Oral picture decayed tooth detection technology based on RetinaNet
By using a deep learning method based on RetinaNet to detect dental caries by taking oral images with a mobile phone, the problem of dental caries detection in areas with insufficient medical resources has been solved, achieving efficient and accurate dental caries detection and alleviating the strain on medical resources.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2024-10-30
- Publication Date
- 2026-05-01
AI Technical Summary
Due to the imbalance of medical resources and the contradiction between supply and demand, many regions are unable to meet people's demand for oral medical services, especially the demand for caries detection. Existing technologies are unable to achieve efficient and accurate caries detection.
A deep learning method based on RetinaNet is used to detect dental caries by taking oral images with a mobile phone. The ResNet50 algorithm is used for feature extraction, and the model is optimized by combining a feature pyramid network and a focus loss function to generate a dental caries detection model. Non-maximum suppression is used to filter bounding boxes to improve detection accuracy.
It enables efficient and accurate caries detection in areas with limited medical resources, avoids the use of high-cost equipment, improves detection speed and accuracy, and alleviates the problem of strained medical resources.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing, specifically relating to a method for detecting dental caries in oral images based on RetinaNet. Background Technology
[0002] Tooth decay is a chronic oral disease affecting most people worldwide. Oral health is an indispensable part of overall health, and the prevention and treatment of oral diseases heavily rely on the diagnosis of dental professionals. However, due to the imbalance of medical resources and the mismatch between supply and demand, many regions are unable to meet the demand for oral healthcare services. With the rapid development of computer technology, applying deep learning technology to the field of oral healthcare has become a trend, which is expected to improve the quality and efficiency of oral healthcare services.
[0003] Deep learning is a powerful artificial intelligence technology that can train models using massive amounts of data to identify and understand patterns in the real world. This project creates models based on deep learning and image processing techniques to analyze tooth shape and color to determine the presence of cavities. These clinical diagnoses can then be used to predict the progression of the disease and promptly alert patients to treatment. Furthermore, this project can facilitate large-scale, automated processing and analysis of oral images and medical records, contributing to oral health screenings and providing a new technological means to improve oral health and the availability of oral healthcare services. Summary of the Invention
[0004] To address the aforementioned issues of imbalanced medical resources, supply-demand imbalances, and the inability of many regions to meet people's demand for oral healthcare services, this invention proposes a RetinaNet-based oral image caries detection technology. This technology allows users to take oral images with a mobile phone and check for caries anytime, anywhere.
[0005] The present invention adopts the following technical solution:
[0006] S1: Obtain a dataset of patients with dental caries, preprocess and expand the dataset, and then divide the dataset into training and testing sets.
[0007] S2: The model loads the image and annotation information, and uses the ResNet50 algorithm for feature extraction.
[0008] S3: Use Feature Pyramid Network (FPN) to build feature pyramids at multiple scales to adapt to the detection of targets of different sizes.
[0009] S4: Apply a classification subnetwork and a bounding box regression subnetwork at each scale of the feature pyramid network to predict the target category and its location;
[0010] S5: The classification task is optimized using the Focal Loss function to reduce the impact of the imbalance between positive and negative samples. The final trained model is obtained by training the model using the training set.
[0011] S6: Input the test set into the trained model to evaluate its performance and continuously optimize the parameters to improve the model.
[0012] Furthermore, the specific content of step S1 is as follows:
[0013] S7: Flip existing images horizontally or vertically, rotate them by a certain angle, and increase their exposure to generate new data that is similar to but not exactly the same as the original image. This increases the richness of the data, solves the CNN overfitting problem caused by data shortage, improves the model's generalization ability, and enhances the model's robustness and reliability.
[0014] S8: Use annotation tools to label images of cavities, targeting teeth with changes in color, defects in shape, or cavities.
[0015] Furthermore, the specific content of step S3 is as follows:
[0016] S9: The feature pyramid network includes bottom-up convolution and top-down feature fusion operations, which makes high-level feature maps have stronger semantic information and low-level feature maps have richer spatial information, thereby improving the performance of multi-scale object detection.
[0017] Furthermore, the specific content of step S4 is as follows:
[0018] S10: The classification subnetwork and the bounding box regression subnetwork are both fully convolutional network structures, and the feature maps at each scale share the same classification subnetwork and bounding box regression subnetwork, which is used to reduce model complexity and improve detection efficiency.
[0019] Furthermore, the specific content of step S5 is as follows:
[0020] S11: RetinaNet generates a large number of anchor boxes, most of which are background (i.e., negative samples), while there are relatively few anchor boxes (positive samples) in the target region.
