An automatic feature extraction method around tooth boundary and an oral lesion recognition method

Abstract

The application discloses a feature automatic extraction method around a tooth boundary and a lesion recognition method, relates to the technical field of oral medicine, and comprises the following steps: acquiring a digital image of an oral X-ray apical film to be processed; performing automatic segmentation and contour extraction on a target tooth in the digital image of the oral X-ray apical film to obtain a tooth contour line of the target tooth; performing normalization processing on the tooth contour line; and traversing each boundary pixel point on the tooth contour line to generate an edge band feature map of the target tooth. The application can effectively extract local features highly related to lesions, especially features of a tooth boundary region, from the oral X-ray apical film, so that the recognition capability for early micro-lesions is improved.

Description

Technical Field

[0001] This invention relates to the field of oral medicine technology, and more specifically to a method for automatic feature extraction around tooth boundaries and a method for lesion identification. Background Technology

[0002] Oral diseases, such as dental caries, periodontitis, and periapical periodontitis, are common conditions affecting human health. Early diagnosis and timely treatment are crucial for halting disease progression and preserving tooth function. Periapical radiographs are one of the most commonly used imaging methods in oral clinical diagnosis. They clearly display the morphology and structure of tooth hard tissues, pulp chambers, periapical tissues, and surrounding alveolar bone, providing key evidence for the diagnosis of these diseases. With the rapid development of artificial intelligence, especially deep learning technology, intelligent assisted diagnostic systems for medical imaging based on deep learning have shown enormous application potential. In the field of oral medicine, researchers have attempted to apply deep learning models such as convolutional neural networks (CNNs) to the automatic identification and diagnosis of common lesions such as dental caries, alveolar bone resorption (periodontitis), and periapical lesions in periapical radiographs.

[0003] Currently, mainstream methods typically include the following steps: 1) preprocessing the original periapical images; 2) (optionally) segmenting the teeth or specific anatomical structures using an image segmentation network; 3) using the entire image or the segmented tooth region as input to train a deep learning model for lesion classification, detection, or segmentation. Although existing technologies have made some progress in identifying more obvious mid-to-late-stage lesions, their effectiveness in identifying early, small lesions remains unsatisfactory. Analysis suggests the main reasons are: The blindness of feature extraction: Raw X-ray images contain complex information such as teeth, alveolar bone, soft tissue, air, and various noises. Lesions, especially early, small lesions, only occur in specific areas of the tooth boundary (such as the crown margin or root margin). Directly learning features from the entire image or the entire tooth region makes the model susceptible to interference from a large amount of irrelevant background information, making it difficult to focus on truly valuable lesion features.

[0004] The significance of features of small lesions is low: the gray-scale changes of early small lesions on X-ray images are subtle and the spatial range is small. Their features are very weak relative to global information at the scale of the original image, making them difficult to be effectively captured and learned by deep learning models.

[0005] Data imbalance and labeling difficulties: Samples of early, minor lesions are relatively scarce and difficult to label, resulting in insufficient learning of such lesions during model training.

[0006] Therefore, in view of the shortcomings of the existing technology, how to provide an automatic feature extraction method around the tooth boundary and an oral lesion identification method, which can effectively extract local features highly related to the lesion from the periapical radiographs of the oral cavity, especially the features of the tooth boundary area, in order to improve the ability to identify early and small lesions, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides an automatic feature extraction method around the tooth boundary and an oral lesion identification method. By focusing on the tooth contour boundary and its surrounding area—that is, the high-incidence area of ​​lesions—highly targeted edge band feature maps are extracted, thereby converging and flattening the information of potential lesion areas scattered on the tooth boundary, highlighting the features of small lesions and enhancing the salience of lesion features; discarding background area information that is not closely related to the lesion improves the efficiency and accuracy of feature learning and reduces interference from irrelevant information; by standardizing the contour extraction and flattening process, the extracted features have better consistency and establish a clear correspondence with the tooth anatomy structure, improving the consistency and interpretability of features; providing higher quality and more targeted training data for deep learning models, thereby significantly improving their intelligent identification ability of early small caries, alveolar bone resorption and periapical lesions in oral X-ray periapical films, and improving the sensitivity and accuracy of early small lesion identification.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: an automatic feature extraction method around the tooth boundary, comprising: Acquire digital images of the periapical radiographs of the oral cavity to be processed; The target tooth in the digital image of the periapical radiograph is automatically segmented and its contour extracted to obtain the tooth contour line of the target tooth; The tooth contour lines are then standardized. By traversing each boundary pixel on the tooth contour line, an edge band feature map of the target tooth is generated.

