-resolution, small lesion regions and small sample categories; S5: Input the enhanced image and the corresponding multi-category lesion labels into the multi-label detection network for training. Optimize the model parameters through an iterative process. Then, input the enhanced image to be analyzed into the trained multi-label detection network to determine its disease diagnosis result.

[0008] Furthermore, a historical panoramic dental X-ray image dataset is obtained, which contains image data of various oral diseases, and is initially screened through a preprocessing unit.

[0009] Further, S31: The filtered image is input into the generator, which includes an encoder and a decoder. First, the encoder processes the filtered image by using a convolutional residual module and downsampling to extract feature maps. The convolutional residual module consists of two 3×3 convolutional layers, a batch normalization layer, and a Leaky ReLU activation function. S32: The feature map output by the encoder is sent to the decoder, and multi-scale features are fused through the upsampling module and convolutional residuals, and gradually restored to the original image size. The upsampling module is the operation module of the convolutional neural network. S33: For the feature maps between the encoder and decoder, information is passed through skip connections, and an attention gating mechanism is built on the skip connection path to filter the features from the encoding layer to the decoding layer; The formula upon which attention gating mechanisms are based is: ; in, This represents the filtered feature map. Let d represent the features of the k-th coding layer, and let d represent the features of the last decoding layer. This represents three 1×1 convolutional layers, where R represents the ReLU function. This represents the Sigmoid function.

[0010] S34: Each layer of the decoder adds an extra path, consisting of a 1×1 convolutional layer and an upsampling layer; S35: Interpolate the feature maps to the same size, concatenate them along the channel dimension, feed the concatenated features into a 3×3 convolutional layer to obtain fused multi-scale information, and add them element-wise to the decoder output to obtain the superimposed features.

[0011] S36: The superimposed feature input spatial channel attention module performs attention allocation on the decoder output features in both channel and spatial dimensions to generate the enhanced image; The enhanced image and the real image are input into the discriminator, which distinguishes between the enhanced image and the real image, extracts and fuses multi-scale features, and outputs the discrimination probability. During the training process, the generator and discriminator are trained through the loss function until the loss function converges.

[0012] S37: The enhanced image obtained in S36 is input into the discriminator along with the real image, and then... , , Max pooling and upsampling are used to obtain feature maps of three different scales. These maps are then concatenated and fused using convolution, batch normalization, and ReLU activation functions to capture global contextual features. Based on these global contextual features, [the following is a separate, unrelated sentence:] ...through... Convolution and Sigmoid activation function generation The true and false probability matrix is ​​used, and the mean of the matrix is ​​used as the final output of the discriminator. S38: The discriminator training process employs adversarial loss, cycle consistency loss, texture fidelity loss, image smoothing loss, and perceptual loss for collaborative supervision. The generator and discriminator are alternately optimized until the loss function converges. The input image is generated by an encoder, and after processing by encoder feature extraction, attention gating mechanism, multi-scale information fusion, and spatial channel attention module, the enhanced image is output.

[0013] Furthermore, the spatial channel attention module in S36 includes: S361: The decoder's output features are subjected to global max pooling and global average pooling in each channel to extract global information in the channel dimension and generate global max pooling feature vectors and average pooling feature vectors containing the number of channels. S362: Input the global max pooling feature vector and the average pooling feature vector into the multilayer perceptron, add the global max pooling feature vector and the average pooling feature vector, and generate channel attention weights through the sigmoid function.

[0014] S363: Multiply the channel attention weights by the superimposed features to output the weighted channel features.

[0015] S364: The weighted channel features are processed based on max pooling and average pooling, and spatial attention weights are generated after convolutional layers and sigmoid activation. S365: Adaptive learning is achieved by multiplying spatial attention weights with the superimposed features; The formula for calculating spatial attention weights is: ; ; in, This represents the weighted channel characteristics. Represents spatial attention weights. Indicates the features after superposition. and These represent global average pooling and max pooling based on spatial dimensions, respectively. and These represent global average pooling and max pooling based on spatial dimensions, respectively. MLP stands for Multilayer Perceptron. This is the Sigmoid activation function.