[0021] S12: When calculating the classification loss, Focal Loss assigns lower weights to these numerous negative samples, reducing their contribution to the loss and preventing the model from overemphasizing the background region. Specifically, this manifests as follows:
[0022]
[0023] Where L flFor Focal-Loss, γ>0 is an adjustable factor, and y=1 indicates that when the predicted value is equal to the label value, we can let:
[0024]
[0025] Then Focal-Loss can be uniformly expressed as:
[0026] L fl =-α t (1-p t ) γ log(p t )
[0027] p t This reflects the proximity of the predicted bounding box to the ground truth bounding box, i.e., category y. t The larger the value of α, the closer it is to category y, indicating a more accurate classification. t It is a balancing factor used to balance the weights of positive and negative samples.
[0028] S13: Improve detection accuracy by training the model to learn more efficiently on small target samples.
[0029] Furthermore, the specific content of step S6 is as follows:
[0030] S14: Use the non-maximum suppression (NMS) method to filter the predicted bounding boxes, eliminate overlapping boxes, and retain only the bounding boxes with the highest confidence to obtain the final detection results.
[0031] S15: Performance evaluation metrics are measured using AP, F1 index, and mAP (mean Average Precision).
[0032] Average Precision (AP) is a commonly used evaluation metric in object detection. It is used to assess the detection performance of a system, specifically its ability to correctly detect targets on a test set of samples at different thresholds. Its formula is as follows:
[0033]
[0034] The F1 index, used to balance precision and recall, is defined as the harmonic mean of precision and recall. Its formula is as follows:
[0035]
[0036] The F1 index ranges from 0 to 1, with the closer to 1 indicating better model performance.
[0037] mAP measures a model's performance across all classes. mAP is calculated by averaging the AP values of all classes. The formula is as follows:
[0038]
[0039] The beneficial effects of this invention are:
[0040] This patent proposes a caries detection technology that uses a mobile phone to capture images of the oral cavity. This method avoids the high cost and equipment requirements of using X-rays and is particularly suitable for areas with limited medical resources. By automatically generating bounding boxes to identify and locate cavities, it not only improves detection speed and accuracy but also helps alleviate the problem of strained medical resources, making dental health management more efficient and widespread. Attached Figure Description
[0041] Figure 1 Roadmap and framework for dental caries detection technology
[0042] Figure 2 RetinaNet architecture diagram
[0043] Figure 3 ResNet50 algorithm architecture diagram Detailed Implementation
[0044] The present invention will now be described in further detail with reference to the accompanying drawings.
[0045] Specific implementation methods are as follows: Figures 1 to 3 As shown:
[0046] according to Figure 1 The roadmap for dental caries detection technology shows that its key features include the following steps:
[0047] S1: Obtain a dataset of patients with dental caries, preprocess and expand the dataset, and then divide the dataset into training and testing sets.
[0048] S7: Flip existing images horizontally or vertically, rotate them by a certain angle, and increase their exposure to generate new data that is similar to but not exactly the same as the original image. This increases the richness of the data, solves the CNN overfitting problem caused by data shortage, improves the model's generalization ability, and enhances the model's robustness and reliability.
[0049] S8: Use annotation tools to label images of cavities, targeting teeth with changes in color, defects in shape, or cavities.
[0050] S2: The model loads the image and annotation information, and uses the ResNet50 algorithm for feature extraction.
[0051] S3: Use Feature Pyramid Network (FPN) to build feature pyramids at multiple scales to adapt to the detection of targets of different sizes.
[0052] S9: The feature pyramid network includes bottom-up convolution and top-down feature fusion operations, which makes high-level feature maps have stronger semantic information and low-level feature maps have richer spatial information, thereby improving the performance of multi-scale object detection.
[0053] S4: Apply a classification subnetwork and a bounding box regression subnetwork at each scale of the feature pyramid network to predict the target category and its location.
[0054] S10: The classification subnetwork and the bounding box regression subnetwork are both fully convolutional network structures, and the feature maps at each scale share the same classification subnetwork and bounding box regression subnetwork, which is used to reduce model complexity and improve detection efficiency.
[0055] S5: The classification task is optimized using the Focal Loss function to reduce the impact of the imbalance between positive and negative samples. The final trained model is obtained by training the model using the training set.
[0056] S11: RetinaNet generates a large number of anchor boxes, most of which are background (i.e., negative samples), while there are relatively few anchor boxes (positive samples) in the target region.
[0057] S12: When calculating the classification loss, Focal Loss assigns lower weights to these numerous negative samples, reducing their contribution to the loss and preventing the model from overemphasizing the background region. Specifically, this manifests as follows:
[0058]
[0059] Where L fl For Focal-Loss, γ>0 is an adjustable factor, and y=1 indicates that when the predicted value is equal to the label value, we can let:
[0060]
[0061] Then Focal-Loss can be uniformly expressed as:
[0062] L fl =-α t (1-p t ) γ log(p t )
[0063] p tThis reflects the proximity of the predicted bounding box to the ground truth bounding box, i.e., category y. t The larger the value of α, the closer it is to category y, indicating a more accurate classification. t It is a balancing factor used to balance the weights of positive and negative samples.