[0009] Preferably, the method further includes: dividing the marginal zone feature map into a crown segment marginal zone feature map and a root segment marginal zone feature map, for the purpose of identifying lesions in different locations.

[0010] Preferably, traversing each boundary pixel on the tooth contour line to generate an edge band feature map of the target tooth includes: calculating the normal vector of each boundary pixel; For each boundary pixel, multiple sampling points are generated at equal intervals along the positive and negative directions of its normal vector, and the pixel values ​​of the sampling points in the original oral X-ray periapical digital image are obtained to form the pixel sequence corresponding to the boundary pixel. The pixel sequences of all boundary pixels are concatenated to form an edge band feature map.

[0011] Preferably, the target tooth in the digital image of the periapical radiograph is automatically segmented and its contour extracted to obtain the tooth contour line of the target tooth, including: The target tooth in the periapical slice is automatically segmented based on the image segmentation neural network to obtain a binary mask image of the tooth; Contour extraction is performed on the binarized mask image to obtain polygonal tooth contour lines; The tooth outline is composed of a series of continuous boundary pixels.

[0012] Preferably, the tooth contour line is standardized, including: Set an anatomical landmark as the starting point for the tooth outline; Starting from the starting point, traverse the entire tooth outline in a uniform direction to ensure that the arrangement order of all boundary pixels on the tooth outline is consistent.

[0013] Preferably, the pixel sequences of all boundary pixels are concatenated to form an edge band feature map, including: By horizontally stitching together the one-dimensional pixel sequences corresponding to all boundary pixels in sequence, a two-dimensional image with a width of l and a height of (2d+1) is formed. Where l represents the number of boundary points, d is a preset positive integer parameter, and 2d+1 is the number of normal sampling points; the pixels in the (d+1)th row correspond to the outline boundary line of the original tooth.

[0014] Preferably, a method for identifying oral lesions around tooth boundaries includes: Based on the lesion area marked on the original oral X-ray periapical digital image, it is mapped to the edge band feature map according to its coordinates in the original image and the positional relationship with the tooth contour. A training dataset is constructed using the labeled edge band feature map as input data and the lesion information obtained by mapping or labeling as supervision information. Construct a deep learning model and train the deep learning model based on the training dataset; The trained deep learning model is used to identify lesions on the edge band feature map of the tooth to be tested, and the lesion identification results on the edge band feature map are obtained.

[0015] Preferably, the method further includes: if the lesion identification result is a pixel-level mask or bounding box on the edge band feature map, then according to the correspondence between the edge band feature map and the original tooth contour boundary points, the lesion identification result is reverse mapped back to the original oral X-ray periapical digital image, thereby marking the name of the lesion, the affected boundary area and the extent on the original oral X-ray periapical digital image.

[0016] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an automatic feature extraction method around the tooth boundary and an oral lesion identification method, which has the following beneficial technical effects: Improving the sensitivity and accuracy of early minute lesion identification: By focusing on the tooth boundary areas where lesions are frequently found, a marginal zone feature map is generated, which enhances and highlights the features of minute lesions. This enables deep learning models to learn and identify these lesions more effectively, significantly improving diagnostic performance and helping to achieve early detection and early treatment of diseases.

[0017] Enhancing the efficiency and robustness of model learning: Edge band feature maps discard a large amount of irrelevant background information, and the data dimensionality is relatively reduced. The model can focus more on learning key features related to lesions, reducing the risk of overfitting and improving learning efficiency and the model's generalization ability.

