Tooth lesion detection method based on global feature optimization Transform
By introducing a Transformer structure with a globally driven semantic enhancement module in dental lesions detection, the problem of insufficient detection of small-sized lesions in the prior art is solved, and higher detection accuracy and richness of feature representation are achieved, which significantly improves dental lesions diagnosis efficiency and accuracy.
Patent Information
- Application Number
- CN202411176167.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-08-12
- Filing Date
- 2024-08-26
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-08-26
AI Technical Summary
The existing dental lesion detection method based on deep learning is insufficient for small-sized lesions in magnetic resonance imaging images, and there are limitations in global feature extraction and insufficient richness of feature representation, making it difficult to accurately detect dental lesions in complex backgrounds.
Using the method of optimizing Transformer based on global features, the improved Transformer structure is constructed by introducing global driver semantic enhancement modules, including encoder, global driver semantic enhancement module and decoder, to capture global features of different scales, and fusion of multi-scale features and Integrate to enhance the detection ability of small-sized dental lesions.
It improves the detection accuracy and efficiency of small-sized lesions, enriches the feature representation ability, reduces the impact of calculation complexity, and significantly improves the accuracy and efficiency of dental lesions diagnosis.
Smart Images

Figure CN119067950A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of tooth image data processing, and in particular to a tooth lesion detection method based on global feature optimization Transformer. Background Art
[0002] Dental lesions include caries, pulpitis and periodontal diseases, which often present diverse morphological and structural characteristics. These lesions have significant diagnostic and therapeutic value in clinical practice. Accurate detection and diagnosis of dental lesions are crucial for formulating effective treatment plans. Among them, magnetic resonance imaging (MRI) technology is widely used in the field of medical imaging, especially in the detection and diagnosis of dental lesions. MRI has the advantages of non-invasive, high contrast and multi-parameter imaging, which can provide detailed information about teeth and their surrounding tissues, and has become an important means of detecting dental lesions. However, in MRI images, these lesions may have small differences in grayscale values and textures from normal tooth tissues and adjacent anatomical structures. In addition, the complex morphology and dense arrangement of teeth make it more difficult to locate the lesion area and accurately determine the boundaries.
[0003] Due to the complexity of MRI images and the diversity of tooth structures, accurate detection of dental lesions, especially small-sized lesions, has always been a challenge. Currently, computer-aided detection (CAD) systems have made some progress in medical imaging analysis, especially deep learning-based methods have shown significant advantages. However, existing deep learning-based tooth detection methods have the following problems: (1) insufficient detection effect on small-scale lesions; (2) limitations in global feature extraction; (3) insufficient richness of feature representation. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a dental lesion detection method based on global feature optimization Transformer, which is used to solve the deficiencies of existing dental detection methods in processing small-scale lesions, especially to solve the problem of difficulty in detection under the complex background of magnetic resonance imaging images.
[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:
[0006] A method for detecting dental lesions based on global feature optimized Transformer, comprising the following steps:
[0007] S1, obtain the magnetic resonance imaging image of the teeth and input it into the backbone network to extract multi-scale features;
[0008] S2, introduce the global driven semantic enhancement module and build an improved Transformer structure;
[0009] The improved Transformer structure includes an encoder, a global driven semantic enhancement module, and a decoder connected in sequence;
[0010] A global-driven semantic enhancement module to capture global features at different scales and enhance the detection of small-sized dental lesions;
[0011] S3, inputting the multi-scale features into the improved Transformer structure for training, and obtaining a trained improved Transformer structure for identifying dental lesions;
[0012] S4. Obtain a new magnetic resonance imaging image of the tooth and input it into the trained improved Transformer structure to obtain the tooth lesion detection result.
[0013] Furthermore, the multi-scale features in step S1 are:
[0014] {S3, S4, S5}
[0015] Among them, S3 represents the third scale feature, S4 represents the fourth scale feature, and S5 represents the fifth scale feature.
[0016] Furthermore, the global driven semantic enhancement module includes a first convolution layer module, a second convolution layer module, a first weight extraction layer module, a second weight extraction layer module, a third weight extraction layer module, a first upsampling module, a second upsampling module, and a convolution block module.