[0064] S13: Improve detection accuracy by training the model to learn more efficiently on small target samples.
[0065] S6: Input the test set into the trained model to evaluate its performance and continuously optimize the parameters to improve the model.
[0066] S14: Use the non-maximum suppression (NMS) method to filter the predicted bounding boxes, eliminate overlapping boxes, and retain only the bounding boxes with the highest confidence to obtain the final detection results.
[0067] S15: Performance evaluation metrics are measured using AP, F1 index, and mAP (mean Average Precision).
[0068] Average Precision (AP) is a commonly used evaluation metric in object detection. It is used to assess the detection performance of a system, specifically its ability to correctly detect targets on a test set of samples at different thresholds. Its formula is as follows:
[0069]
[0070] The F1 index, used to balance precision and recall, is defined as the harmonic mean of precision and recall. Its formula is as follows:
[0071]
[0072] The F1 index ranges from 0 to 1, with the closer to 1 indicating better model performance.
[0073] mAP measures a model's performance across all classes. mAP is calculated by averaging the AP values of all classes. The formula is as follows:
[0074]
Claims
1. A deep learning-based object detection method, using the RetinaNet model architecture for multi-object detection, characterized in that... Includes the following steps: S1: Obtain a dataset of patients with dental caries, preprocess and expand the dataset, and then divide the dataset into training and testing sets. S2: The model loads the image and annotation information, and uses the ResNet50 algorithm for feature extraction. S3: Use Feature Pyramid Network (FPN) to build feature pyramids at multiple scales to adapt to the detection of targets of different sizes. S4: Apply a classification subnetwork and a bounding box regression subnetwork at each scale of the feature pyramid network to predict the target category and its location; S5: The classification task is optimized using the Focal Loss function to reduce the impact of the imbalance between positive and negative samples. The final trained model is obtained by training the model using the training set. S6: Input the test set into the trained model to evaluate its performance and continuously optimize the parameters to improve the model.
2. The target detection method according to claim 1, wherein step S1 comprises: S7: Flip existing images horizontally or vertically, rotate them by a certain angle, and increase their exposure to generate new data that is similar to but not exactly the same as the original image. This increases the richness of the data, solves the CNN overfitting problem caused by data shortage, improves the model's generalization ability, and enhances the model's robustness and reliability. S8: Use annotation tools to label images of cavities, targeting teeth with changes in color, defects in shape, or cavities.
3. The target detection method according to claim 1, wherein step S3 comprises: S9: The feature pyramid network includes bottom-up convolution and top-down feature fusion operations, which makes high-level feature maps have stronger semantic information and low-level feature maps have richer spatial information, thereby improving the performance of multi-scale object detection.
4. The target detection method according to claim 1, wherein step S4 comprises: S10: The classification subnetwork and the bounding box regression subnetwork are both fully convolutional network structures, and the feature maps at each scale share the same classification subnetwork and bounding box regression subnetwork, which is used to reduce model complexity and improve detection efficiency.
5. The target detection method according to claim 1, wherein step S5 comprises: S11: RetinaNet generates a large number of anchor boxes, most of which are background (i.e., negative samples), while there are relatively few anchor boxes (positive samples) in the target region. S12: When calculating the classification loss, Focal Loss assigns lower weights to these numerous negative samples, reducing their contribution to the loss and preventing the model from overemphasizing the background region. Specifically, this manifests as follows: Where L fl For Focal-Loss, γ>0 is an adjustable factor, and y=1 indicates that when the predicted value is equal to the label value, we can let: Then Focal-Loss can be uniformly expressed as: L fl =-a t (1-p t ) γ log(p t ) p t This reflects the proximity of the predicted bounding box to the ground truth bounding box, i.e., category y. t The larger the value of α, the closer it is to category y, indicating a more accurate classification. t It is a balancing factor used to balance the weights of positive and negative samples. S13: Improve detection accuracy by training the model to learn more efficiently on small target samples.
6. The target detection method according to claim 1, wherein step S6 comprises: S14: Use the non-maximum suppression (NMS) method to filter the predicted bounding boxes, eliminate overlapping boxes, and retain only the bounding boxes with the highest confidence to obtain the final detection results. S15: Performance evaluation metrics are measured using AP, F1 index, and mAP (mean Average Precision). Average Precision (AP) is a commonly used evaluation metric in object detection. It is used to assess the detection performance of a system, specifically its ability to correctly detect targets on a test set of samples at different thresholds. Its formula is as follows: The F1 index, used to balance precision and recall, is defined as the harmonic mean of precision and recall. Its formula is as follows: The F1 index ranges from 0 to 1, with the closer to 1 indicating better model performance. mAP measures a model's performance across all classes. mAP is calculated by averaging the AP values of all classes. The formula is as follows: