A method for caries detection in dental panoramic radiographs

Abstract

The application discloses a caries recognition method for an oral panoramic film, which comprises the following steps: preprocessing the oral panoramic film, extracting a preselected region and marking the same as a sample set; constructing a U-Net segmentation model and training the same by using the sample set; and testing the segmentation model. In combination with the characteristics of the oral panoramic film image, the application can achieve accurate recognition of a pixel-level caries lesion region on the oral panoramic film. The image pre-processing effectively eliminates some unnecessary information on the oral panoramic film, thereby effectively improving the accuracy of the caries segmentation model. The image enhancement strategy during the training of the segmentation model is combined with the characteristics of the oral panoramic film, thereby improving the robustness of the segmentation model. The image post-processing on the prediction result can effectively eliminate some false positive regions.

Description

Technical Field

[0001] This invention is applied to the field of artificial intelligence medical image processing, specifically a method for identifying dental caries from panoramic oral radiographs. Background Technology

[0002] Dental caries is one of the most common infectious chronic dental diseases in humans. It is a dynamic disease process caused by the metabolic activity of the tooth biofilm. However, dental caries is a preventable disease; if detected in its early stages, it can be stopped and potentially reversed. X-rays are one of the methods used by dentists to assess oral health and diagnose dental diseases (such as dental caries). They are also the most common imaging method in dental clinical practice, helping to identify dental problems that are difficult to detect through visual examination alone. There are several types of dental X-rays, each recording different anatomical views of the teeth, such as apical, panoramic, and periapical radiographs. Among these, panoramic radiographs are the most common due to their low cost and low radiation exposure. Panoramic radiographs cover the entire dentition of the patient, as well as the surrounding bone and jaw structure. However, compared to the other two types of dental X-rays, panoramic radiographs have higher image noise, lower resolution, and cannot provide a detailed view of each tooth, making the identification of dental caries based on panoramic radiographs more difficult.

[0003] Currently, most computer-aided diagnostic (CAD) systems for dental caries identification developed based on artificial intelligence, especially deep learning technology, are only applicable to periodontal radiographs and occlusal radiographs. There is still no high-performance method for automatically identifying dental caries and classifying the degree of decay on panoramic radiographs. Existing methods for automatically identifying dental caries on panoramic radiographs mostly employ methods similar to those used on the other two types of X-rays: training a semantic segmentation model or a detection model to directly segment / detect the lesion area of ​​the caries. However, panoramic radiographs also contain information about other oral tissues besides teeth, and the lesion area of ​​the caries is only a small percentage of the entire panoramic radiograph. This makes direct segmentation of the caries more difficult, resulting in poor algorithm performance. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for identifying dental caries using panoramic radiographs, addressing the shortcomings of existing technologies.

[0005] To solve the above-mentioned technical problems, the present invention provides a method for identifying dental caries using panoramic radiographs, comprising the following steps:

[0006] Preprocess the panoramic oral radiographs, extract and label the pre-selected regions, and use them as a sample set;

[0007] Construct a U-Net segmentation model and train it using a sample set;

[0008] The segmentation model was tested.

[0009] As one possible implementation, the step of constructing the U-Net segmentation model and training it using a sample set specifically includes:

[0010] S21. Perform convolution and pooling operations on the extracted sample set images. After n pooling operations, a total of n feature maps of different sizes are obtained.

[0011] S22. Upsample the feature map and concatenate it with the feature map of the same size from the above n feature maps of different sizes on the channel side. Then convolve and upsample the concatenated feature map. After n upsampling operations, a prediction result map with the same size as the input sample set image can be obtained.

[0012] As one possible implementation, further, step S21 specifically involves: the encoder part performing convolution and pooling operations on the extracted 224×224 sample set image, and obtaining four feature maps of different sizes of 112×112, 56×56, 28×28 and 14×14 after four pooling operations.