[0017] Furthermore, step S3 specifically includes:
[0018] S31, after position encoding, the S3 scale feature in the multi-scale feature is added to the S3 scale feature, and then input into the encoder of the improved Transformer structure for encoding operation to obtain the global feature F3;
[0019] S32, inputting the global feature F3, the S4 scale feature in the multi-scale feature, and the S5 scale feature into the global driven semantic enhancement module for multi-scale feature fusion to obtain a multi-scale fusion feature image;
[0020] S33. Input the multi-scale fusion feature image into the decoder of the improved Transformer structure to classify and locate the area of interest of dental lesions, and obtain a trained improved Transformer structure for identifying dental lesions.
[0021] Further, step S32 specifically includes:
[0022] S321, input the global feature F3 into the first convolutional layer module for size alignment operation, obtain the global feature F3 of the same size as the scale feature S5, and input it into the first weight extraction layer module for convolution activation operation to obtain the weight W s5 , and the weight W s5 After element-by-element multiplication with the S5 scale feature, element-by-element addition with the S5 scale feature is performed to obtain the first weighted feature S′5, that is:
[0023]
[0024] Among them, conv represents the convolution operation, ↓ represents the downward resizing operation, and sigmoid represents the activation function;
[0025] S322, after the first weighted feature S′5 is input into the first upsampling module for upsampling operation, it is concatenated with the scale feature S4 to obtain the concatenated feature S′4, that is:
[0026] S′4=C(↑S′5,S4)
[0027] Among them, C represents concatenation, ↑ represents upsampling;
[0028] S323, input the global feature F3 into the second convolutional layer module for size alignment operation, obtain the global feature F3 of the same size as the scale feature S4, and input it into the second weight extraction layer module for convolution activation operation to obtain the weight W s4 , and the weight W s4 After element-by-element multiplication with the concatenated feature S′4, element-by-element addition with the concatenated feature S′4 is performed to obtain the second weighted feature S″4, namely:
[0029]
[0030] S324, input the second weighted feature S″4 into the second upsampling module for upsampling operation, and then concatenate it with the global feature F3 to obtain the concatenated global feature F′3, and at the same time input the global feature F3 into the third weight extraction layer module for convolution activation operation to obtain the weight W s3 , the weight W s3 After element-by-element multiplication with the concatenated global feature F′3, element-by-element addition with the concatenated global feature F′3 is performed to obtain the third weighted feature F″3, namely:
[0031]
[0032] S325. Input the third weighted feature F″3 into the convolution block module for feature integration and optimization to obtain a multi-scale fusion feature image.
[0033] Further, step S33 specifically includes:
[0034] S331, position encoding the multi-scale fusion feature image to generate input features containing position information;
[0035] S332, embedding the decoder into a multi-head self-attention mechanism to capture the correlation within the decoder embedding and generate new decoder embedding features;
[0036] S333, interacting the input features containing the position information with the new decoder embedding features through a multi-head cross attention mechanism, obtaining the encoder information related to the current decoding position, and sequentially inputting the feedforward neural network, sum normalization, and multi-layer perceptron to perform multi-layer full connection operations, residual addition and normalization processing, and classification and positioning operations, respectively, to generate the predicted classification confidence and the positioning bounding box offset;
[0037] S334, generating a new bounding box according to the predicted localized bounding box offset and the predefined anchor box, and obtaining the localization result and classification result of the dental lesion region of interest;
[0038] S335. The positioning result of the dental lesion region of interest, the classification result and the true label are simultaneously input into the loss function, and the loss function is used to calculate the gradient to update the network parameters of the improved Transformer structure to obtain a trained improved Transformer structure.
[0039] The present invention has the following beneficial effects:
[0040] 1. The present invention proposes a dental lesion detection method based on global feature optimization Transformer. By introducing a global driven semantic enhancement module, an improved Transformer structure is constructed for dental lesion detection. It can not only improve the detection accuracy and efficiency of small-scale lesions, but also enrich the feature representation capability, reduce the influence of computational complexity, and improve the efficiency and accuracy of dental lesion diagnosis.
[0041] 2. In order to help the encoder of the improved Transformer structure distinguish the position information of the input features and thus better capture the relative relationship between features, position encoding is added to the S3 scale features before encoding, and then added to the S3 scale features and input into the encoder for extracting global features, effectively capturing the global context within the S3 scale features and providing a comprehensive understanding of the medical scenario. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 A schematic diagram of the process of a dental lesion detection method based on global feature optimization Transformer proposed in the present invention;
[0043] Figure 2 Schematic diagram of the improved Transformer structure. DETAILED DESCRIPTION
[0044] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0045] like Figure 1 As shown, a method for detecting dental lesions based on global feature optimized Transformer includes the following steps S1-S4:
[0046] S1. Obtain the magnetic resonance imaging image of the teeth and input it into the backbone network to extract multi-scale features.
