A method for detecting dental disease regions based on image segmentation and reconstruction algorithms

Abstract

This invention provides a method for detecting dental lesion regions based on image segmentation and reconstruction algorithms, comprising: inputting an original image; sampling regions of the original image using a target detection algorithm to generate N detection boxes; filtering out detection boxes with low lesion probability using threshold suppression; obtaining the overlap of the detection boxes using non-maximum suppression and determining whether the overlap exceeds the suppression threshold; segmenting the remaining n detection boxes using a segmentation algorithm to obtain n Boolean matrices belonging to the lesion region; reconstructing the n Boolean matrices based on the region information to generate a discrete array; and analyzing the pixel position and lesion type at the position based on the values ​​of the discrete array. This invention reduces missegmentation and noise, avoids interference from irrelevant regions in the image, and improves the accuracy and overall performance of lesion region segmentation by predicting high-probability lesion regions in the image and refining their segmentation and reconstruction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a dental disease area detection method based on image segmentation and reconstruction algorithm. BACKGROUND

[0002] With the rapid development of artificial intelligence technology in the medical industry, oral diseases can also be analyzed through oral image data such as panoramic films, CBCT images and intraoral photographs, and AI algorithms can automatically identify oral problems in a short time, thereby assisting oral doctors in accurate diagnosis, greatly improving the efficiency of diagnosis and reducing errors caused by human factors.

[0003] The existing scheme for processing oral images is usually through a deep learning image segmentation algorithm and inputting the complete image into a convolutional neural network similar to the U-Net structure. The network extracts feature information through a series of image processing steps. Specifically, the input oral image is first extracted through a convolutional layer to extract low-level features, then the data distribution is balanced through a batch normalization layer, the network expression ability is enhanced through an activation function nonlinear transformation, and then a down-sampling layer is used to reduce the size of the feature map to extract higher-level features. In order to avoid overfitting in deep learning, a Dropout layer is inserted to reduce overfitting by randomly discarding neurons to improve the generalization ability of the model. Then, through further processing of these high-dimensional feature maps, they are sequentially enhanced through a convolutional layer, a batch normalization layer and an activation function, and then an up-sampling layer is used to restore the size of the feature map to be consistent with the original image, and finally a segmentation result with the same size as the input image is generated.

[0004] However, the above-mentioned scheme based on pixel-by-pixel equal processing of images has some obvious shortcomings. First, since the convolutional neural network in the above process treats each pixel of each image equally, it leads to misclassification in the prediction results, i.e. pixels not belonging to the target area are mistakenly classified as target objects. In addition, for noise points in the image, such as low-quality or blurred areas, they may also be mistakenly considered as part of the target area without distinction, which leads to a large number of noise points and errors in the final segmentation result. In summary, the robustness and accuracy of the existing technology still need to be improved when dealing with complex or variable scene image processing.

[0005] Therefore, there is an urgent need for a dental disease area detection method that can selectively process key information in images and ignore irrelevant areas when dealing with complex or variable scene image processing, thereby helping to reduce misclassification and noise and improve overall performance. SUMMARY

[0006] In order to solve the problems existing in the prior art, the present application aims to provide a dental disease area detection method based on image segmentation and reconstruction algorithm, which can selectively process key information in the image and ignore irrelevant areas during image processing by pre-evaluating and processing the disease area, thereby reducing missegmentation and noise, avoiding interference from irrelevant areas in the image, and improving the accuracy and overall performance of disease area segmentation.

[0007] The present application achieves the above-mentioned purpose by the following technical solutions:

[0008] A dental disease area detection method based on image segmentation and reconstruction algorithm, comprising:

[0009] S1: input an original image, sample regions of the original image by a target detection algorithm, and generate N detection boxes containing region information of all suspected disease areas in the original image.

[0010] S2: filter out the detection boxes with low disease probability by threshold suppression processing.

[0011] S3: set a suppression threshold, obtain the overlap degree of the detection boxes by non-maximum suppression processing, and determine whether the overlap degree of the detection boxes exceeds the suppression threshold, if yes, filter out the detection boxes, and if no, keep the detection boxes.

[0012] S4: segment the n detection boxes kept by the segmentation algorithm, and determine whether the pixels of the segmented regions belong to the lesion area, thereby obtaining n Boolean matrices belonging to the lesion area.