[0016] Furthermore, the method for training the network in the detection network is as follows: S41: The enhanced image is input into the detection network, and feature extraction is performed through the backbone network and embedded spatial channel collaborative attention. The outputs of the last three stages of the detection network are used as the input to the hybrid encoder, and the outputs of the three stages include the fifth deep feature. Fourth deep features and third deep features ; S42: The final three stages of the extracted output-input hybrid encoder, deep features. After being converted into a one-dimensional vector, it is input into the intra-scale interaction module for feature enhancement. The output is adjusted to a two-dimensional sequence for cross-scale feature fusion to generate fused features. The intra-scale interaction module is used to realize contextual interaction of features within the same scale, strengthen the correlation and information transmission between features, and improve the discriminability and robustness of feature expression. S43: Enhanced features of the intra-scale interaction module The third deep features extracted from the backbone network Fourth deep feature Cross-scale fusion in pyramid networks, third deep features and fourth deep features After fusion, the data are fed into a full-dimensional dynamic convolution module for collaborative modeling using global and local features. S44: The full-dimensional dynamic convolution module is then integrated into the pyramid network for cross-scale fusion after processing. S45: Minimum uncertainty query selection of the fused feature map input, which selects a fixed number of features from the feature sequence output by the encoder as the target query; S46: After being input into the decoder, the prediction head maps the confidence score and bounding box. The training process is supervised by adaptive threshold focus loss. Learning of difficult samples is enhanced by dynamic threshold decoupling and adaptive weight adjustment. S47: The network continuously iterates and updates its parameters to ensure that the loss parameters converge to a minimum value, thus completing the entire training process and obtaining the optimal weights.

[0017] Furthermore, the method for feature enhancement through the spatial channel collaborative attention module is as follows: S411: After the enhanced image input detection network, the input channel attention of the features extracted from the convolutional layers is decoupled along the height and width dimensions. Global average pooling is applied to each dimension to generate two unidirectional one-dimensional sequences. and sequence ; S412: Sequence and sequence Divide into K independent sub-features of the same size Features of the sub-score 'i' is used to index the sub-features. The sub-feature decomposition process is as follows: ; ; in, This represents the i-th sub-feature in the height dimension. Let K represent the i-th sub-feature in the width dimension, K represent the number of sub-feature divisions, and C represent the number of channels; S413: Based on sub-features Features of the sub-score Feature extraction is performed using lightweight convolution, and the features are concatenated. Normalization is then performed using K sets of normalization methods, and spatial attention weights are generated using the sigmoid activation function. and Intermediate features are obtained by multiplying spatial attention with features extracted from convolutional layers. ; The spatial attention weights are calculated as follows: ; ; in, The spatial attention weights in the height dimension are calculated as follows: ; in, express Sigmoid Normalization These represent K sets of normalization operations performed along the height and width dimensions, respectively. This represents the spatial structure information of the i-th sub-feature obtained after lightweight convolution operation, where X is the feature extracted by the convolutional layer. S414: The intermediate features are downsampled using average pooling to obtain the feature map. ; Feature map after downsampling Represented as: ; in, This indicates that pooling operations are performed using convolutional kernels of size K×K, with resolution ranging from... Readjusted to ; S415: Feature map after downsampling Normalization and 1×1 depthwise convolution generate query Q, key K, and value V; ; ; ; in, , , These represent linear projection operations that generate queries, keys, and values, respectively. S416: Perform relevance analysis based on query Q, key K, and value V to generate channel attention. Channel attention Multiply by the intermediate features to obtain the weighted features. , as the channel-enhanced feature map output; Channel attention The calculation method is as follows: ; The channel attention-weighted features are represented as follows: ; in, This indicates that a pooling operation is performed using a convolutional kernel of size 1×1. This indicates the aforementioned intermediate characteristics.

[0018] Further, S44 includes: S441: Features after fusion After convolution processing, the data is fed into three branches, which include a local branch, a large kernel branch, and a global branch. S442: The big core branch uses a size of [value missing]. Deep convolutions acquire global structural information and large-scale contextual information; S443: Local branch usage Deep convolutional layers are used to modulate local signals; S444: The global branch captures global information through dual-domain channel attention and frequency channel attention modules. First, the input features are subjected to global average pooling and spatial dimension compression to obtain channel information. Channel weights are obtained through 1×1 convolution. Simultaneously, the frequency domain features are transformed to the frequency domain by Fourier transform. After element-wise multiplication of the frequency domain features with the convolution output, the features are transformed back to the spatial domain by inverse Fourier transform to obtain the features. Based on features The output features of the dual-domain channel attention are generated by multiplying the channel weights element-wise. ; Output features of dual-channel attention The calculation formula is as follows: ; in, This represents element-wise multiplication. This represents the output characteristics after the inverse Fourier transform. This represents a 1×1 convolutional layer, and GAP represents global average pooling. S445: Secondly, the output features of dual-domain channel attention. The data is fed into the frequency channel attention module, and after depthwise separable convolutional dimensionality reduction in the frequency channel attention module, enhanced intermediate features are obtained. The enhanced intermediate features are obtained by using Fast Fourier Transform. Map to the frequency domain and construct a Query, Key, Value triplet; S446: The query and key elements are multiplied together and then fed into the ingress control unit Sigmoid, which is then multiplied element-wise with the value element to obtain the attention weight. After performing attention calculations on the spectrum, the attention weights are inversely Fourier transformed and returned to the spatial domain. S447: Enhanced intermediate feature parallel input frequency channel enhancement module (FWM). The frequency channel enhancement module includes convolution, GELU activation function and Sigmoid activation function processing. It performs residual repair in the frequency domain through gating mechanism, concatenates the frequency domain attention result with the frequency gating path, and compresses and fuses it into the original dimension through the second convolutional layer to achieve unified modeling of multi-source semantics.