[0013] As one possible implementation, further, step S22 specifically involves: the decoder part performing deconvolution on the 14×14 feature map obtained in step S21 to obtain a 28×28 feature map, and then concatenating it with the 28×28 feature map obtained in step S21 on the channel side. Then, convolution and deconvolution are performed on the concatenated feature map, and after four deconvolutions, a prediction result map of 224×224 with the same size as the input sample set image is obtained.

[0014] As a possible implementation, the step of constructing the U-Net segmentation model and training it using a sample set further includes:

[0015] S23. Design the loss function for training the U-Net segmentation model;

[0016] Specifically, the loss function combines the binary cross-entropy (BCE) function with the Dice loss function:

[0017] L seg =L BCE +L Dice

[0018] Among them, L BCE The calculation method is as follows:

[0019]

[0020] Where f represents the number of pixels, and m j With n jThese represent the predicted value and its corresponding actual value (Ground Truth), respectively.

[0021] Among them, L Dice The calculation method is as follows:

[0022]

[0023] Where, m j With n j These represent the predicted value and its corresponding actual value, respectively.

[0024] As a possible implementation, the step of constructing the U-Net segmentation model and training it using a sample set further includes:

[0025] Image enhancement operations can be used to improve the robustness of the model, including random image rotation, contrast adjustment, random image scaling, random flipping, and random offsetting.

[0026] As one possible implementation, the random zoom-in or zoom-out image operation is further configured with a zoom range of 10% and a zoom range of 50%.

[0027] As one possible implementation, the step of preprocessing the panoramic oral radiographs, extracting and labeling pre-selected regions, and using them as a sample set includes:

[0028] For preprocessing, the Sobel operator is used, with the order of the x-direction derivative set to 0 and the order of the y-direction derivative set to 1, and the kernel size set to 3×3 to perform vertical edge detection on the panoramic image, highlighting the vertical edges on the image, and a bilateral filter is used for edge sharpening.

[0029] As one possible implementation, the step of preprocessing the panoramic oral radiographs, extracting and labeling pre-selected regions, and using them as a sample set further includes:

[0030] The pre-selected region is extracted, and the image is subjected to horizontal integral intensity projection. The first obvious positive slope on the projection map represents the edge of the left angle of the mandible, and the initial pre-selected region is cropped out. The cropped image is binarized using the Otsu method, and a Gaussian filter is used to reduce noise in the binarized image. Then, the processed binary image is subjected to horizontal integral intensity projection to identify the outer oblique lines of the maxilla and mandible in the image, and the final pre-selected region is extracted.

[0031] As one possible implementation, the step of testing the segmentation model further includes:

[0032] S31. Extract the pre-selected region from the panoramic oral radiograph;

[0033] S32. Input the pre-selected region into the trained segmentation model to obtain a prediction map of the same size; set the confidence threshold to obtain a binary map of the prediction result.

[0034] S33. Perform image post-processing on the prediction results; calculate the maximum connected component on the binary image, remove some connected components, and obtain the prediction results.

[0035] The present invention adopts the above technical solution and has the following beneficial effects:

[0036] 1. Combining the characteristics of panoramic oral radiographs, this invention can achieve pixel-level accurate identification of caries lesion areas on panoramic oral radiographs.

[0037] 2. Image preprocessing effectively removes some unnecessary information from panoramic oral radiographs, thus improving the accuracy of the dental caries segmentation model.

[0038] 3. The image enhancement strategy used when training the segmentation model incorporates the characteristics of panoramic dental radiographs to improve the robustness of the segmentation model.

[0039] 4. Post-processing of the predicted results can effectively remove some false positive areas. Attached Figure Description

[0040] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0041] Figure 1 This is a schematic diagram of the U-Net network structure of the present invention;

[0042] Figure 2 This is a panoramic oral radiograph of the present invention and its corresponding prediction result diagram. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0044] Example 1

[0045] This invention provides a method for identifying dental caries using panoramic radiographs, comprising the following steps:

[0046] The steps of preprocessing panoramic dental images, extracting and labeling pre-selected regions, and using them as a sample set include:

[0047] For preprocessing, the Sobel operator is used, with the order of the x-direction derivative set to 0 and the order of the y-direction derivative set to 1, and the kernel size set to 3×3 to perform vertical edge detection on the panoramic image, highlighting the vertical edges on the image, and a bilateral filter is used for edge sharpening.