[0013] S5: reconstruct the n Boolean matrices according to the region information to generate a discrete array with the same size as the original image.

[0014] S6: analyze the pixel position represented by the value of the discrete array and the disease type of the position.

[0015] According to the dental disease area detection method based on image segmentation and reconstruction algorithm provided by the present application, image preprocessing is performed before inputting the original image, and the processing process includes:

[0016] S11: read the original image and perform pixel normalization and data enhancement processing on the original image.

[0017] S12: input the processed original image into a neural network and output after data format processing by the neural network.

[0018] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, wherein the target detection algorithm in step S1 adopts a region proposal network, including:

[0019] S13: generating a candidate box containing all region information of the original image through the region proposal network.

[0020] S14: performing feature extraction and classification on each candidate box.

[0021] S15: mapping the candidate box to a position and size close to a real detection box through a bounding box regression to predict the position and size of the detection box.

[0022] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, wherein the target detection algorithm in step S1 adopts a selective search, including:

[0023] S131: performing region segmentation on the original image.

[0024] S141: creating a bounding box for the segmented region.

[0025] S151: setting a search rule and merging similar regions according to the search rule to obtain the detection box.

[0026] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, wherein the threshold suppression processing method in step S2 includes:

[0027] S21: setting a minimum threshold according to the image content of the dental disease.

[0028] S22: comparing the pixels in the detection box with the minimum threshold, retaining the detection box greater than or equal to the minimum threshold, and filtering out the detection box less than the minimum threshold.

[0029] S23: repeating step S22 until all N detection boxes are processed.

[0030] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, wherein the non-maximum suppression processing method in step 3 includes:

[0031] S31: calculating the confidence of each detection box.

[0032] S32: scoring and sorting the detection boxes according to the value of the confidence.

[0033] S33: selecting the detection box with the highest score and calculating the overlap degree of the detection box with the remaining detection boxes.

[0034] S34: Determine whether the overlap is greater than the suppression threshold, if yes, keep the bounding box. If no, set the score of the bounding box to 0.

[0035] S35: Repeat steps S33-S34 until all bounding boxes are processed.

[0036] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, and the confidence calculation method in step S31 comprises:

[0037] S311: Obtain the lesion existence probability of the bounding box.

[0038] S312: Calculate the overlap of the bounding box and the real bounding box by the YOLO algorithm.

[0039] S313: Estimate the confidence of the bounding box according to the maximum overlap.

[0040] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, and after the threshold suppression and non-maximum suppression, the number of the bounding boxes is reduced to n, which is consistent with the number of the real target area.

[0041] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, and the segmentation algorithm comprises:

[0042] S41: Perform binary segmentation on the area of the bounding box.

[0043] S42: Set a lesion area threshold.

[0044] S43: Compare the pixels of the segmented area with the lesion area threshold, if greater than or equal to the lesion area threshold, it means that it belongs to the lesion area, then set the pixel to True. If less than, set it to False.

[0045] S44: The pixels with True value form a two-dimensional array.

[0046] S45: Repeat steps S43-S44 until all segmented areas are compared, and obtain n Boolean matrices.

[0047] According to the application, a dental disease area detection method based on image segmentation and reconstruction algorithm is provided, and step S6 further comprises:

[0048] S61: Define a set of color palettes corresponding to different lesion types.

[0049] S62: Map the discrete array to a superimposed layer of corresponding colors according to the color palette.

[0050] S63: Add the pixel values of the lesion area to the overlay layer in proportion, so as to analyze the area where the lesion is located through the color displayed.

[0051] Therefore, compared with the prior art, the present application has the following beneficial effects:

[0052] 1. The present application evaluates the position of the lesion area through pre-processing of the image and specific algorithms or strategies. Compared with the traditional image processing scheme, the present application avoids the interference of irrelevant areas in the whole image through the way of prior screening, thereby improving the accuracy of lesion area segmentation.

[0053] 2. The present application positions the lesion area and divides it into multiple sub-regions for more detailed segmentation processing through efficient segmentation of local areas. Compared with the traditional image processing scheme, the segmentation strategy of the present application enables the model to focus on the details of the key area, thereby reducing the influence of missegmentation and noise points.

[0054] 3. The present application realizes the visualization of the overall prediction result by reconstructing and integrating the segmentation results of multiple sub-regions and splicing them back to the original image. Compared with the traditional image processing scheme, the present application ensures the segmentation accuracy and global consistency through image reconstruction.