[0019] S448: Add the final output to the original input to obtain the output features. The formula used is: ; in, This indicates the inverse Fourier transform operation. Indicates a convolutional layer. This indicates the first convolutional layer. Indicates Fourier transform, This indicates the second convolutional layer. , This represents element-wise addition.

[0020] S449: The outputs of the local branch, the big kernel branch, and the global branch are weighted and fused, and then output after convolutional modulation.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes generative adversarial networks (GANs) to deeply capture the semantic features of images through adversarial training between the generator and discriminator. This generates high-quality images that more closely resemble the distribution of real-world images, avoiding the common problems of traditional histogram equalization and filtering algorithms failing to distinguish foreground from background when processing image regions, resulting in over-enhancement of high-quality regions and under-enhancement of low-quality regions. Furthermore, this invention learns complex mapping relationships from low-quality to high-quality images through GANs, making it suitable for enhancement tasks of unpaired images. Effective training can be achieved without relying on strictly paired high- and low-quality images. Based on this, a multi-scale attention mechanism and loss function are introduced. This invention constrains the quality of generated images from multiple dimensions, reducing noise and artifacts while highlighting important regions in the image. It employs an improved multi-label convolutional neural network for oral disease diagnosis, using spatial channel collaborative attention to enhance feature extraction and improve the network's representation of image features. It utilizes full-dimensional dynamic convolutional kernels and a three-branch parallel structure combined with frequency enhancement technology to improve the collaborative modeling of global and local features, enhancing the network's ability to perceive edge information. It strengthens the learning of difficult samples through an adaptive threshold focus loss function, improving the sample imbalance problem and making it more sensitive to rare disease lesions and low-resolution images with limited sample size, resulting in more accurate detection results. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the spatial channel attention module of the present invention. Figure 3 This is a schematic diagram of the multi-category detection network of the present invention; Figure 4 This is a schematic diagram of the spatial channel collaborative attention module of the present invention; Figure 5 This is a schematic diagram of the full-dimensional dynamic convolution module of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0024] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example: Please see Figure 1 The present invention provides a technical solution: A diagnostic method for oral diseases based on multi-path feature extraction and frequency enhancement, comprising the following steps: S1: Collect historical panoramic dental X-ray image datasets, and automatically filter and remove low-quality images with motion artifacts, overexposure, and severe metal artifacts based on the preprocessing unit, so as to retain the filtered images; the preprocessing unit is a quality assessment system based on texture features, edge density, and brightness histogram statistics, which is existing technology.

[0026] A historical panoramic dental X-ray image dataset is acquired. This dataset contains image data of various oral diseases. Due to the data being collected from different hospitals and different equipment, there are differences in the grayscale information of the same image, as well as the influence of differences in operators. In addition, there may be damaged or poorly imaged samples. Therefore, a preprocessing unit is needed to perform preliminary screening to automatically identify and remove low-quality images with motion artifacts, overexposure, and severe metal artifacts, so as to ensure the reliability and validity of the data.

[0027] S2: Establish 13 categories of oral abnormality detection standards covering periapical lesions, caries, wisdom teeth, missing teeth, metal fillings, root canal fillings, implants, metal crowns, metal bridges, metal brackets, ceramic brackets, malocclusion, and supernumerary teeth, in order to determine the multi-category lesion labels of each screened image; Based on the standards for detecting oral abnormalities, annotation rules were formulated to clarify the characteristic definition and selection range of each type of lesion, ensuring that the annotation personnel maintain consistent identification standards for each type of lesion.

[0028] Import the filtered images into the labelimg tool, use the rectangular selection function to accurately outline the lesion area, and after completing the annotation of all lesions in a single image, save the annotation information in a format supported by the tool along with the original image to form a dataset containing the image and corresponding multi-category annotation data.

[0029] To ensure annotation quality, multiple annotators were assigned to independently annotate the same batch of images. The annotation results were compared through cross-checking. For annotated areas with discrepancies, the annotators jointly reviewed the results and determined the final annotation results based on the evaluation criteria, thereby reducing annotation errors.