[0047] Specifically, the multi-scale features in step S1 are:
[0048] {S3, S4, S5}
[0049] Among them, S3 represents the third scale feature, S4 represents the fourth scale feature, and S5 represents the fifth scale feature.
[0050] In this embodiment, a magnetic resonance imaging image of the teeth is first obtained and input into the backbone network to extract multi-scale features, wherein the extracted multi-scale features are {S3, S4, S5}; the multi-scale features {S3, S4, S5} extracted from the backbone network differ in spatial resolution and semantic information, i.e., S3 has higher spatial resolution and lower semantic information, and is suitable for detecting small-sized objects, so it can capture finer details; S4 and S5 have higher semantic information but lower spatial resolution, which helps to provide rich semantic context to identify and classify regions of interest (ROI).
[0051] S2. Introduce a global-driven semantic enhancement module and construct an improved Transformer structure; wherein the improved Transformer structure includes an encoder, a global-driven semantic enhancement module and a decoder connected in sequence; the global-driven semantic enhancement module is used to capture global features of different scales and enhance the detection of small-sized dental lesions.
[0052] Specifically, the global driven semantic enhancement module includes a first convolution layer module, a second convolution layer module, a first weight extraction layer module, a second weight extraction layer module, a third weight extraction layer module, a first upsampling module, a second upsampling module, and a convolution block module.
[0053] In this embodiment, the improved Transformer structure is as follows Figure 2 As shown, Figure 2 The improved Transformer structure includes an encoder, a global driven semantic enhancement module and a decoder connected in sequence, wherein the global driven semantic enhancement module includes a first convolutional layer module, a second convolutional layer module, a first weight extraction layer module, a second weight extraction layer module, a third weight extraction layer module, a first upsampling module, a second upsampling module and a convolutional block module; the working principle of the improved Transformer structure is: after position encoding S3, the element-by-element summation with S3 is performed, and the global feature F3 is extracted by inputting the encoder, and the global feature F3, S4 and S5 are integrated and optimized by using the global driven semantic enhancement module to enhance the overall feature representation capability, and the details of S3 are combined with the semantic richness of S4 and S5, and finally the entire global feature information (multi-scale fusion feature image) is obtained and input into the decoder to generate classification and positioning results. Among them, the first convolutional layer module and the second convolutional layer module are used to match and align the size of the global feature F3 with the multi-scale feature; the first weight extraction layer module, the second weight extraction layer module, and the third weight extraction layer module are used to extract different weight information respectively; the first upsampling module and the second upsampling module are used to perform upsampling and downsampling operations respectively, and the convolution block module is used to integrate and optimize the features to generate rich global feature representations, that is, multi-scale fusion feature images. Therefore, the global driven semantic enhancement module significantly improves the ability of the improved Transformer structure to detect small lesions by focusing on enriching the global context features with high-level semantic features, enriches the representation ability of the features, and also reduces the impact of computational complexity.
[0054] S3. Input the multi-scale features into the improved Transformer structure for training to obtain a trained improved Transformer structure for identifying dental lesions.
[0055] Specifically, step S3 specifically includes S31-S33:
[0056] S31, after position encoding the S3 scale feature in the multi-scale feature, add it to the S3 scale feature, and input it into the encoder of the improved Transformer structure for encoding operation to obtain the global feature F3.
[0057] In this embodiment, in order to help the encoder of the improved Transformer structure distinguish the position information of the input features and better capture the relative relationship between the features, the position encoding is added to the S3 scale feature before encoding, and then added to the S3 scale feature and input into the encoder for extracting the global feature. The purpose of this design is to effectively capture the global context within the S3 scale feature and provide a comprehensive understanding of the medical scene.
[0058] S32, input the global feature F3, the S4 scale feature in the multi-scale feature, and the S5 scale feature into the global driven semantic enhancement module for multi-scale feature fusion to obtain a multi-scale fusion feature image.
[0059] In this embodiment, the global driven semantic enhancement module is used to integrate and optimize the global feature F3, the S4 scale feature and the S5 scale feature in the multi-scale feature to enhance the overall feature representation, so as to combine the details of the S3 scale feature with the semantic richness of the S4 scale feature and the S5 scale feature, thereby realizing dental lesion detection.