[0018] Enhancing the interpretability of the model: Since there is a clear correspondence between the edge band feature map and the tooth outline, the model's recognition results on the feature map can be directly mapped back to the tooth boundary of the original image, allowing doctors to intuitively see the lesion area that the model is focusing on, thus enhancing the credibility and interpretability of AI-assisted diagnosis.

[0019] Assisting clinical decision-making and improving diagnostic and treatment efficiency: More accurate and sensitive intelligent recognition results can provide powerful assistance to clinicians, helping them to quickly discover potential lesions, especially early and small lesions that are easily overlooked, thereby improving diagnostic efficiency, optimizing treatment plans, and reducing the workload of doctors.

[0020] Good versatility and scalability: This method is not only applicable to the identification of caries, alveolar bone resorption, and periapical lesions, but can also be theoretically extended to intelligent identification tasks of other lesion types occurring at or near the tooth boundary. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of tooth region segmentation and contour extraction provided in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the contour edge band sampling method provided in an embodiment of the present invention.

[0024] Figure 3This is a schematic diagram of edge band feature map generation provided in an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of an oral lesion identification method around the tooth boundary provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention discloses an automatic feature extraction method around tooth boundaries, comprising: Acquire digital images of the periapical radiographs of the oral cavity to be processed; The target tooth in the digital image of the periapical radiograph is automatically segmented and its contour extracted to obtain the tooth contour line of the target tooth; The tooth contour lines are then standardized. By traversing each boundary pixel on the tooth contour line, an edge band feature map of the target tooth is generated.

[0028] Specifically, the process involves traversing each boundary pixel on the tooth contour line to generate an edge band feature map of the target tooth, including: calculating the normal vector of each boundary pixel. For each boundary pixel, multiple sampling points are generated at equal intervals along the positive and negative directions of its normal vector, and the pixel values ​​of the sampling points in the original oral X-ray periapical digital image are obtained by interpolation to form the pixel sequence corresponding to the boundary pixel. The pixel sequences of all boundary pixels are concatenated to form an edge band feature map.

[0029] Specifically, the interpolation method is bilinear interpolation.

[0030] Specifically, the target tooth in the digital image of the periapical radiograph is automatically segmented and its contour extracted to obtain the tooth contour line of the target tooth, including: The target tooth in the periapical slice is automatically segmented based on the image segmentation neural network to obtain a binary mask image of the tooth; Contour extraction is performed on the binarized mask image to obtain polygonal tooth contour lines; The tooth outline is composed of a series of continuous boundary pixels.

[0031] Specifically, the tooth contour lines are standardized by uniformly designating anatomical landmarks as the starting points of the contour lines and standardizing the traversal direction.

[0032] Specifically, the standardization process for the tooth contour lines includes: Set an anatomical landmark as the starting point for the tooth outline; Starting from the starting point, traverse the entire tooth outline in a uniform direction to ensure that the arrangement order of all boundary pixels on the tooth outline is consistent.

[0033] Specifically, the pixel sequences of all boundary pixels are concatenated to form an edge band feature map, including: By horizontally stitching together the one-dimensional pixel sequences corresponding to all boundary pixels in sequence, a two-dimensional image with a width of l and a height of (2d+1) is formed. Where l represents the number of boundary points, d is a preset positive integer parameter, and 2d+1 is the number of normal sampling points; the pixels in the (d+1)th row correspond to the outline boundary line of the original tooth.

[0034] Specifically, the middle row of pixels in the edge band feature map corresponds to the outline boundary line of the tooth.

[0035] Specifically, it also includes the step of segmenting the marginal zone feature map according to the anatomical structure of the crown and root.

[0036] In one specific embodiment of the present invention, an automatic feature extraction method around a tooth boundary includes: Step 1: As Figure 1 As shown, tooth region segmentation and contour extraction 1.1 Input Image Acquisition: Acquire the digital image of the periapical radiograph of the oral cavity to be processed.

[0037] 1.2 Tooth segmentation: Using existing mainstream image segmentation neural networks (such as U-Net, DeepLab series, etc., which have been studied and have high segmentation accuracy and reliable performance), the target teeth in the periapical radiographs are automatically segmented to obtain a binary mask image of the teeth.