[0030] S3: Image enhancement is performed on the selected images based on generative adversarial networks to obtain enhanced images with clear boundary morphology and prominent lesion features; The structure of the augmented network is as follows: S31: The filtered image is input into the generator, which includes an encoder and a decoder. First, the encoder processes the filtered image by using a convolutional residual module and downsampling to extract feature maps. The convolutional residual module is followed by a pooling layer to reduce the feature map size by 1 / 2. After six downsampling operations, a deep semantic feature with a scale of 1 / 32 is obtained. The convolutional residual module consists of two 3×3 convolutional layers, a batch normalization layer, and a LeakyReLU activation function. Skip connections are used between the input and output values ​​to realize residual calculation, which can avoid network degradation, gradient vanishing, and gradient exploding problems caused by increasing depth. S32: The feature map output by the encoder is sent to the decoder. Multi-scale features are fused through the upsampling module and convolutional residual, and the original image size is gradually restored. The upsampling module is the operation module of the convolutional neural network. Multi-scale feature fusion is in the decoding process. Not only are the features obtained by upsampling in the current layer used, but also the features of the corresponding layer of the encoder are combined. Features of different scales are fused together through skip connections. S33: For the feature maps between the encoder and decoder, information is passed through skip connections, and an attention gating mechanism is built on the skip connection path to filter the features from the encoding layer to the decoding layer, retaining important information to participate in the subsequent decoding process. Compared with simple fusion operations, this can suppress irrelevant activation responses, highlight important image regions, and effectively avoid redundant low-level features. The convolution kernel size of the gating mechanism is 1×1 and the stride is set to 1.

[0031] The formula upon which attention gating mechanisms are based is: ; in, This represents the filtered feature map. Let d represent the features of the k-th coding layer, and let d represent the features of the last decoding layer. This represents three 1×1 convolutional layers, where R represents the ReLU function. This represents the Sigmoid function.

[0032] S34: To enhance information sharing of features at different scale levels, an extra path is added to each layer of the decoder, consisting of a 1×1 convolutional layer and an upsampling layer. The convolutional layer extracts features at different scales and controls the balance between deep abstract semantics and high resolution. The upsampling layer eliminates scale differences.

[0033] S35: Interpolate the feature maps to the same size, concatenate them along the channel dimension, feed the concatenated features into a 3×3 convolutional layer to obtain fused multi-scale information, and add them element-wise to the decoder output to obtain the superimposed features.

[0034] S36: The superimposed feature input spatial channel attention module performs attention allocation on the decoder output features in both channel and spatial dimensions to generate the enhanced image; Please see Figure 2 The spatial channel attention module in S36 includes: S361: The decoder's output features are subjected to global max pooling and global average pooling in each channel to extract global information in the channel dimension and generate global max pooling feature vectors and average pooling feature vectors containing the number of channels. S362: Input the global max pooling feature vector and the average pooling feature vector into the multilayer perceptron, add the global max pooling feature vector and the average pooling feature vector, and generate channel attention weights through the sigmoid function.

[0035] S363: Multiply the channel attention weights by the superimposed features to output the weighted channel features.

[0036] S364: The weighted channel features are processed based on max pooling and average pooling, and spatial attention weights are generated after convolutional layers and sigmoid activation. S365: Adaptive learning is achieved by multiplying spatial attention weights with the superimposed features, thereby weighting the features at each spatial location to highlight important regions; The formula for calculating spatial attention weights is: ; ; in, This represents the weighted channel characteristics. Represents spatial attention weights. Indicates the features after superposition. and These represent global average pooling and max pooling based on spatial dimensions, respectively. and These represent global average pooling and max pooling based on spatial dimensions, respectively. MLP represents a multilayer perceptron, which is composed of... Convolutional layers, ReLU and Convolutional layers consist of This is the Sigmoid activation function.

[0037] The enhanced image and the real image are input into the discriminator, which distinguishes between the enhanced image and the real image, extracts and fuses multi-scale features, and outputs the discrimination probability. During the training process, the generator and discriminator are trained through the loss function until the loss function converges.

[0038] S37: The enhanced image obtained in S36 is input into the discriminator along with the real image, and then... , , Max pooling and upsampling are used to obtain feature maps of three different scales. These maps are then concatenated and fused using convolution, batch normalization, and ReLU activation functions to capture global contextual features. Based on these global contextual features, [the following is a separate, unrelated sentence:] ...through... Convolution and Sigmoid activation function generation The true / false probability matrix is ​​used, and the mean of the matrix is ​​used as the final output of the discriminator. The global context feature is a feature representation of the image that is fused from local details to global information through multi-layer processing of the network. It contains the interrelationships between various regions in the image and the overall semantic information, rather than just the features of local pixels or local regions.

[0039] S38: The discriminator training process employs adversarial loss, cycle consistency loss, texture fidelity loss, image smoothing loss, and perceptual loss for collaborative supervision. The generator and discriminator are alternately optimized until the loss function converges. The loss function expression is: ; ; ; ; ; ; Where x represents the enhanced image and y represents the original image. and These represent image transformations performed by a generator network from a real image to an enhanced image, and vice versa. and These are used to distinguish whether the input image is an enhanced image or a real image, respectively. and These represent the expected values ​​of the input image, This represents the feature extraction of the j-th channel in the i-th layer of a generative adversarial network, where H and W represent the height and width of the image, respectively. Image pixel values, and Let represent the horizontal and vertical gradients, respectively, and let c be a smoothing constant used to prevent the denominator from being zero. It is the standard deviation, used for statistical analysis of texture similarity. This is used to represent the summation operation performed on the K sub-regions into which the feature map is divided. Indicates basic loss. , Indicating resistance to loss, Indicates the loss of cycle consistency. Indicates perceived loss. Indicates texture fidelity loss. Indicates image smoothing loss. , , , , All of these represent hyperparameters. Used to refer to the loss function; The input image is generated by an encoder, and after processing by encoder feature extraction, attention gating mechanism, multi-scale information fusion, and spatial channel attention module, the enhanced image is output.