[0060] Specifically, step S32 specifically includes S321-S325:
[0061] S321, input the global feature F3 into the first convolutional layer module for size alignment operation, obtain the global feature F3 of the same size as the scale feature S5, and input it into the first weight extraction layer module for convolution activation operation to obtain the weight W s5 , and the weight W s5 After element-by-element multiplication with the S5 scale feature, element-by-element addition with the S5 scale feature is performed to obtain the first weighted feature S′5, that is:
[0062]
[0063] Among them, conv represents the convolution operation, ↓ represents the downward resizing operation, and sigmoid represents the activation function.
[0064] In this embodiment, the stride of the first convolutional layer module (conv4×4) is 4, and the role of this module is to match the size of the global feature F3 with the scale feature S5; the first weight extraction layer module consists of a conv1×1 layer and a sigmoid activation function, and its role is to generate a weight W specific to the global feature. s5 ; and use the weight W s5 Multiply it with the S5 scale feature to embed the global context information into the S5 scale feature, and then obtain the first weighted feature S′5 through residual connection, that is, addition operation; this process embodies the semantic enhancement driven by global features.
[0065] S322, after the first weighted feature S′5 is input into the first upsampling module for upsampling operation, it is concatenated with the scale feature S4 to obtain the concatenated feature S′4, that is:
[0066] S′4=C(↑S′5,S4)
[0067] Among them, C represents concatenation and ↑ represents upsampling.
[0068] In this embodiment, the purpose of upsampling is to match the size of the first weighted feature S′5 with the size feature S4.
[0069] S323, input the global feature F3 into the second convolutional layer module for size alignment operation, obtain the global feature F3 of the same size as the scale feature S4, and input it into the second weight extraction layer module for convolution activation operation to obtain the weight W s4 , and the weight W s4 After element-by-element multiplication with the concatenated feature S′4, element-by-element addition with the concatenated feature S′4 is performed to obtain the second weighted feature S″4, namely:
[0070]
[0071] In this embodiment, the purpose of this process is to use the extracted weight W s4 The concatenated feature S′4 is further enhanced. The second weight extraction layer module is also composed of a conv1×1 layer and a sigmoid activation function.
[0072] S324, input the second weighted feature S″4 into the second upsampling module for upsampling operation, and then concatenate it with the global feature F3 to obtain the concatenated global feature F′3, and at the same time input the global feature F3 into the third weight extraction layer module for convolution activation operation to obtain the weight W s3 , the weight W s3 After element-by-element multiplication with the concatenated global feature F′3, element-by-element addition with the concatenated global feature F′3 is performed to obtain the third weighted feature F″3, namely:
[0073]
[0074] In this embodiment, the purpose of upsampling is to align the size of the second weighted feature S″4 with that of F3; at the same time, using the extracted weight W s3 The concatenated global feature F′3 is optimized to integrate the global feature information. The third weight extraction layer module is also composed of a conv1×1 layer and a sigmoid activation function.
[0075] S325. Input the third weighted feature F″3 into the convolution block module for feature integration and optimization to obtain a multi-scale fusion feature image.
[0076] In this embodiment, the convolutional block module is used to integrate and optimize F″3 to generate a rich global feature representation, that is, a multi-scale fused feature image. Therefore, the global driven semantic enhancement module significantly improves the ability of the improved Transformer structure to detect small lesions by focusing on enriching the global context features with high-level semantic features. Among them, the convolutional block module consists of a convolutional layer and a tanh activation function.
[0077] S33. Input the multi-scale fusion feature image into the decoder of the improved Transformer structure to classify and locate the area of interest of dental lesions, and obtain a trained improved Transformer structure for identifying dental lesions.
[0078] In this embodiment, the multi-scale fusion feature image is input into the decoder of the improved Transformer structure, and is processed through a multi-layer structure, namely, position encoding, decoder embedding and anchor box definition, to finally generate classification and positioning results; wherein the position encoding can help the decoder distinguish the position information of the input features, so as to better capture the relative relationship between the features; the anchor box is a predefined set of borders that cover the image at different scales and aspect ratios, and help the encoder predict the position and size of the area of interest of the dental lesion; the decoder embedding uses the learnable position encoding as the input embedding, and is processed through multiple layers of structures, and finally converted into the output classification result and bounding box offset; finally, the predefined anchor box is updated using the learned bounding box offset to obtain the final positioning result, that is, the trained improved Transformer structure is obtained for identifying dental lesions.
[0079] Specifically, step S33 specifically includes S331-S335:
[0080] S331, position encoding the multi-scale fusion feature image to generate input features containing position information.
[0081] S332. The decoder embedding is input into the multi-head self-attention mechanism to capture the correlation within the decoder embedding and generate new decoder embedding features.