[0048] The pre-selected region is extracted, and the image is subjected to horizontal integral intensity projection. The first obvious positive slope on the projection map represents the edge of the left angle of the mandible, and the initial pre-selected region is cropped out. The cropped image is binarized using the Otsu method, and a Gaussian filter is used to reduce noise in the binarized image. Then, the processed binary image is subjected to horizontal integral intensity projection to identify the outer oblique lines of the maxilla and mandible in the image, and the final pre-selected region is extracted.

[0049] Construct a U-Net segmentation model and train it using a sample set;

[0050] The segmentation model was tested.

[0051] As one possible implementation, the step of constructing the U-Net segmentation model and training it using a sample set specifically includes:

[0052] S21. Perform convolution and pooling operations on the extracted sample set images. After n pooling operations, a total of n feature maps of different sizes are obtained.

[0053] S22. Upsample the feature map and concatenate it with the feature map of the same size from the above n feature maps of different sizes on the channel side. Then convolve and upsample the concatenated feature map. After n upsampling operations, a prediction result map with the same size as the input sample set image can be obtained.

[0054] Specifically, step S21 involves the encoder performing convolution and pooling operations on the extracted 224×224 sample set image. After four pooling operations, four feature maps of different sizes, namely 112×112, 56×56, 28×28 and 14×14, are obtained in sequence.

[0055] Step S22 specifically involves the decoder section performing deconvolution on the 14×14 feature map obtained in step S21 to obtain a 28×28 feature map. This 28×28 feature map is then concatenated with the 28×28 feature map obtained in step S21 along the same channel. Finally, convolution and deconvolution are performed on the concatenated feature map. After four deconvolutions, a 224×224 prediction result map with the same size as the input sample set image is obtained.

[0056] The steps also include:

[0057] S23. Design the loss function for training the U-Net segmentation model;

[0058] Specifically, the loss function combines the binary cross-entropy (BCE) function with the Dice loss function:

[0059] L seg =L BCE +L Dice

[0060] Among them, L BCE The calculation method is as follows:

[0061]

[0062] Where f represents the number of pixels, and m j With n j These represent the predicted value and its corresponding actual value (Ground Truth), respectively.

[0063] Among them, L Dice The calculation method is as follows:

[0064]

[0065] Where, m j With n j These represent the predicted value and its corresponding actual value, respectively.

[0066] Also includes:

[0067] The robustness of the model is improved using image enhancement operations, including random image rotation, contrast adjustment, random image magnification or reduction, random flipping, and random offsetting. The random image magnification or reduction operation is set with a magnification range of 10% and a reduction range of 50%.

[0068] The specific steps for testing the segmentation model include:

[0069] S31. Extract the pre-selected region from the panoramic oral radiograph;

[0070] S32. Input the pre-selected region into the trained segmentation model to obtain a prediction map of the same size; set the confidence threshold to obtain a binary map of the prediction result.

[0071] S33. Perform image post-processing on the prediction results; calculate the maximum connected component on the binary image, remove some connected components, and obtain the prediction results.

[0072] Example 2

[0073] A method for identifying dental caries using panoramic radiographs involves extracting the region of interest (ROI) from the panoramic radiograph using traditional image processing methods and filtering out unwanted information. Then, a segmentation model called U-Net is used to identify caries lesions within the ROI. This method effectively improves the accuracy of caries identification.

[0074] The specific steps are as follows:

[0075] Step 1: Perform image preprocessing on the panoramic dental radiograph to extract the region of interest (RoI).

[0076] Step 1.1: Since it is necessary to extract the upper and lower jaw regions containing all teeth as RoI from the original panoramic image, some image preprocessing operations need to be performed on the panoramic image first. The purpose is to improve the image quality and highlight the useful details on the panoramic image.