[0040] S4: Construct a multi-label detection network including a spatial channel collaborative attention module, a full-dimensional dynamic convolution module, and an adaptive threshold focus loss to ensure the multi-label detection network's ability to perceive low-resolution, small lesion regions and small sample categories; Please see Figure 3 The process of training the detection branch of the oral disease diagnostic model is as follows: S41: The enhanced image is input into the detection network, and feature extraction is performed through the backbone network and embedded spatial channel collaborative attention. The outputs of the last three stages of the detection network are used as the input to the hybrid encoder, and the outputs of the three stages include the fifth deep feature. Fourth deep features and third deep features ; The method for feature enhancement using a spatial channel collaborative attention module is as follows: Please see Figure 4 Spatial channel collaborative attention module, such as Figure 4 As shown: S411: After the enhanced image input detection network, the input channel attention of the features extracted from the convolutional layers is decoupled along the height and width dimensions. Global average pooling is applied to each dimension to generate two unidirectional one-dimensional sequences. and sequence ; S412: In order to learn different spatial distributions and contextual relationships, the sequence and sequence Divide into K independent sub-features of the same size Features of the sub-score 'i' is used to index sub-features. Depthwise separable convolutions enrich semantic information, while shared convolutions are used for alignment. The sub-feature decomposition process is as follows: ; ; in, This represents the i-th sub-feature in the height dimension. Let K represent the i-th sub-feature in the width dimension, K represent the number of sub-feature divisions, and C represent the number of channels; S413: Based on sub-features Features of the sub-score Feature extraction is performed using lightweight convolution, and the features are concatenated. Normalization is then performed using K sets of normalization methods, and spatial attention weights are generated using the sigmoid activation function. and Intermediate features are obtained by multiplying spatial attention with features extracted from convolutional layers. ; The spatial attention weights are calculated as follows: ; ; in, The spatial attention weights in the height dimension are calculated as follows: ; in, express Sigmoid Normalization These represent K sets of normalization operations performed along the height and width dimensions, respectively. This represents the spatial structure information of the i-th sub-feature obtained after lightweight convolution operation, where X is the feature extracted by the convolutional layer. S414: The intermediate features are downsampled using average pooling to obtain the feature map. ; Feature map after downsampling Represented as: ; in, This indicates that pooling operations are performed using convolutional kernels of size K×K, with resolution ranging from... Readjusted to ; S415: Feature map after downsampling Normalization and 1×1 depthwise convolution generate query Q, key K, and value V; ; ; ; in, , , These represent linear projection operations that generate queries, keys, and values, respectively. S416: Perform relevance analysis based on query Q, key K, and value V to generate channel attention. Channel attention Multiply by the intermediate features to obtain the weighted features. , as the channel-enhanced feature map output; Channel attention The calculation method is as follows: ; The channel attention-weighted features are represented as follows: ; in, This indicates that a pooling operation is performed using a convolutional kernel of size 1×1. This indicates the aforementioned intermediate characteristics.

