Dental caries identification method and equipment
Through the cascaded setting and feature map processing of multiple image recognition mechanisms, characterization feature maps are generated, which solves the accuracy problem of the caries recognition algorithm under complex and variable image conditions, and realizes efficient caries recognition under different shooting conditions.
Patent Information
- Application Number
- CN202510503895.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-07-25
AI Technical Summary
When faced with dental images under different shooting conditions, the existing caries recognition algorithm has a high probability of missed detection and missed detection, making it difficult to accurately identify caries under complex and changeable image conditions.
Multiple graph identification mechanisms are used to cascade the features of teeth are extracted through one-dimensional and covariant processing, and the confidence feature map is adjusted using perturbation feature maps to generate characterization feature maps, and the caries category is determined in combination with a classifier.
It improves the accuracy and reliability of caries recognition, reduces misjudgment and misjudgment, and can accurately identify caries under different shooting conditions.
Smart Images
Figure CN120374583A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of auxiliary diagnosis of dental caries, and particularly relates to a method and device for identifying dental caries. Background Art
[0002] In China, dental caries is still one of the most common dental diseases. Doing a good job in the whole chain management of "prevention - screening - treatment" is an important measure for preventing and treating dental caries. Some studies have shown that early screening and remineralization intervention for dental caries can effectively cure dental caries at low cost and reduce the medical burden to the greatest extent. In the face of the reality of insufficient total medical resources and uneven distribution in China, using advanced technologies such as artificial intelligence to assist in screening dental caries can greatly reduce the workload of medical staff, narrow the difference in medical conditions in different regions, and promote medical equality. However, in the face of large differences in the quality distribution of input dental images, the current algorithms for identifying dental caries still have a high probability of missed detection and false detection, and there is still much room for improvement. Summary of the Invention
[0003] In view of this, the present invention provides a method and device for identifying dental caries to improve the reliability of dental caries identification when facing complex and changeable dental images.
[0004] To achieve the above object, the solution adopted by the present invention is: A method for identifying dental caries, comprising the following steps: Extracting image features of the image to be recognized according to a plurality of configured image recognition mechanisms; the image to be recognized includes an image of a tooth to be diagnosed, and the plurality of image recognition mechanisms are cascaded; Performing one-dimensional processing on the latent features output by the selected image recognition mechanism Figure 1 To obtain a confidence feature map; Performing covariant processing on the latent feature maps output by at least one of the image recognition mechanisms to obtain at least one perturbation feature map; Adjusting the confidence feature map using the perturbation feature map to obtain a characterization feature map; Determining the category to which the image to be recognized belongs according to the characterization feature map.
[0005] Further, the one-dimensional processing includes: respectively calculating the average value of the largest M eigenvalues in each channel of the feature map to obtain a first transition feature map; the first transition feature map undergoes a first adaptation process to obtain the confidence feature map; the value of M is equal to the height of the latent feature map output by the selected image recognition mechanism.
[0006] Further, the first adaptation process includes a first activation layer, a first linear layer, and a second activation layer that are performed in sequence.
[0007] Further, the covariant processing includes: calculating the average value of the largest N eigenvalues at each spatial position of the feature map to obtain a second transitional feature map; the second transitional feature map undergoes a second adaptation process to obtain the perturbation feature map; the value of N is less than the channel size of the feature map for covariant processing.
[0008] Further, the second adaptation process includes, in sequence, a first convolutional layer, a third activation layer, a flatten layer, a second linear layer, and a fourth activation layer.
[0009] During the process of one-dimensionalizing the feature map, the average value of the largest M eigenvalues in each channel of the feature map is calculated respectively, rather than simply obtaining the largest eigenvalue. Correspondingly, during the covariant processing, the average value of the largest N eigenvalues at each spatial position of the feature map is calculated. In combination with the first adaptation process and the second adaptation process, the dental caries recognition method can adapt to dental images under various different shooting conditions, such as different lighting intensities, angles, and changes in the position and size of the teeth in the image, etc. This helps to improve the robustness of the dental caries recognition method in practical applications, and can extract effective features as accurately as possible regardless of the image quality.