[0055] 4. The present application realizes efficient segmentation and detection of common dental diseases through the dental disease area detection method based on image segmentation and reconstruction algorithm, thereby realizing the preliminary screening function of dental diseases.

[0056] The present application will be further described in detail below in combination with the drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0057] Fig. 1 is a flow chart of an embodiment of the dental disease area detection method based on image segmentation and reconstruction algorithm of the present application.

[0058] Fig. 2 is an image processing process chart of steps S1-S3 of an embodiment of the dental disease area detection method based on image segmentation and reconstruction algorithm of the present application.

[0059] Fig. 3 is an image processing process chart of S4-S5 in an embodiment of the dental disease area detection method based on image segmentation and reconstruction algorithm of the present application. DETAILED DESCRIPTION

[0060] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0061] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. Those skilled in the art will appreciate that embodiments described herein can be combined with other embodiments.

[0062] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. Those skilled in the art will appreciate that embodiments described herein can be combined with other embodiments. Figs. 1-3 The present application provides a dental disease area detection method based on image segmentation and reconstruction algorithm, a dental disease area detection method based on image segmentation and reconstruction algorithm, comprising:

[0063] S1: input an original image, perform region sampling on the original image by a target detection algorithm, and generate N detection boxes containing region information of all suspected diseased areas in the original image.

[0064] S2: adopt threshold suppression processing to filter out the detection boxes with low disease probability.

[0065] S3: set a suppression threshold, obtain the overlap degree of the detection boxes by non-maximum suppression processing, and judge whether the overlap degree of the detection boxes exceeds the suppression threshold, if yes, filter out the detection boxes, if not, keep the detection boxes.

[0066] S4: segment the n detection boxes kept by the segmentation algorithm, and judge whether the pixels of the segmented areas belong to the lesion area, so as to obtain n Boolean matrices belonging to the lesion area.

[0067] S5: reconstruct the n Boolean matrices according to the region information to generate a discrete array with the same size as the original image.

[0068] S6: analyze the pixel position represented by the value of the discrete array and the disease type of the position according to the value of the discrete array. In this embodiment, image preprocessing is performed before inputting the original image, and the processing process includes:

[0069] S11: read the original image, and perform pixel normalization and data enhancement processing on the original image.

[0070] S12: input the processed original image into the neural network, and output the processed original image after data format processing by the neural network.

[0071] Specifically, the pixel value normalization in the embodiment is used to map the pixel value in the interval [0, 1], so that the pixel value change in the image keeps a relatively stable proportional relationship in a certain range, so as to improve the numerical stability and convergence speed of the image processing algorithm.

[0072] Specifically, the pixel value normalization in the embodiment adopts the maximum and minimum value normalization method, and the process includes:

[0073] S111: obtaining the minimum value and the maximum value of all pixel points in the original image;

[0074] S112: calculating the difference between the value of each pixel point in the original image and the minimum value;

[0075] S113: comparing the difference with the difference between the maximum value and the minimum value, so as to obtain the mapping of the pixel value in the interval [0, 1].

[0076] Specifically, the data enhancement processing of the image in step S11 in the embodiment includes image dimension adjustment, dimension exchange, scaling and format conversion, etc., which is used to improve the diversity of image data, so as to improve the generalization ability of the image data model. The above image enhancement processing scheme is only exemplary and is not the only processing method. For example, it also includes image contrast adjustment, cropping, filtering, etc., so as to achieve the following effects according to the actual dental disease image: (1) eliminating the background interference of the image, improving the definition and contrast of the image; (2) enhancing the image features, so as to reduce the consumption of computing resources and reduce the complexity of subsequent processing; (3) removing unimportant parts, focusing on key areas, and adjusting the image size to adapt to the application scene and algorithm requirements of dental disease diagnosis; (4) enhancing the detectability of the information about the lesion area in the original image, improving the reliability of feature extraction, image segmentation, matching and recognition.

[0077] Specifically, the prediction data format of the graph neural network in step S12 in the embodiment adopts a multi-dimensional array form, and the specific format depends on the color mode and dimension of the original image. For example, if the original image is a color image, it adopts a three-dimensional array format, and the representation form of the array is (H, W, C), wherein H is the height of the original image, W is the width of the original image, and C is the channel number of the original image. If the original image is an RGB image, C is 3.