[0038] 1.3 Contour Extraction: Contour extraction is performed on the segmented binary mask image of the teeth to obtain the polygonal contour line of the target teeth. This contour line consists of a series of continuous boundary pixels (x_i, y_i), where i = 1, 2, ..., l, and l is the total number of boundary pixels.

[0039] Step 2: Standardization of tooth contour lines

[0040] 2.1 Determine the starting point of the contour: A single anatomical landmark is designated as the starting point of the contour line. Preferably, the mesial cementoenamel junction (CEJ mesial) of the tooth is selected as the starting point of the contour line. If this landmark is unclear or difficult to automatically identify, other easily identifiable anatomical landmarks or preset regular points can be selected.

[0041] 2.2 Unified Traversal Direction: Starting from the starting point, traverse the entire tooth contour line in a unified direction (e.g., clockwise) to ensure that the arrangement order of all boundary pixels (x_i, y_i) on the contour line is consistent. Figure 2 As shown.

[0042] Specifically, step 3: Generation of edge band feature maps

[0043] 3.1 Traverse boundary points: Following the order determined in step 2.2, traverse each boundary pixel point P_i(x_i, y_i) on the tooth contour line.

[0044] 3.2 Calculate the normal vector: For each boundary point P_i, calculate its normal vector on the contour line. The direction of the normal vector is defined as: positive direction from the center of the overall tooth contour outward (or uniformly defined as positive direction from the outside towards the tooth center, which must be consistent throughout the method). The normal vector can be derived from the tangent direction of adjacent boundary points (for example, for point P_i, its tangent direction can be obtained by fitting or differencing three points P_{i-1}, P_i, and P_{i+1}, and the normal vector is perpendicular to the tangent direction).

[0045] 3.3 Sampling along the normal direction and obtaining pixel values: 3.3.1 Sampling Point Generation: Centered on the boundary point P_i, take a distance of d pixels (d is a preset positive integer parameter) along the positive and negative directions of its normal vector (i.e., the inner and outer sides of the tooth), and generate a total of (2d+1) sampling points at equal intervals on this normal segment. The coordinates of these sampling points may be non-integer.

[0046] 3.3.2 Bilinear Interpolation: For each sampling point with non-integer coordinates, the pixel value of that sampling point is calculated using bilinear interpolation, utilizing its neighboring pixel values ​​in the original X-ray image. This solves the problem of insufficient sampling points or difficulty in directly obtaining pixel values ​​when the normal crosses a pixel.

[0047] 3.3.3 Forming a pixel sequence: Each boundary point P_i can obtain a one-dimensional pixel sequence S_i=[v_1,v_2,...,v_{2d+1}] containing (2d+1) pixel values ​​through the above method, where v_{d+1} corresponds to the pixel value of the boundary point P_i itself.

[0048] 3.4 Constructing the Edge Band Feature Map: The one-dimensional pixel sequence S_i corresponding to all boundary points P_i (i=1 to l) is sequentially and horizontally concatenated to form a two-dimensional image with a width of l (number of boundary points) and a height of (2d+1) (number of normal sampling points). This image is defined as the "edge band feature map" of the tooth. In this feature map, the pixels in the (d+1)th row (i.e., the middle row of pixels) correspond exactly to the outline boundary line of the original tooth, such as... Figure 3 As shown.

[0049] Specifically, it also includes: dividing the marginal zone feature map into crown segment marginal zone feature map and root segment marginal zone feature map, which are used to identify lesions in different locations.

[0050] Specifically, step 4: (optional) Segmentation of edge band feature map

[0051] Based on the anatomical characteristics of teeth, the marginal zone feature map generated in step 3 can be further divided along the width direction (i.e., the direction of the original outline) into crown segment marginal zone feature maps and root segment marginal zone feature maps. For example, the cementoenamel junction (CEJ) can be identified as the segmentation point. This allows the crown portion to focus more on the characteristics of lesions such as caries, while the root portion focuses more on the characteristics of alveolar bone resorption (periodontitis) and periapical lesions, providing more targeted training data for the intelligent diagnosis of different lesions.