[0077] First, the Sobel operator is used, with the x-direction derivative order set to 0 and the y-direction derivative order set to 1, and the kernel size set to 3×3, to perform vertical edge detection on the panoramic image, highlighting the vertical edges on the image. Then, a bilateral filter is used to sharpen the edges.

[0078] Next, horizontal integral intensity projection is performed on the image. The first obvious positive slope on the projection map represents the edge of the left angle of the mandible, and the initial RoI is cropped out.

[0079] Step 1.2: The cropped image is binarized using the Otsu algorithm. Then, a Gaussian filter is used to reduce noise in the binarized image. Finally, a horizontal integral intensity projection is performed on the processed binary image to identify the outer slope lines of the maxilla and mandible (the starting point of the gap between the maxilla and mandible), which is the first obvious negative slope appearing in the projection image. The final RoI is then extracted.

[0080] Step 2: Train a segmentation model for the identification and segmentation of dental caries.

[0081] Step 2.1: Use the RoI extracted in the previous step and its corresponding caries lesion region annotations to train the caries lesion region segmentation network. In order to improve the robustness of the model, image enhancement strategies were also used during training: such as random image rotation, contrast adjustment, random image enlargement or reduction, random flipping, and random offset (vertical or horizontal).

[0082] Reducing the image size helps the model see different proportions of caries lesions, while increasing the size might cause missed cavities. Therefore, a 10% increase in magnification and a 50% reduction in size were set. For positive samples with caries annotations, the same rule applies to random offsets to prevent data augmentation strategies from causing omissions of image regions containing caries. Since caries mainly occur at the edges of teeth, the offset must be set to a small value, such as 20%.

[0083] Step 2.2: U-Net is used as the model for dental caries segmentation. It is the most widely adopted network framework for medical image segmentation because its unique encoder and decoder design can extract effective information while preserving the original features of the image. In the encoder part, convolution and pooling operations are performed on the extracted 224×224 RoI image. After four pooling operations, four feature maps of different sizes are obtained: 112×112, 56×56, 28×28, and 14×14. Then, in the decoder part, the 14×14 feature map is upsampled to obtain a 28×28 feature map. This 28×28 feature map is then concatenated with the 28×28 feature map from the encoder part on a channel-wise basis. Then, convolution and upsampling are performed on the concatenated feature map. After four upsampling operations, a prediction result of 224×224 with the same size as the input image is obtained.

[0084] Step 2.3: Design the loss function for training the caries segmentation network. Since the proportion of caries lesions on panoramic images is unbalanced compared to other parts, the loss function combines the binary cross-entropy (CE) function with the Dice loss function:

[0085] L seg =L BCE +L Dice

[0086] Among them, L BCE The calculation method is as follows:

[0087]

[0088] f represents the number of pixels, while m j With n j These represent the predicted value and its corresponding actual value (Ground Truth), respectively.

[0089] Because the cross-entropy loss function is highly sensitive to class imbalance, the Dice loss is also used as a loss function during model training. The calculation method is as follows:

[0090]

[0091] m j With n j These represent the predicted value and its corresponding actual value (Ground Truth), respectively.

[0092] Step 3: Test the dental caries identification method based on deep learning and image processing panoramic radiographs.

[0093] Step 3.1: Extract a 224×224 area of ​​RoI from the panoramic radiograph using the method described in Step 1.

[0094] Step 3.2: Input the RoI into the trained caries segmentation model to obtain a prediction map of the same size. Set the confidence threshold T1 = 0.5 to obtain a binary map of the prediction result.

[0095] Step 3.3: Post-process the prediction results. Calculate the maximum connected component on the binary image and remove some connected components smaller than 10 pixels to effectively reduce false positives. After removal, the final prediction result is obtained.

[0096] The above description is an embodiment of the present invention. For those skilled in the art, any equivalent changes, modifications, substitutions and variations made in accordance with the teachings of the present invention without departing from the principles and spirit of the present invention should be covered by the present invention.