[0078] Specifically, the target detection algorithm in step S1 of the embodiment adopts a deep learning convolutional neural network (CNN), which includes R-CNN algorithm, Fast R-CNN algorithm and Faster R-CNN algorithm. The target detection process of the Faster R-CNN algorithm adopted in the embodiment includes:

[0079] S13: A region proposal network (RPN) is introduced, and the RPN is used to generate candidate boxes containing all region information of the original image.

[0080] S14: Feature extraction and classification are performed on each candidate box.

[0081] S15: The candidate boxes are mapped to positions and sizes close to the real detection boxes through bounding box regression to predict the positions and sizes of the detection boxes.

[0082] Specifically, the RPN network can quickly generate high-quality candidate boxes, thereby avoiding the cumbersome process of generating candidate regions using external algorithms, and improving the speed and accuracy of target detection.

[0083] Specifically, the target detection process of another embodiment adopting the R-CNN algorithm includes:

[0084] S131: The original image is regionally segmented.

[0085] S141: A bounding box is created for the segmented region.

[0086] S151: A search rule is set, and similar regions are merged according to the search rule to obtain the detection box.

[0087] Specifically, the target detection process of another embodiment adopting the Fast R-CNN algorithm includes:

[0088] S132: The entire original image is input into a neural network, and a feature map thereof is calculated.

[0089] S142: A Rol pooling process is performed on the corresponding candidate region extracted from the feature map to obtain a fixed-size feature vector.

[0090] S152: Class prediction and position prediction are respectively performed according to the feature vector.

[0091] In the embodiment, the threshold suppression processing method in step S2 includes:

[0092] S21: A minimum threshold is set according to the image content of the dental disease.

[0093] S22: comparing the pixels in the detection frame with the minimum threshold value, retaining the detection frame greater than or equal to the minimum threshold value, and filtering out the detection frame less than the minimum threshold value.

[0094] S23: repeating step S22 until all N detection frames are processed.

[0095] Specifically, the minimum threshold value in the embodiment is set according to the data characteristics of the original image and combined with actual dental disease diagnosis application, and is adjusted and optimized multiple times to determine the most suitable threshold value through multiple clinical trials and comparisons. By setting the minimum threshold value, the characteristics of the dental disease area can be screened out, ensuring that the probability of the existence of lesions in all detection frames in the target detection result is higher than the minimum threshold value, thereby facilitating more accurate dental disease diagnosis.

[0096] In the embodiment, the non-maximum suppression processing method in step 3 includes:

[0097] S31: calculating the confidence of each detection frame.

[0098] S32: sorting the detection frames in descending order according to the value of the confidence.

[0099] S33: selecting the detection frame with the highest score and calculating the overlap degree of the detection frame with the remaining detection frames.

[0100] S34: determining whether the overlap degree is greater than the suppression threshold value, if yes, retaining the detection frame, if no, setting the score of the detection frame to 0.

[0101] S35: repeating steps S33-S34 until all detection frames are processed.

[0102] In the embodiment, the calculation method of the confidence in step S31 includes:

[0103] S311: obtaining the probability of the existence of lesions in the detection frame.

[0104] S312: calculating the overlap degree of the detection frame with the real detection frame by the YOLO algorithm.

[0105] S313: estimating the confidence of the detection frame according to the maximum overlap degree.

[0106] Specifically, in the target detection, step S3 of the embodiment further includes filtering the detection result by the confidence, specifically, setting a confidence threshold value, only when the confidence of the detection frame is higher than the threshold value, the detection result will be considered valid, so as to improve the accuracy and reliability of the detection target.

[0107] Specifically, in another embodiment, the non-maximum suppression processing method in step 3 includes:

[0108] S3-1: Set the IOU threshold of the detection frame, and sort the class probability of the classifier in descending order.

[0109] S3-2: From the maximum probability, judge whether the overlap of the detection frame exceeds the IOU threshold in turn, if it exceeds, suppress the detection frame, thereby completing the traversal of one detection frame;

[0110] S3-3: Find the maximum probability in the remaining detection frame, and repeat step S3-2 until all detection frames are processed.