[0010] Further, the process of adjusting the confidence feature map using one of the perturbation feature maps includes: using a second convolutional layer and a fifth activation layer to process the result of the element-wise summation of the perturbation feature map and the confidence feature map; The process of adjusting the confidence feature map using multiple perturbation feature maps includes: multiple perturbation feature maps are stacked row by row and then processed by a third convolutional layer to obtain an integrated feature map; then, a fourth convolutional layer and a sixth activation layer are used to process the result of the element-wise summation of the integrated feature map and the confidence feature map.
[0011] During the process of stacking these perturbation feature maps row by row to obtain an integrated feature map, these diverse feature information can be concentrated together, enabling the model to learn the combined relationships between these different features, thereby more comprehensively understanding the feature expression of the dental image. This integration method greatly enriches the dimension of the features, laying a foundation for more accurate subsequent dental caries recognition. Further, determining the category to which the image to be recognized belongs according to the characterization feature map includes: using a classifier to calculate the characterization feature map to obtain the category to which the image to be recognized belongs.
[0012] Further, the image recognition mechanism is the Inception module in GoogLeNet, or, The image recognition mechanism is the residual module in ResNe101, or, The image recognition mechanism is the DenseBlock module in DenseNet.
[0013] Furthermore, the internal calculation process of the image recognition mechanism includes: The feature maps input into the image recognition mechanism are respectively processed by a max pooling module and an average pooling module, and a first feature map and a second feature map are correspondingly obtained; The result obtained by taking the difference between the elements of the first feature map and the second feature map undergoes first feature capture to obtain a third feature map; The result obtained by multiplying the elements of the second feature map and the third feature map undergoes second feature capture to obtain a fourth feature map; The result obtained by summing the elements of the first feature map and the fourth feature map undergoes third feature capture to obtain the latent feature map output by the image recognition mechanism.
[0014] Compared with the existing feature extraction module, the above-mentioned internal mechanism of the image recognition mechanism uses differential analysis to highlight key features, greatly improving the sensitivity and recognition ability of the model to caries features, and enabling it to more accurately "understand" the details of tooth images. For example, it can better identify the subtle differences between carious parts and normal teeth in terms of texture, color transition, etc., and has a better recognition effect on the tiny lesions in the early stage of caries.
[0015] A caries recognition device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned method are implemented.
[0016] The beneficial effects of the present invention are: The popularization of smartphones has made it very convenient to collect and transmit tooth images. Through the combination of image media and artificial intelligence technology, caries can be quickly screened. However, the image quality of tooth images taken by mobile phones with different prices and brands varies greatly, and most users lack professional photography knowledge. These factors result in a large gap between the obtained tooth image quality and the ideal situation. The present invention uses the perturbation feature map to directly adjust the one-dimensionalized confidence feature map, so that the representation feature map not only contains the basic features of the original image, but also integrates the key features highlighted after covariant processing and stable feature representations, making the image feature representation in the representation feature map more rich and perfect. In the face of complex and variable tooth images in practical applications, the model can make more stable judgments based on this optimized feature map, improving the accuracy and reliability of recognition. Even if there are some noises in the image or the shooting conditions are not ideal, the model can accurately identify caries relying on the rich feature information in the representation feature map, reducing the occurrence of misjudgment and missed judgment. Description of the Drawings
[0017] Figure 1 It is a schematic flowchart of a caries recognition method for an implementation manner. Detailed implementation manners
[0018] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are some, but not all, of the embodiments of the technical solutions of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments described in this application document without creative efforts shall fall within the scope of protection of the technical solutions of the present application.
[0019] As Figure 1 shown, a dental caries recognition method includes the following steps: S1. Obtain an image to be recognized by using an electronic device such as a smart phone. The image to be recognized includes an image of a tooth to be diagnosed, and the image of the tooth to be diagnosed occupies a part of the area of the image to be recognized. In some implementation manners, the captured image is directly input into a configured image recognition mechanism to extract image features, and the image to be recognized includes images of multiple teeth. In some implementation manners, technologies such as object detection can be used to preprocess the captured image so that only an image of one tooth is included in the processed image to be recognized, and then the preprocessed image to be recognized is input into the image recognition mechanism. In some implementation manners, technologies such as image segmentation can be used to preprocess the captured image, separate and remove irrelevant backgrounds around the tooth, and then input the preprocessed image to be recognized into the image recognition mechanism.