[0052] In a specific embodiment of the present invention, a method for identifying oral lesions around tooth boundaries employs an edge band feature map extracted based on the above method, including: Based on the lesion area marked on the original oral X-ray periapical digital image, it is mapped to the edge band feature map according to its coordinates in the original image and the positional relationship with the tooth contour. A training dataset is constructed using the labeled edge band feature map as input data and the lesion information obtained by mapping or labeling as supervision information. Construct a deep learning model and train the deep learning model based on the training dataset; train the deep learning model based on the edge band feature map and its corresponding lesion annotation information; The trained deep learning model is used to identify lesions on the edge band feature map of the tooth to be tested, and the lesion identification results on the edge band feature map are obtained.

[0053] Specifically, it also includes the step of mapping the lesion identification results from the edge band feature map back to the original oral X-ray periapical image, specifically: If the lesion identification result is a pixel-level mask or bounding box on the edge band feature map, then according to the correspondence between the edge band feature map and the original tooth contour boundary points, the lesion identification result is inversely mapped back to the original oral X-ray periapical digital image, thereby marking the name of the lesion, the affected boundary area and the extent on the original oral X-ray periapical digital image.

[0054] Specifically, the lesions include caries, alveolar bone resorption (periodontitis), and / or periapical lesions.

[0055] In one specific embodiment of the present invention, a method for identifying oral lesions around the tooth boundary, such as... Figure 4 As shown, the specific steps include: Step 5. Lesion annotation and model training based on edge band feature maps 5.1 Lesion annotation mapping / feature icon labeling: 5.1.1 If the original periapical radiograph already has lesion markings: Map the lesion areas (such as caries extent, alveolar bone resorption area, periapical lesion area) marked on the original X-ray periapical radiograph to the edge band feature map generated in step 3 (or the segmented feature map in step 4) according to their positional relationship with the tooth contour in the original image. Specifically, pixels in the original lesion area located within the normal sampling band of a certain boundary point P_i will correspond to the corresponding row in the i-th column of the feature map.

[0056] 5.1.2 If the original root tip slice does not have lesion markings: the lesion area can be marked directly on the generated edge zone feature map (or segmented feature map).

[0057] 5.2 Model training data construction: The training dataset is constructed using edge band feature maps (or segmented feature maps) as input data and lesion information obtained by mapping or annotation (which can be pixel-level segmentation masks, minimum bounding rectangles of lesion regions, or lesion classification labels) as supervision information.

[0058] 5.3 Deep Learning Model Training: Design and train one or more deep learning models (such as CNN, Transformer, etc.) to enable them to identify lesions based on the input edge band feature map (or segmented feature map), achieving pixel-level segmentation, object detection (such as minimum bounding box localization) or lesion classification.

[0059] Step 6: Lesion identification and outcome mapping (if needed)

[0060] 6.1 Feature map lesion identification: The edge band feature map of the new tooth to be detected (generated through step 1.3) is input into the trained deep learning model to obtain the lesion identification results on the feature map (such as pixel-level mask of the lesion area, bounding box coordinates or classification results).

[0061] 6.2 Mapping the results back to the original image: If the recognition result is a pixel-level mask or bounding box on the feature map, then based on the correspondence between the edge band feature map and the boundary points of the original tooth contour (i.e., the column of the feature map corresponds to the boundary point P_i, and the row corresponds to the sampling position of P_i in the normal direction), the lesion recognition result is inversely mapped back to the original periapical radiograph of the oral cavity, thereby marking the name of the lesion, the affected boundary area and the extent on the original image.

[0062] If the model only outputs lesion classification results (such as whether there is a lesion, the type of lesion, and the degree of lesion), then the classification result can be output directly without step 6.2.

[0063] This invention relates to image processing and analysis of periapical radiographs, particularly a deep learning-based intelligent recognition technology for dental lesions (such as caries, alveolar bone resorption / periodontitis, and periapical lesions). It flattens the three-dimensional, curled tooth contour and focuses on the contour boundary and its adjacent normal region, constructing an edge band feature map through bilinear interpolation sampling. This is fundamentally different from existing technologies that directly use the original image or the entire tooth region as input. It significantly condenses information related to the lesion-related area, eliminates a large amount of irrelevant background, and makes minute lesion features stand out from the complex background, making them easier for deep learning models to capture.