[0041] S42: The final three stages of the extracted output-input hybrid encoder, deep features. After being converted into a one-dimensional vector, it is input into the intra-scale interaction module for feature enhancement. The output is adjusted to a two-dimensional sequence for cross-scale feature fusion to generate fused features. The intra-scale interaction module is used to realize contextual interaction of features within the same scale, strengthen the correlation and information transmission between features, and improve the discriminability and robustness of feature expression. S43: Enhanced features of the intra-scale interaction module The third deep features extracted from the backbone network Fourth deep feature Cross-scale fusion is performed in the pyramid network to enhance the model's focus on small-scale lesion regions, and third-layer deep features are used. and fourth deep features After fusion, the data is fed into a full-dimensional dynamic convolution module, which solves the problem of information loss of small targets due to insufficient receptive field of multi-scale targets by using global-local feature collaborative modeling. S44: The full-dimensional dynamic convolution module is then integrated into the pyramid network for cross-scale fusion after processing. Please see Figure 5 Full-dimensional dynamic convolution module, such as Figure 5 As shown: S441: Features after fusion After convolution, the data is fed into three branches: a local branch, a large kernel branch, and a global branch, to enhance multi-scale representation. S442: The big core branch uses a size of [value missing]. Depthwise convolutions acquire global structural information and large-scale contextual information. and Depth-separable convolutions and square convolutions are used in parallel to capture strip-shaped contextual information; S443: Local branch usage Deep convolutional layers are used to modulate local signals and supplement features; S444: The global branch captures global information through dual-domain channel attention and frequency channel attention modules. First, the input features are subjected to global average pooling and spatial dimension compression to obtain channel information. Channel weights are obtained through 1×1 convolution. Simultaneously, the frequency domain features are transformed to the frequency domain by Fourier transform. After element-wise multiplication of the frequency domain features with the convolution output, the features are transformed back to the spatial domain by inverse Fourier transform to obtain the features. Based on features The output features of the dual-domain channel attention are generated by multiplying the channel weights element-wise. This enables global coarse-grained feature enhancement along the spatial-channel dimension; Output features of dual-channel attention The calculation formula is as follows: ; in, This represents element-wise multiplication. This represents the output characteristics after the inverse Fourier transform. This represents a 1×1 convolutional layer, and GAP represents global average pooling. S445: Secondly, the output features of dual-domain channel attention. The data is fed into the frequency channel attention module, and after depthwise separable convolutional dimensionality reduction in the frequency channel attention module, enhanced intermediate features are obtained. The enhanced intermediate features are obtained by using Fast Fourier Transform. Map to the frequency domain and construct a Query, Key, Value triplet; S446: The query and key elements are multiplied together and then fed into the ingress control unit Sigmoid, which is then multiplied element-wise with the value element to obtain the attention weight. After performing attention calculations on the spectrum, the attention weights are inversely Fourier transformed and returned to the spatial domain; thus, long-distance pixel dependencies are obtained while preserving edge structure information.

[0042] S447: Enhanced intermediate feature parallel input frequency channel enhancement module (FWM). The frequency channel enhancement module includes convolution, GELU activation function and sigmoid activation function processing. It performs residual repair in the frequency domain through a gating mechanism to enhance the response capability to high-frequency structures. It concatenates the frequency domain attention results with the frequency gating path and compresses and fuses them into the original dimension through the second convolutional layer to achieve unified modeling of multi-source semantics.

[0043] S448: Add the final output to the original input to obtain the output features. The formula used is: ; in, This indicates the inverse Fourier transform operation. Indicates a convolutional layer. This indicates the first convolutional layer. Indicates Fourier transform, This indicates the second convolutional layer. , This represents element-wise addition.

[0044] S449: The outputs of the local branch, the big kernel branch, and the global branch are weighted and fused, and then output after convolutional modulation.

[0045] S45: Minimum uncertainty query selection of the fused feature map input, which selects a fixed number of features from the feature sequence output by the encoder as the target query; S46: After being input into the decoder, the prediction head maps the confidence score and bounding box. The training process is supervised by adaptive threshold focus loss. Learning of difficult samples is enhanced by dynamic threshold decoupling and adaptive weight adjustment. S47: The network continuously iterates and updates its parameters to ensure that the loss parameters converge to a minimum value, thus completing the entire training process and obtaining the optimal weights.

[0046] S5: Input the enhanced image and the corresponding multi-category lesion labels into the multi-label detection network for training. Optimize the model parameters through an iterative process. Then, input the enhanced image to be analyzed into the trained multi-label detection network to determine its disease diagnosis result.

[0047] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0048] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0049] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0050] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for diagnosing oral diseases based on multi-path feature extraction and frequency enhancement, characterized in that, The specific steps include: S1: Collect a historical panoramic dental X-ray image dataset. Based on the preprocessing unit, automatically filter and remove low-quality images in the dataset that have motion artifacts, overexposure, or severe metal artifacts, so as to keep the filtered images. S2: Establish 13 categories of oral abnormality detection standards covering periapical lesions, caries, wisdom teeth, missing teeth, metal fillings, root canal fillings, implants, metal crowns, metal bridges, metal brackets, ceramic brackets, malocclusion, and supernumerary teeth, in order to determine the multi-category lesion labels of each screened image; S3: Image enhancement is performed on the selected images based on generative adversarial networks to obtain enhanced images with clear boundary morphology and prominent lesion features; S4: Construct a multi-label detection network including a spatial channel collaborative attention module, a full-dimensional dynamic convolution module, and an adaptive threshold focus loss to ensure the multi-label detection network's ability to perceive low-resolution, small lesion regions and small sample categories; S5: Input the enhanced image and the corresponding multi-category lesion labels into the multi-label detection network for training. Optimize the model parameters through an iterative process. Then, input the enhanced image to be analyzed into the trained multi-label detection network to determine its disease diagnosis result.

2. The oral disease diagnosis method based on multi-path feature extraction and frequency enhancement according to claim 1, characterized in that: A historical panoramic dental X-ray image dataset was obtained, which contains image data of various oral diseases.