[0082] S333. The input features containing the position information are interacted with the new decoder embedding features through a multi-head cross attention mechanism to obtain the encoder information related to the current decoding position, and are sequentially input into the feedforward neural network, sum normalization, and multi-layer perceptron to perform multi-layer full connection operations, residual addition and normalization processing, and classification and positioning operations, respectively, to generate the confidence of the predicted classification and the bounding box offset of the positioning.
[0083] S334. Generate a new bounding box based on the predicted localized bounding box offset and the predefined anchor box, and obtain the localization result and classification result of the dental lesion region of interest.
[0084] S335. The positioning result of the dental lesion region of interest, the classification result and the true label are simultaneously input into the loss function, and the loss function is used to calculate the gradient to update the network parameters of the improved Transformer structure to obtain a trained improved Transformer structure.
[0085] In this embodiment, when training the improved Transformer structure, the Hungarian method is used to determine the best binary match between the predicted target and the true target; at the same time, the focal loss is used to supervise the classification loss between the prediction and the true target, and the L1 loss and the GIoU loss are used to supervise the positioning loss between the prediction and the true target. The above three losses constitute a loss function for gradient calculation and back propagation. The network parameters of the improved Transformer structure are updated using the loss function to obtain a trained improved Transformer structure.
[0086] S4. Obtain a new magnetic resonance imaging image of the tooth and input it into the trained improved Transformer structure to obtain the tooth lesion detection result.
[0087] In this embodiment, after the training of the improved Transformer structure is completed, the acquired magnetic resonance imaging image of the new tooth can be input into the trained improved Transformer structure to detect and identify the target of interest of the dental lesion.
[0088] In summary, the present invention proposes a method for detecting dental lesions based on global feature optimized Transformer, and proposes a Transformer structure of an encoder combined with a global driven semantic enhancement module, which has the following advantages: (1) Improving the detection accuracy of small-sized lesions; that is, it can effectively capture and integrate global features of different scales, thereby enhancing the detection ability of small-sized lesions; at the same time, the encoder can establish long-distance dependencies between images and capture global contextual information, so that small-sized lesions can be accurately identified and detected in complex backgrounds; (2) Enriching feature representation; that is, the global driven semantic enhancement module further enriches the feature representation by integrating high-level semantic information; so that the improved Transformer structure, by combining global contextual features with high-level semantic features, can better understand and distinguish different types of lesions; this multi-scale feature enhancement method not only improves the detection accuracy, but also significantly enhances The ability to detect small targets; (3) reduce the impact of computational complexity; that is, although the Transformer structure performs well in global feature extraction, its computational complexity increases with the square of the input data size and is limited by hardware conditions; the improved Transformer structure proposed in the present invention effectively alleviates the problem of the complexity of the traditional Transformer structure by combining the encoder with the global driven semantic enhancement module, so that the model can still achieve efficient dental lesion detection under the condition of limited hardware resources; (4) improve diagnostic efficiency and accuracy; the combination of the encoder and the global driven semantic enhancement module improves feature extraction and enhancement, significantly improves the overall performance of the improved Transformer structure, and makes the detection and segmentation of dental lesions more accurate; it can not only improve diagnostic efficiency, but also provide clinicians with more accurate diagnostic information to assist them in early diagnosis and treatment of diseases. Therefore, the present invention uses the improved Transformer structure to detect dental lesions, significantly improving the accuracy and efficiency of dental lesion detection, overcoming the shortcomings of existing methods in small-size lesion detection, and providing an innovative and effective solution for the field of medical image analysis.
[0089] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
[0090] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.
Claims
1. A method for detecting dental lesions based on global feature optimized Transformer, characterized in that: The following steps are involved: S1, obtain the magnetic resonance imaging image of the teeth and input it into the backbone network to extract multi-scale features; S2, introduce the global driven semantic enhancement module and build an improved Transformer structure; The improved Transformer structure includes an encoder, a global driven semantic enhancement module, and a decoder connected in sequence; A global-driven semantic enhancement module to capture global features at different scales and enhance the detection of small-sized dental lesions; S3, inputting the multi-scale features into the improved Transformer structure for training, and obtaining a trained improved Transformer structure for identifying dental lesions; S4. Obtain a new magnetic resonance imaging image of the tooth and input it into the trained improved Transformer structure to obtain the tooth lesion detection result.