[0064] By standardizing the starting point of the contour (e.g., the mesial cementoenamel junction) and the traversal direction (e.g., clockwise), the spatial consistency and comparability of the edge band feature maps are ensured. This improves the stability of the features and the effectiveness of model training, providing a unified reference system for feature maps from different teeth and images.

[0065] The marginal zone feature map is segmented by crown and root to provide customized feature input for lesions in different common sites. This further enhances the specificity of the features and helps the model learn more precisely about lesions in specific locations.

[0066] By employing a method of sampling at equal intervals along the normal direction combined with bilinear interpolation, the accuracy and continuity of pixel values ​​at the sampling points are ensured, thus resolving the problem of oblique sampling. This guarantees the quality of the edge band feature map and avoids information loss or distortion.

[0067] There is a clear one-to-one correspondence between the marginal zone feature map and the original tooth contour points, which makes it easy to map the recognition results back to the original image and provide an intuitive clinical reference.

[0068] Specifically, Example 1: Early Caries Identification via Periapical Radiograph

[0069] Experimental objective: To verify whether the edge band feature map method proposed in this invention is superior to traditional methods based on original images in identifying early microcaries.

[0070] Experimental data: 1000 periapical radiographs containing early microcartilage (superficial enamel caries, small lesion area, and insignificant grayscale changes) were collected, along with 1000 normal periapical radiographs without caries as a control group. All images were annotated by at least two senior dentists to determine the location and extent of the caries (as the gold standard).

[0071] Image preprocessing: uniform size and grayscale normalization.

[0072] Experimental methods: Tooth segmentation and contour extraction: The target teeth in all periapical slices were segmented using a pre-trained U-Net segmentation network, and their contours were extracted.

[0073] Edge band feature map generation: Set the starting point of the contour to the mesial cementoenamel junction, and traverse the contour clockwise. Set d=5 (i.e., ±5 pixels along the normal, for a total of 11 sampling points).

[0074] For each contour boundary point, 11 sampling points are generated at equal intervals along its normal vector direction. Pixel values ​​are obtained through bilinear interpolation to construct an edge band feature map (width is the number of contour points l, and height is 11).

[0075] Model Training and Evaluation: Experimental Group (Method of this Invention): Using the generated edge band feature map as input, a lightweight CNN model (such as an improved model based on MobileNet or ResNet18) is constructed for training binary classification (present / absent) and pixel-level segmentation of caries.

[0076] Control group (traditional method): The original periapical slice image (or only the segmented tooth region image) was used as input and trained using a CNN model with the same structure.

[0077] The training set, validation set, and test set are divided in a 7:1:2 ratio.

[0078] Evaluation metrics: accuracy, sensitivity / recall, specificity, F1 score, and intersection-over-union (IoU) ratio for the segmentation task.

[0079] Expected experimental results: The experimental group will show a significant improvement in recall sensitivity for early microcaries compared to the control group. The accuracy and F1 score of the experimental group should also be better than those of the control group.

[0080] In the pixel-level segmentation task, the experimental group showed significantly higher accuracy (IoU) in locating small caries areas than the control group. Statistical analysis showed that the difference between the two groups was statistically significant (p<0.05).

[0081] Experimental conclusion: The edge band feature extraction method around the tooth boundary proposed in this embodiment of the invention can effectively improve the recognition performance of deep learning models for early microcaries in periapical radiographs.

[0082] Example 2: Intelligent grading of alveolar bone resorption (periodontitis)

[0083] Experimental objective: To verify the accuracy of the method of the present invention in classifying alveolar bone resorption of different degrees.

[0084] Experimental data: Periapical radiographs of 500 cases with different degrees of alveolar bone resorption (mild, moderate, severe, and no resorption) were collected and graded by periodontologists.

[0085] Experimental method: Tooth segmentation and contour extraction were performed in the same manner as in Case 1.