[0020] The number of image recognition mechanisms can be set to 3, 4 or more. The first image recognition mechanism receives the image to be recognized as input. The multiple image recognition mechanisms are cascaded, which means that the feature map output at the end of the previous image recognition mechanism is directly input into the subsequent image recognition mechanism. In the present invention, the feature map output at the end after internal calculation of each image recognition mechanism is called a latent feature map.
[0021] In some implementation manners, the image recognition mechanism can be directly implemented by using an existing trained feature extraction module. For example, the image recognition mechanism can be the Inception module in GoogLeNet, or the residual module in ResNe101, or the DenseBlock module in DenseNet. As a preferred implementation manner, the internal calculation process of the image recognition mechanism can include the following steps: The feature maps input to the image recognition mechanism are processed by a max pooling module and an average pooling module respectively, and a first feature map and a second feature map are obtained correspondingly. Inside the max pooling module, there is at least a locally max pooling layer and an activation function (such as ReLU, GELU, Swish or sigmoid, etc.) arranged in cascade. For example, the pooling window of the locally max pooling layer can be 2*2, the stride is equal to the width of the max pooling window, and the pooling window slides along the spatial dimension of the feature map. Inside the average pooling module, there is at least a locally average pooling layer and an activation function (such as ReLU, GELU, Swish or sigmoid, etc.) arranged in cascade. For example, the pooling window of the locally average pooling layer can be 2*2, the stride is equal to the width of the average pooling window, and the pooling window slides along the spatial dimension of the feature map.
[0022] The elements in the first feature map are subtracted from the corresponding elements in the second feature map, and then the feature map obtained by subtracting (the first feature map and the second feature map) undergoes a first feature capture to obtain a third feature map; the result obtained by multiplying the corresponding elements of the second feature map and the third feature map undergoes a second feature capture to obtain a fourth feature map; the result obtained by summing the corresponding elements of the first feature map and the fourth feature map undergoes a third feature capture to obtain the implicit feature map output by the image recognition mechanism.
[0023] The above-mentioned feature capture can be implemented by many well-known algorithms, as long as appropriate feature information can be captured from the feature map. For example, the first feature capture can include a first convolution operation and a first activation operation in sequence, the second feature capture can include a second convolution operation and a second activation operation in sequence, and the third feature capture can include a third convolution operation and a third activation operation in sequence. The first convolution operation, the second convolution operation, and the third convolution operation can all be completed by a common convolution operation layer, and there is no special limitation on the size of its convolution kernel. For example, it can be 2*2 or 3*3 or 4*4, etc. The first activation operation, the second activation operation, and the third activation operation can all be conventional non-linear activation functions.
[0024] S2. Perform dimensionality reduction processing on the implicit features output by the selected image recognition mechanism Figure 1 to obtain a confidence feature map. Preferably, perform dimensionality reduction processing on the implicit features output by the image recognition mechanism at the later position. Figure 1 In this embodiment, perform dimensionality reduction processing on the implicit features output by the last image recognition mechanism Figure 1 to obtain a confidence feature map.
[0025] Specifically, the one-dimensional processing includes: calculating the average value of the largest M eigenvalues in each channel of the latent feature map respectively to obtain a first transitional feature map; the first transitional feature map undergoes a first adaptation process to obtain a confidence feature map; the value of M is equal to the height of the latent feature map output by the selected image recognition mechanism. For example, assuming the height * width * channel size of the feature map is 200 * 200 * 64, then this feature Figure 1 contains a total of 64 channels, and each channel contains 40,000 eigenvalues. During the one-dimensional processing, the average value of the largest 200 eigenvalues among the 40,000 eigenvalues in each channel is calculated respectively, and the obtained first transitional feature map is a one-dimensional vector with a length of 64.
[0026] Furthermore, the first adaptation process includes: calculating the first transitional feature map by using a cascaded first activation layer, a first linear layer, and a second activation layer. The first linear layer can be implemented by a fully connected layer, and the first activation layer and the second activation layer can each independently select a single activation function such as ReLU, GELU, Swish, or sigmoid to implement, or can each independently select multiple activation functions to be combined in series, parallel, or series-parallel manners.