2. The method for detecting dental lesions based on global feature optimized Transformer according to claim 1, characterized in that: The multi-scale features in step S1 are: {S3, S4, S5} Among them, S3 represents the third scale feature, S4 represents the fourth scale feature, and S5 represents the fifth scale feature.
3. The method for detecting dental lesions based on global feature optimized Transformer according to claim 2, characterized in that: The global driven semantic enhancement module includes a first convolution layer module, a second convolution layer module, a first weight extraction layer module, a second weight extraction layer module, a third weight extraction layer module, a first upsampling module, a second upsampling module, and a convolution block module.
4. The method for detecting dental lesions based on global feature optimized Transformer according to claim 3, characterized in that: Step S3 specifically includes: S31, after position encoding, the S3 scale feature in the multi-scale feature is added to the S3 scale feature, and then input into the encoder of the improved Transformer structure for encoding operation to obtain the global feature F3; S32, inputting the global feature F3, the S4 scale feature in the multi-scale feature, and the S5 scale feature into the global driven semantic enhancement module for multi-scale feature fusion to obtain a multi-scale fusion feature image; S33. Input the multi-scale fusion feature image into the decoder of the improved Transformer structure to classify and locate the area of interest of dental lesions, and obtain a trained improved Transformer structure for identifying dental lesions.
5. The method for detecting dental lesions based on global feature optimized Transformer according to claim 4, characterized in that: Step S32 specifically includes: S321, input the global feature F3 into the first convolutional layer module for size alignment operation, obtain the global feature F3 of the same size as the scale feature S5, and input it into the first weight extraction layer module for convolution activation operation to obtain the weight W s5 , and the weight W s5 After element-by-element multiplication with the S5 scale feature, element-by-element addition with the S5 scale feature is performed to obtain the first weighted feature S′5, that is: Among them, conv represents the convolution operation, ↓ represents the downward resizing operation, and sigmoid represents the activation function; S322, after the first weighted feature S′5 is input into the first upsampling module for upsampling operation, it is concatenated with the scale feature S4 to obtain the concatenated feature S′4, that is: S′4=C(↑S′5,S4) Among them, C represents concatenation, ↑ represents upsampling; S323, input the global feature F3 into the second convolutional layer module for size alignment operation, obtain the global feature F3 of the same size as the scale feature S4, and input it into the second weight extraction layer module for convolution activation operation to obtain the weight W s4 , and the weight W s4 After element-by-element multiplication with the concatenated feature S′4, element-by-element addition with the concatenated feature S′4 is performed to obtain the second weighted feature S″4, namely: S324, input the second weighted feature S″4 into the second upsampling module for upsampling operation, and then concatenate it with the global feature F3 to obtain the concatenated global feature F′3, and at the same time input the global feature F3 into the third weight extraction layer module for convolution activation operation to obtain the weight W s3 , the weight W s3 After element-by-element multiplication with the concatenated global feature F′3, element-by-element addition with the concatenated global feature F′3 is performed to obtain the third weighted feature F″3, namely: S325. Input the third weighted feature F″3 into the convolution block module for feature integration and optimization to obtain a multi-scale fusion feature image.
6. The method for detecting dental lesions based on global feature optimized Transformer according to claim 5, characterized in that: Step S33 specifically includes: S331, position encoding the multi-scale fusion feature image to generate input features containing position information; S332, embedding the decoder into a multi-head self-attention mechanism to capture the correlation within the decoder embedding and generate new decoder embedding features; S333, interacting the input features containing the position information with the new decoder embedding features through a multi-head cross attention mechanism, obtaining the encoder information related to the current decoding position, and sequentially inputting the feedforward neural network, sum normalization, and multi-layer perceptron to perform multi-layer full connection operations, residual addition and normalization processing, and classification and positioning operations, respectively, to generate the predicted classification confidence and the positioning bounding box offset; S334, generating a new bounding box according to the predicted localized bounding box offset and the predefined anchor box, and obtaining the localization result and classification result of the dental lesion region of interest; S335. The positioning result of the dental lesion region of interest, the classification result and the true label are simultaneously input into the loss function, and the loss function is used to calculate the gradient to update the network parameters of the improved Transformer structure to obtain a trained improved Transformer structure.
Citation Information
Patent Citations
Deep learning-based root tip X-ray film disease identification method and system
CN117132835A
Particle detection sensor and particle detection device
KR1020240125531A
Cited By
Oral periodontal image analysis method and system based on multi-scale feature fusion
CN120953746A
A Multi-Scale Feature Fusion Method and System for Oral and Periodontal Image Analysis
CN120953746B