3. The oral disease diagnosis method based on multi-path feature extraction and frequency enhancement according to claim 2, characterized in that: The following method is used to enhance the selected images based on generative adversarial networks: S31: Input the filtered image into the generator, which includes an encoder and a decoder. First, the encoder processes the filtered image by using a convolutional residual module and downsampling to extract feature maps. The convolutional residual module consists of two 3×3 convolutional layers, a batch normalization layer, and a Leaky ReLU activation function. S32: The feature map output by the encoder is sent to the decoder, and multi-scale features are fused through the upsampling module and convolutional residuals, and gradually restored to the original image size. The upsampling module is the operation module of the convolutional neural network. S33: For the feature maps between the encoder and decoder, information is passed through skip connections, and an attention gating mechanism is built on the skip connection path to filter the features from the encoding layer to the decoding layer; The formula upon which attention gating mechanisms are based is: ； in, This represents the feature map after filtering by the attention gating mechanism. d represents the feature of the k-th coding layer of the encoder, and d represents the feature of the output of the last decoder layer. This represents three 1×1 convolutional layers, where R represents the ReLU function. This represents the Sigmoid function; S34: Each layer of the decoder adds an extra path, which consists of a 1×1 convolutional layer and an upsampling layer. The feature map input to the decoder is processed through the extra path, and the feature map output by the extra path is output. S35: The feature map output by the decoder and the feature map output by the extra path are interpolated to the same size by bilinear interpolation, and then concatenated in the channel dimension. The concatenated features are fed into a 3×3 convolutional layer to obtain fused multi-scale information, and then added element-wise to the decoder output to obtain the superimposed features. S36: The superimposed feature input spatial channel attention module performs attention allocation on the decoder output features in both channel and spatial dimensions to generate the enhanced image.

4. The oral disease diagnosis method based on multi-path feature extraction and frequency enhancement according to claim 3, characterized in that: The spatial channel attention module in S36 includes: S361: The decoder's output features are subjected to global max pooling and global average pooling in each channel to extract global information in the channel dimension and generate global max pooling feature vectors and average pooling feature vectors containing the number of channels. S362: Input the global max pooling feature vector and the average pooling feature vector into the multilayer perceptron, add the global max pooling feature vector and the average pooling feature vector, and generate channel attention weights through the Sigmoid function; S363: Multiply the channel attention weights by the superimposed features to output the weighted channel features; S364: The weighted channel features are processed based on max pooling and average pooling, and spatial attention weights are generated after convolutional layers and sigmoid activation. S365: Adaptive learning is achieved by multiplying spatial attention weights with the superimposed features; The formula for calculating spatial attention weights is: ； ； in, This represents the weighted channel characteristics. Represents spatial attention weights. Indicates the features after superposition. and These represent global average pooling and max pooling based on spatial dimensions, respectively. and These represent global average pooling and max pooling based on spatial dimensions, respectively. MLP stands for Multilayer Perceptron. This is the Sigmoid activation function.

5. The oral disease diagnosis method based on multi-path feature extraction and frequency enhancement according to claim 4, characterized in that: The methods for training a network in a detection network are as follows: S41: The enhanced image is input into the detection network, and feature extraction is performed through the backbone network and embedded spatial channel collaborative attention. The outputs of the last three stages of the detection network are used as the input to the hybrid encoder. The detection network includes 5 convolutional layers, and the outputs of the 5 convolutional layers correspond to the fifth deep feature layer. Fourth deep features Third deep features Second deep features First deep feature The output of the last three stages is the fifth deep feature. Fourth deep features and third deep features ; S42: The final three stages of the extracted output-input hybrid encoder, deep features. After being converted into a one-dimensional vector, it is input into the intra-scale interaction module for feature enhancement. The output is adjusted to a two-dimensional sequence for cross-scale feature fusion to generate fused features. The intra-scale interaction module is used to realize contextual interaction of features within the same scale, strengthen the correlation and information transmission between features, and improve the discriminability and robustness of feature expression. S43: Enhanced features of the intra-scale interaction module The third deep features extracted from the backbone network Fourth deep feature Cross-scale fusion in pyramid networks, third deep features and fourth deep features After fusion, the data are fed into a full-dimensional dynamic convolution module for collaborative modeling using global and local features. S44: The full-dimensional dynamic convolution module is then integrated into the pyramid network for cross-scale fusion after processing. S45: Minimum uncertainty query selection of the fused feature map input, which selects a fixed number of features from the feature sequence output by the encoder as the target query; S46: After being input into the decoder, the prediction head maps the confidence score and bounding box. The training process is supervised by adaptive threshold focus loss. Learning of difficult samples is enhanced by dynamic threshold decoupling and adaptive weight adjustment. S47: The network continuously iterates and updates its parameters to ensure that the loss parameters converge to a minimum value, thus completing the entire training process and obtaining the optimal weights.