[0086] Marginal band feature maps are generated and further divided into crown and root segments. Training is performed using only the root segment marginal band feature maps.

[0087] A classification model was constructed to classify the degree of alveolar bone resorption on the feature map of the root segment margin (Level 0: no resorption, Level 1: mild, Level 2: moderate, Level 3: severe).

[0088] The control group used images of the entire tooth region or raw images.

[0089] Expected experimental results: The experimental group was significantly better than the control group in terms of the accuracy of alveolar bone resorption grading and the Kappa coefficient, especially in the ability to identify early mild resorption.

[0090] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0091] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for automatically extracting features around tooth boundaries, characterized in that, include: Acquire digital images of the periapical radiographs of the oral cavity to be processed; The target tooth in the digital image of the periapical radiograph is automatically segmented and its contour extracted to obtain the tooth contour line of the target tooth; The tooth contour lines are then standardized. By traversing each boundary pixel on the tooth contour line, an edge band feature map of the target tooth is generated.

2. The method for automatically extracting features around a tooth boundary according to claim 1, characterized in that, Also includes: The marginal zone feature map is divided into crown segment marginal zone feature map and root segment marginal zone feature map, which are used to identify lesions in different locations.

3. The method for automatically extracting features around a tooth boundary according to claim 1, characterized in that, Traverse each boundary pixel on the tooth contour line to generate an edge band feature map of the target tooth, including: calculating the normal vector of each boundary pixel; For each boundary pixel, multiple sampling points are generated at equal intervals along the positive and negative directions of its normal vector, and the pixel values ​​of the sampling points in the original oral X-ray periapical digital image are obtained to form the pixel sequence corresponding to the boundary pixel. The pixel sequences of all boundary pixels are concatenated to form an edge band feature map.

4. The method for automatically extracting features around a tooth boundary according to claim 1, characterized in that, The target tooth in the periapical radiograph digital image is automatically segmented and its contour extracted to obtain the tooth contour line of the target tooth, including: The target tooth in the periapical slice is automatically segmented based on the image segmentation neural network to obtain a binary mask image of the tooth; Contour extraction is performed on the binarized mask image to obtain polygonal tooth contour lines; The tooth outline is composed of a series of continuous boundary pixels.

5. The method for automatically extracting features around a tooth boundary according to claim 4, characterized in that, The standardization process for the tooth contour lines includes: Set an anatomical landmark as the starting point for the tooth outline; Starting from the starting point, traverse the entire tooth outline in a uniform direction to ensure that the arrangement order of all boundary pixels on the tooth outline is consistent.

6. The method for automatically extracting features around a tooth boundary according to claim 2, characterized in that, The pixel sequences of all boundary pixels are concatenated to form an edge band feature map, including: By horizontally stitching together the one-dimensional pixel sequences corresponding to all boundary pixels in sequence, a two-dimensional image with a width of l and a height of 2d+1 is formed. Where l represents the number of boundary points, d is a preset positive integer parameter, and 2d+1 is the number of normal sampling points; the pixels in the (d+1)th row correspond to the outline boundary line of the original tooth.

7. A method for identifying oral lesions around tooth boundaries, employing the automatic feature extraction method around tooth boundaries as described in any one of claims 1-6, characterized in that, include: Based on the lesion area marked on the original oral X-ray periapical digital image, it is mapped to the edge band feature map according to its coordinates in the original image and the positional relationship with the tooth contour. A training dataset is constructed using the labeled edge band feature map as input data and the lesion information obtained by mapping or labeling as supervision information. Construct a deep learning model and train the deep learning model based on the training dataset; The trained deep learning model is used to identify lesions on the edge band feature map of the tooth to be tested, and the lesion identification results on the edge band feature map are obtained.

8. The method for identifying oral lesions around the tooth boundary according to claim 7, characterized in that, Also includes: If the lesion identification result is a pixel-level mask or bounding box on the edge band feature map, then according to the correspondence between the edge band feature map and the original tooth contour boundary points, the lesion identification result is inversely mapped back to the original oral X-ray periapical digital image, thereby marking the name of the lesion, the affected boundary area and the extent on the original oral X-ray periapical digital image.

Images