[0111] Specifically, the embodiment sets the score of the detection frame overlapping in the same target region but with non-maximum score to 0 by the non-maximum suppression method, thereby eliminating the overlapping detection frame to realize secondary suppression, improving the accuracy of the overall detection and reducing the calculation complexity, which helps to improve the speed and efficiency of target detection.

[0112] In the embodiment, after the threshold suppression and the non-maximum suppression, the number of the detection frames is n, which is consistent with the number of the real target regions.

[0113] In the embodiment, the segmentation algorithm includes:

[0114] S41: Binary segmentation is performed on the region of the detection frame.

[0115] S42: A lesion region threshold is set.

[0116] S43: The pixel of the segmented region is compared with the lesion region threshold, if it is greater than or equal to the lesion region threshold, it means that it belongs to the lesion region, and the pixel is set to True. If it is less than it, it is set to False.

[0117] S44: The pixels with True value form a two-dimensional array.

[0118] S45: Steps S43-S44 are repeated until the comparison of all segmented regions is completed, and n Boolean matrices are obtained, which represent the specific shape of each lesion.

[0119] Specifically, the binary segmentation in step S41 of the embodiment adopts a local binary algorithm, which includes:

[0120] S411: Select a pixel point of the detection frame as the center, and dynamically calculate the threshold of the pixel point according to the gray mean value and standard deviation of the r*r neighborhood of the current pixel point.

[0121] S412: Repeat step S411, traverse the pixels of the detection frame using the above-mentioned r*r template.

[0122] S413: Perform binaryzation calculation using the Sauvola algorithm formula.

[0123] Specifically, in step S43 of the embodiment, if the pixel value is True, it indicates that it is a target pixel, and if the pixel value is False, it indicates that it is a background pixel.

[0124] Specifically, the embodiment can make the calculation model more focused on the details of the key area by efficiently segmenting the local area, positioning the lesion area, and then dividing it into multiple sub-areas for more detailed segmentation processing. This segmentation strategy can reduce the influence of missegmentation and noise points.

[0125] Specifically, in step S5 of the embodiment, image reconstruction includes:

[0126] S51: Obtain the coordinate information and disease category information of each region through the target detection algorithm.

[0127] S52: Reconstruct the n Boolean matrices back to the original image according to the above-mentioned region information, and generate a discrete array with the same size as the original image.

[0128] Specifically, in the discrete array of the embodiment, 0 represents the current pixel as a background pixel, and other numbers represent lesion types, such as: 1 represents the current pixel position as gingivitis, 2 represents the current pixel position as caries, 3 represents the current pixel as calculus, etc. The lesion types of dental diseases specifically include: gingivitis, caries, calculus, wear plane, malformation of lingual fossa, residual root, missing tooth, dental plaque, and wedge-shaped defect.

[0129] Specifically, the embodiment reassembles the segmentation results of multiple sub-areas back to the original image through image reconstruction, thereby realizing the visualization of the overall prediction result and ensuring the segmentation accuracy and global consistency.

[0130] In the embodiment, step S6 further includes:

[0131] S61: Define a set of palettes corresponding to different lesion types.

[0132] S62: Map the discrete array to a superimposed layer of corresponding colors according to the palette.

[0133] S63: Add the pixel values of the lesion area to the superimposed layer in proportion, thereby analyzing the area where the lesion is located through the displayed color.

[0134] Specifically, taking an RGB color picture as an example, a set of color palettes is first defined, for example, 1 represents red, and 2 represents green. A discrete array in the format of ori_H*ori_W is mapped to a color superimposed layer with the color palettes, wherein different colors respectively represent regions corresponding to different diseases. The pixel values of the background region are kept unchanged, and the pixel values of the disease region are added to the superimposed layer color in a certain proportion, so that the disease region can be displayed in the original image in the self-defined color.

[0135] Any combination of the technical features in the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present disclosure.

[0136] The above-described embodiments are only preferred embodiments of the present application, and cannot be used to limit the scope of protection of the present application. Any non-essential changes and replacements made by those skilled in the art on the basis of the present application shall fall within the scope of protection of the present application.