[0027] S3. Covariant processing is performed on the latent feature maps output by at least one image recognition mechanism to obtain at least one perturbed feature map (one corresponding perturbed feature map is obtained after covariant processing of one latent feature map). Preferably, covariant processing is performed on the latent feature maps output by multiple image recognition mechanisms to obtain multiple perturbed feature maps. In this embodiment, covariant processing is performed on all the latent feature maps output by the image recognition mechanisms. For example, if the number of image recognition mechanisms is set to 4, then covariant processing is independently performed on the latent feature maps output by the 4 image recognition mechanisms respectively.
[0028] Specifically, the covariant processing includes: calculating the average value of the largest N eigenvalues at each spatial position of the feature map to obtain a second transitional feature map, and the second transitional feature map is essentially a two-dimensional matrix with 1 channel. The second transitional feature map undergoes a second adaptation process to obtain a perturbed feature map. The value of N is less than the channel size of the feature map for covariant processing, and N is a positive integer. That is to say, for which latent feature map the covariant processing is performed, the value of N is set to be less than the channel size of this latent feature map during the covariant processing. Preferably, the value range of N is (Z / 2) ≤ N ≤ (Z / 16), where Z is the channel size of the feature map for covariant processing. Further preferably, Z is a multiple of 8, and N = (Z / 8). For example, if the height * width * channel size of the feature map for covariant processing is 200 * 200 * 64, then each spatial position of this feature map contains 64 eigenvalues. Preferably, the average value of the largest 8 eigenvalues among the 64 eigenvalues at each spatial position of the feature map is calculated, and the obtained second transitional feature map has a height * width * channel size of 200 * 200 * 1.
[0029] The second adaptation process includes: calculating the second intermediate feature map using a cascaded first convolutional layer, third activation layer, flatten layer (unfolding layer), second linear layer, and fourth activation layer. The first convolutional layer can use conventional convolution calculation, but preferably deformable convolution calculation. The second linear layer can be implemented using a fully connected layer. The third activation layer and the fourth activation layer can each independently select a single activation function such as ReLU, GELU, Swish, or sigmoid for implementation, or can each independently select multiple activation functions and combine them in a series, parallel, or series-parallel manner for implementation.
[0030] S4. Use the perturbed feature map to adjust the confidence feature map to obtain the characterization feature map. The perturbed feature map and the confidence feature map can be vectors of equal length.
[0031] When using one perturbed feature map to adjust the confidence feature map, the adjustment process includes: first summing the corresponding elements of the perturbed feature map and the confidence feature map, and then using a cascaded second convolutional layer (the kernel size can be 1*1) and fifth activation layer to process the summation result (of the corresponding elements of the perturbed feature map and the confidence feature map) to obtain the characterization feature map.
[0032] When using multiple perturbed feature maps to adjust the confidence feature map, the adjustment process includes: first stacking the multiple perturbed feature maps row by row. Stacking row by row means arranging multiple vectors in sequence as the rows of a matrix. For example, if A perturbed feature maps of length B are stacked row by row, a two-dimensional matrix with a width of B and a height of A is obtained. The matrix obtained after stacking is processed by a third convolutional layer to obtain an integrated feature map, and the integrated feature map is a vector. Exemplarily, in some embodiments, the width size of the kernel of the third convolutional layer is 1, the height size of the kernel is A, and the convolutional window slides along the width direction of the matrix (stride is 1) to obtain an integrated feature map of length B. In other embodiments, the width size of the kernel of the third convolutional layer is A, the height size of the kernel is A, and the convolutional window slides along the width direction of the matrix (stride is 1). At this time, 0 padding can be performed on the horizontal two ends of the matrix generated after stacking so that the length of the obtained integrated feature map is still B. Then use a cascaded fourth convolutional layer (the kernel size can be 1*1) and sixth activation layer to process the result of summing the corresponding elements of the integrated feature map and the confidence feature map to obtain the characterization feature map. The fifth activation layer and the sixth activation layer can each independently select a single activation function such as ReLU, GELU, Swish, or sigmoid for implementation, or can each independently select multiple activation functions and combine them in a series, parallel, or series-parallel manner for implementation.