Claims

1. A method for identifying dental caries using panoramic dental radiographs, characterized in that, It includes the following steps: Preprocess the panoramic oral radiographs, extract and label the pre-selected regions, and use them as a sample set; Construct a U-Net segmentation model and train it using a sample set; Test the segmentation model; The step of preprocessing panoramic oral radiographs, extracting and labeling pre-selected regions, and using them as a sample set includes: Preprocessing uses the Sobel operator, setting the order of the x-direction derivative to 0, the order of the y-direction derivative to 1, and the kernel size to 3×3 to perform vertical edge detection on the panoramic image, highlighting the vertical edges on the image, and using a bilateral filter for edge sharpening. Also includes: The pre-selected region is extracted, and the image is subjected to horizontal integral intensity projection. The first obvious positive slope on the projection map represents the edge of the left angle of the mandible, and the initial pre-selected region is cropped out. The cropped image is binarized using the Otsu method, and a Gaussian filter is used to reduce noise in the binarized image. Then, the processed binary image is subjected to horizontal integral intensity projection to identify the external oblique lines of the maxilla and mandible in the image, and the final pre-selected region is extracted. The step of constructing the U-Net segmentation model and training it using a sample set further includes: S23. Design the loss function for training the U-Net segmentation model; Specifically, the loss function combines the binary cross-entropy (BCE) function with the Dice loss function: in, The calculation method is as follows: Where f represents the number of pixels, and and These represent the predicted value and its corresponding actual value (Ground Truth), respectively. in, The calculation method is as follows: in, and These represent the predicted value and its corresponding actual value, respectively.

2. The method for identifying dental caries using panoramic dental radiographs according to claim 1, characterized in that: The steps of constructing the U-Net segmentation model and training it using a sample set specifically include: S21. Perform convolution and pooling operations on the extracted sample set images. After n pooling operations, a total of n feature maps of different sizes are obtained. S22. Upsample the feature map and concatenate it with the feature map of the same size from the above n feature maps of different sizes on the channel side. Then convolve and upsample the concatenated feature map. After n upsampling operations, a prediction result map with the same size as the input sample set image can be obtained.

3. The method for identifying dental caries using panoramic dental radiographs according to claim 2, characterized in that: Specifically, step S21 involves the encoder performing convolution and pooling operations on the extracted 224×224 sample set image. After four pooling operations, four feature maps of different sizes, namely 112×112, 56×56, 28×28, and 14×14, are obtained in sequence.

4. The method for identifying dental caries using panoramic dental radiographs according to claim 3, characterized in that: Specifically, step S22 involves the following steps: In the decoder section, the 14×14 feature map obtained in step S21 is deconvolved to obtain a 28×28 feature map, which is then concatenated with the 28×28 feature map obtained in step S21 along the same channel. The concatenated feature map is then convolved and deconvolved, and after four deconvolutions, a 224×224 prediction result map with the same size as the input sample set image is obtained.

5. The method for identifying dental caries using panoramic dental radiographs according to claim 1, characterized in that: The step of constructing the U-Net segmentation model and training it using a sample set further includes: Image enhancement operations can be used to improve the robustness of the model, including random image rotation, contrast adjustment, random image scaling, random flipping, and random offsetting.

6. The method for identifying dental caries using panoramic dental radiographs according to claim 5, characterized in that: The random zoom-in or zoom-out image operation is set with a zoom range of 10% and a zoom range of 50%.

7. The method for identifying dental caries using panoramic dental radiographs according to claim 1, characterized in that: The specific steps for testing the segmentation model include: S31. Extract the pre-selected region from the panoramic oral radiograph; S32. Input the pre-selected region into the trained segmentation model to obtain a prediction map of the same size; set the confidence threshold to obtain a binary map of the prediction result. S33. Perform image post-processing on the prediction results; calculate the maximum connected component on the binary image, remove some connected components, and obtain the prediction results.

Images