Claims

1. A method for detecting dental disease area based on image segmentation and reconstruction algorithm, characterized in that, The method comprises the following steps: S1: input an original image, perform region sampling on the original image by a target detection algorithm, and generate N detection boxes containing region information of all suspected diseased regions in the original image; S2: perform threshold suppression processing to filter out the detection boxes with low disease probability; S3: set a suppression threshold, obtain the overlap degree of the detection boxes by non-maximum suppression processing, and determine whether the overlap degree of the detection boxes exceeds the suppression threshold; if yes, filter out the detection boxes; if no, retain the detection boxes; wherein, after the threshold suppression and non-maximum suppression, the number of the detection boxes is reduced to n, which is consistent with the number of real target regions; S4: perform segmentation on the n retained detection boxes by a segmentation algorithm, and determine whether the pixels of the segmented regions belong to the lesion regions, so as to obtain n Boolean matrices belonging to the lesion regions; S5: reconstruct the n Boolean matrices according to the region information to generate a discrete array with the same size as the original image; S6: analyze the pixel position represented by the value of the discrete array and the disease type of the position according to the value of the discrete array; wherein, step S6 further comprises: S61: define a palette corresponding to different disease types; S62: map the discrete array to a superimposed layer of corresponding colors according to the palette; S63: add the pixel values of the lesion regions to the superimposed layer in proportion, so as to analyze the region where the lesion is located through the displayed color.

2. The method of claim 1, wherein the image segmentation and reconstruction algorithm is based on a level set method. Before inputting the original image, perform image preprocessing, and the processing process comprises: S11: read the original image, and perform pixel normalization and data enhancement processing on the original image; S12: input the processed original image into a neural network, and output after performing data format processing on the original image by the neural network.

3. The method of claim 1, wherein the image segmentation and reconstruction algorithm is based on a level set method. The target detection algorithm in step S1 adopts a region proposal network, which comprises: S13: generate a candidate box containing all region information of the original image by a region proposal network; S14: perform feature extraction and classification on each candidate box; S15: map the candidate box to a position and size close to the real detection box by boundary regression to predict the position and size of the detection box.

4. The method for detecting dental disease area based on image segmentation and reconstruction algorithm according to claim 1, characterized in that, The target detection algorithm in step S1 adopts a selective search, which comprises: S131: perform region segmentation on the original image; S141: create a boundary box for the segmented region; S151: set a search rule, and merge similar regions according to the search rule to obtain the detection box.

5. The method of claim 3 or 4, wherein the image segmentation and reconstruction algorithm is based on a level set method. The threshold suppression processing method in step S2 comprises: S21: set a minimum threshold according to the image content of dental diseases; S22: compare the pixels in the detection box with the minimum threshold, retain the detection box greater than or equal to the minimum threshold, and filter out the detection box less than the minimum threshold; S23: repeat step S22 until the N detection boxes are completely processed.

6. The method of claim 5, wherein the image segmentation and reconstruction algorithm is based on a level set method. The non-maximum suppression processing method in step 3 comprises: S31: calculate the confidence of each detection box; S32: Score ranking the detection boxes according to the value of the confidence; S33: Select the detection box with the highest score and calculate the overlap degree of the detection box and the remaining detection boxes; S34: Determine whether the overlap degree is greater than the suppression threshold, if yes, keep the detection box; if no, set the score of the detection box to 0; S35: Repeat steps S33-S34 until all detection boxes are processed.

7. The method of claim 6, wherein the image segmentation and reconstruction algorithm is based on a level set method. The calculation method of the confidence in step S31 includes: S311: Obtain the lesion existence probability of the detection box; S312: Calculate the overlap degree of the detection box and the real detection box by the YOLO algorithm; S313: Estimate the confidence of the detection box according to the maximum overlap degree.

8. The method of claim 7, wherein the image segmentation and reconstruction algorithm is based on a level set method. The segmentation algorithm includes: S41: Binary segmentation of the region of the detection box; S42: Set the lesion region threshold; S43: Compare the pixel of the segmented region with the lesion region threshold, if greater than or equal to the lesion region threshold, it means that it belongs to the lesion region, then set the pixel to True; if less than, set it to False; S44: The pixels with True value form a two-dimensional array; S45: Repeat steps S43-S44 until all segmented regions are compared, and obtain n Boolean matrices.

9. The method of claim 1, wherein the image segmentation and reconstruction algorithm-based dental disease area detection method is characterized by, Step S5 image reconstruction includes: S51: Obtain the coordinate information and disease category information of each region by the target detection algorithm; S52: According to the above region information, reconstruct the n Boolean matrices back to the original image, and generate a discrete array with the same size as the original image.

Images