6. The oral disease diagnosis method based on multi-path feature extraction and frequency enhancement according to claim 5, characterized in that: The method for feature enhancement using a spatial channel collaborative attention module is as follows: S411: After the enhanced image input detection network, the input channel attention of the features extracted from the convolutional layers is decoupled along the height and width dimensions. Global average pooling is applied to each dimension to generate two unidirectional one-dimensional sequences. and sequence ; S412: Sequence and sequence Divide into K independent sub-features of the same size Features of the sub-score 'i' is used to index the sub-features. The sub-feature decomposition process is as follows: ； ； in, This represents the i-th sub-feature in the height dimension. Let K represent the i-th sub-feature in the width dimension, K represent the number of sub-feature divisions, and C represent the number of channels; S413: Based on sub-features Features of the sub-score Feature extraction is performed using lightweight convolution, and the features are concatenated. Normalization is then performed using K sets of normalization methods, and spatial attention weights are generated using the sigmoid activation function. and Intermediate features are obtained by multiplying spatial attention with features extracted from convolutional layers. ; The spatial attention weights are calculated as follows: ； ； in, For spatial attention weights in the height dimension, The spatial attention weights are defined in the width dimension, and the intermediate features are calculated as follows: ； in, express Sigmoid Normalization These represent K sets of normalization operations performed along the height and width dimensions, respectively. This represents the spatial structure information of the i-th sub-feature obtained after lightweight convolution operation. The subscript H is used to index the height dimension, the subscript W is used to index the width dimension, and X is the feature extracted by the convolutional layer. S414: The intermediate features are downsampled using average pooling to obtain the feature map. ; Feature map after downsampling Represented as: ； in, This indicates that pooling operations are performed using convolutional kernels of size K×K, with resolution ranging from... Readjusted to ; S415: Feature map after downsampling Normalization and 1×1 depthwise convolution generate query Q, key K, and value V; ； ； ； in, , , These represent linear projection operations that generate queries, keys, and values, respectively. S416: Perform relevance analysis based on query Q, key K, and value V to generate channel attention. Channel attention Multiply by the intermediate features to obtain the weighted features. , as the channel-enhanced feature map output; Channel attention The calculation method is as follows: ； The channel attention-weighted features are represented as follows: ； in, This indicates that a pooling operation is performed using a convolutional kernel of size 1×1. This indicates the aforementioned intermediate characteristics.

7. The oral disease diagnosis method based on multi-path feature extraction and frequency enhancement according to claim 6, characterized in that: S44 includes: S441: Features after fusion After convolution processing, the data is fed into three branches, which include a local branch, a large kernel branch, and a global branch. S442: The big core branch uses a size of [value missing]. Deep convolutions acquire global structural information and large-scale contextual information; S443: Local branch usage Deep convolutional layers are used to modulate local signals; S444: The global branch captures global information through dual-domain channel attention and frequency channel attention modules. First, the input features are subjected to global average pooling and spatial dimension compression to obtain channel information. Channel weights are obtained through 1×1 convolution. Simultaneously, the frequency domain features are transformed to the frequency domain by Fourier transform. After element-wise multiplication of the frequency domain features with the convolution output, the features are transformed back to the spatial domain by inverse Fourier transform to obtain the features. Based on features The output features of the dual-domain channel attention are generated by multiplying the channel weights element-wise. ; Output features of dual-channel attention The calculation formula is as follows: ； in, This represents element-wise multiplication. This represents the output characteristics after the inverse Fourier transform. This represents a 1×1 convolutional layer, and GAP represents global average pooling. S445: Secondly, the output features of dual-domain channel attention. The data is fed into the frequency channel attention module, and after depthwise separable convolutional dimensionality reduction in the frequency channel attention module, enhanced intermediate features are obtained. The enhanced intermediate features are obtained by using Fast Fourier Transform. Map to the frequency domain and construct a Query, Key, Value triplet; S446: The query and key elements are multiplied together and then fed into the ingress control unit Sigmoid, which is then multiplied element-wise with the value element to obtain the attention weight. After performing attention calculations on the spectrum, the attention weights are inversely Fourier transformed and returned to the spatial domain. S447: Enhanced intermediate feature parallel input frequency channel enhancement module (FWM). The frequency channel enhancement module includes convolution, GELU activation function and Sigmoid activation function processing. It performs residual repair in the frequency domain through a gating mechanism, concatenates the frequency domain attention result with the frequency gating path, and compresses and fuses it into the original dimension through the second convolutional layer to achieve unified modeling of multi-source semantics. S448: Add the final output to the original input to obtain the output features. The formula used is: ； in, This indicates the inverse Fourier transform operation. Indicates a convolutional layer. This indicates the first convolutional layer. Indicates Fourier transform, This indicates the second convolutional layer. , This represents element-wise addition. S449: The outputs of the local branch, the big kernel branch, and the global branch are weighted and fused, and then output after convolutional modulation.

Images