[0033] S5. Determine the category to which the image to be recognized belongs according to the representation feature map. Specifically, use an existing classifier (such as a softmax layer) to calculate the representation feature map, and the category to which the image to be recognized belongs can be obtained. The representation feature map is a one-dimensional vector, and its length matches the number of possible categories of the image to be recognized. For example, in some embodiments, the image to be recognized can be binary-classified, as caries or healthy, and at this time the length of the representation feature map can be 2. In other embodiments, the image to be recognized can also be four-classified, as high risk, medium risk, low risk, and healthy, and at this time the length of the representation feature map can be 4.
[0034] The above-mentioned multiple image recognition mechanisms extract the image features of the image to be recognized and perform dimensionality reduction processing on the implicit features output by the selected image recognition mechanism Figure 1 These steps of covariance processing on the implicit feature maps output by at least one image recognition mechanism, adjusting the confidence feature map using the perturbation feature map, and determining the category to which the image to be recognized belongs according to the representation feature map can be automatically calculated and completed by an artificial neural network algorithm model programmed to run on devices such as computers. As is well known, by training the network, the configuration of internal parameters of parts such as the image recognition mechanism can be completed.
[0035] The above-described embodiments only represent the specific implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.
Claims
1. A method for identifying dental caries, characterized in that, It includes the following steps: Multiple image recognition mechanisms configured as completed extract the image features of the image to be recognized; the image to be recognized contains the image of the tooth to be diagnosed, and the multiple image recognition mechanisms are cascaded; Perform one-dimensional processing on the latent feature map output by the selected image recognition mechanism to obtain a confidence feature map; Perform covariant processing on the latent feature maps output by at least one of the image recognition mechanisms to obtain at least one perturbation feature map; Use the perturbation feature map to adjust the confidence feature map to obtain a characterization feature map; Determine the category to which the image to be recognized belongs according to the characterization feature map.
2. The caries recognition method according to claim 1, wherein The one-dimensional processing includes: respectively calculating the average value of the largest M eigenvalue in each channel of the feature map to obtain a first transition feature map; the first transition feature map undergoes a first adaptation process to obtain the confidence feature map; the value of M is equal to the height of the latent feature map output by the selected image recognition mechanism.
3. The caries recognition method according to claim 2, characterized in that, The first adaptation process includes a first activation layer, a first linear layer, and a second activation layer in sequence.
4. The caries recognition method according to claim 1, characterized in that The covariant processing includes: calculating the average value of the largest N eigenvalue at each spatial position of the feature map to obtain a second transition feature map; the second transition feature map undergoes a second adaptation process to obtain the perturbation feature map; the value of N is less than the channel size of the feature map for covariant processing.
5. The caries identification method according to claim 4, characterized in that, The second adaptation process includes a first convolutional layer, a third activation layer, a flatten layer, a second linear layer, and a fourth activation layer in sequence.
6. The caries identification method according to claim 1, characterized in that The process of using one perturbation feature map to adjust the confidence feature map includes: using a second convolutional layer and a fifth activation layer to process the result of the element-wise summation of the perturbation feature map and the confidence feature map; The process of using multiple perturbation feature maps to adjust the confidence feature map includes: after the multiple perturbation feature maps are stacked row by row and processed by a third convolutional layer to obtain an integrated feature map; then using a fourth convolutional layer and a sixth activation layer to process the result of the element-wise summation of the integrated feature map and the confidence feature map.
7. The caries identification method according to claim 1, characterized in that Determining the category to which the image to be recognized belongs according to the characterization feature map includes: using a classifier to calculate the characterization feature map to obtain the category to which the image to be recognized belongs.
8. The dental caries recognition method according to claim 1, wherein the image recognition mechanism is the Inception module in GoogLeNet, or the image recognition mechanism is the residual module in ResNe101, or the image recognition mechanism is the DenseBlock module in DenseNet.
9. The caries recognition method according to claim 1, wherein, The internal calculation process of the image recognition mechanism includes: The feature map input to the image recognition mechanism is respectively processed by a max pooling module and an average pooling module to correspondingly obtain a first feature map and a second feature map; The result obtained by subtracting the elements of the first feature map and the second feature map undergoes a first feature capture to obtain a third feature map; The result obtained by multiplying the elements of the second feature map and the third feature map undergoes a second feature capture to obtain a fourth feature map; The result obtained by summing the elements of the first feature map and the fourth feature map undergoes a third feature capture to obtain the latent feature map output by the image recognition mechanism.
10. A dental caries recognition device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.