Automatic measuring system and device for included angle of long axis of molar

Abstract

The invention discloses a molar long axis included angle automatic measurement system and device, relates to the field of machine learning, and is suitable for contour recognition of a second molar and a wisdom tooth, key anatomical point detection and automatic measurement of a tooth center axis included angle. Aiming at the problems of strong subjectivity, low efficiency and insufficient quantification precision of current clinical subjective analysis and poor robustness and incomplete feature utilization of an existing deep learning scheme, the system comprises four layers of image standardization, target recognition, parameter calculation and result output, wherein the image standardization layer is used for preprocessing an image through denoising, enhancing and zooming to adapt to a model; the target identification layer realizes tooth positioning and key point detection by means of YOLOv8-pose; the parameter calculation layer fits a tooth center axis through anatomical correction and calculates an included angle; and the result output layer realizes visual annotation and structured report output. According to the method, the low-quality image detection robustness is improved, the implementation cost is reduced, the included angle measurement precision is ensured, and a quantitative basis is provided for wisdom tooth impacted risk assessment, orthodontic treatment planning and oral surgery preoperative analysis.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a deep learning model-based oral surface tomography processing technology. It is applicable to the location of key anatomical points (crown apex and root apex) of the second and third molars and the automated measurement of the long axis angle of the teeth. It can be applied to clinical scenarios such as impacted wisdom teeth risk assessment, orthodontic treatment planning, and preoperative analysis of oral surgery, providing dentists with objective and quantitative diagnostic evidence. Background Technology

[0002] As human diets become more refined, uneven degeneration of the jawbone, muscles, and teeth has occurred, leading to an imbalance between tooth and bone volume, manifesting as crowding and impacted teeth. The third molar, being the last tooth to erupt, has an impaction rate of 9.5%-39%, and is most common in people aged 18-25 (Note: The third molar, also known as the wisdom tooth, is the last permanent tooth to erupt in the human mouth; both are different names for the same tooth, and will be referred to as "wisdom tooth" below for simplicity). Impacted wisdom teeth not only easily induce complications such as pericoronitis, dental caries, and jaw cysts, but may also affect the stability of the dentition after orthodontic treatment, leading to relapse of malocclusion and posing long-term oral health risks to patients.

[0003] Current clinical assessments of wisdom tooth development still rely primarily on subjective qualitative analysis by dentists: relying on clinical observation and examination of panoramic imaging, it is impossible to automatically measure the long axis angle between the second molar and the wisdom tooth. This method has significant limitations and is severely out of touch with the needs of clinical diagnosis and research.

[0004] First, it is highly subjective and lacks quantitative evidence. Different doctors have individual differences in their judgment of key anatomical landmarks such as the position of the crown and root, and there is no unified quantitative standard. It is impossible to objectively compare the long axis angle of molars in different cases, which is not conducive to efficacy evaluation and treatment plan optimization in clinical research.

[0005] Secondly, it is inefficient and has a low degree of data digitization. Manual analysis of curved tomographic images is time-consuming and places a relatively heavy workload on doctors. At the same time, manual analysis requires manual recording of results, which is difficult to convert into structured electronic data. Subsequent scientific research statistics require secondary data entry, which increases the workload and is prone to recording errors.

[0006] Third, the accuracy is insufficient. Traditional methods cannot accurately extract key related parameters such as the long axis angle of the wisdom tooth: on the one hand, it is difficult to establish a dynamic tracking database of wisdom tooth growth from adolescence to adulthood, and it is impossible to analyze the evolution of the angle with jawbone development; on the other hand, due to the lack of quantitative indicators, doctors cannot predict the risk of impaction during the critical period of jawbone development in adolescence (12-16 years old), miss the best intervention time, and ultimately can only resort to invasive treatments such as tooth extraction.

[0007] Although some studies in recent years have introduced computer vision and deep learning technologies, existing solutions have not yet broken through the bottleneck of clinical application and are difficult to adapt to the needs of primary healthcare institutions. First, the use of features is incomplete and detached from clinical logic, relying solely on image texture features for model training without integrating the anatomical structure of teeth and the characteristics of jawbone development during puberty. Second, the problem of angular distortion in panoramic tomographic imaging has not been solved. Due to the systematic projection error of horizontal reduction and vertical magnification in the molar area of ​​panoramic radiographs, existing technologies lack targeted anatomical correction mechanisms and cannot accurately correct the coordinate deviation of key points. Consequently, they cannot achieve accurate anatomical positioning of key points of the wisdom tooth and the second molar and accurate measurement of the long axis angle, which are the key information needed for accurate prediction of impacted wisdom teeth in adolescents.

[0008] In summary, there is an urgent need for an intelligent technology solution that combines the measurement of related parameters, data processing, and low deployment costs to achieve accurate identification of key points of molars in curved tomographic images and simultaneously complete the automated measurement of the long axis angle of molars, providing quantitative evidence for the early diagnosis and clinical intervention of impacted wisdom teeth in adolescence. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention proposes an automated measurement system and device for the long axis angle of molars.

[0010] An automated measurement system and device for the long axis angle of molars comprises four layers: image standardization, target recognition, parameter calculation, and result output. Its core function is to automate the measurement of the long axis angle between molars. The image standardization layer includes a raw image acquisition module and an image preprocessing module; the target recognition layer relies on multi-model collaborative recognition to extract and learn key image parameters; the parameter calculation layer includes anatomical correction and quantitative calculation modules; and the result output layer uses a dedicated module to achieve image annotation and structured report output, as detailed below:

[0011] I. Image Standardization Layer

[0012] The raw image acquisition module acquires raw panoramic images of the oral cavity where the molars are located using clinical X-ray equipment (such as mainstream panoramic radiography machines). It supports data interface docking with different models of dental imaging equipment to ensure that the acquired raw images completely cover the molar area, providing compliant initial data for subsequent standardized image processing.

[0013] The image preprocessing module is used to convert non-standardized raw images into standard data for the adaptive model, providing high-quality image input for subsequent automated measurement of the long axis angle. Its core functions include: 1. Receiving clinically acquired curved tomographic images (supporting JPG, BMP, and DICOM formats, resolution 920*1080~3072*1536 pixels); 2. Suppressing X-ray scattering artifacts using an adaptive threshold denoising algorithm and enhancing the grayscale difference between teeth and jawbone using contrast-limited adaptive histogram equalization; 3. Uniformly scaling the image to 640*640 pixels, establishing a coordinate mapping relationship before and after scaling to ensure that subsequent parameters can be back-mapped to the original image, and outputting standardized images and a coordinate mapping table.

[0014] II. Target Recognition Layer

[0015] By employing a detection model based on the YOLOv8-pose architecture, tooth region localization and key point detection are achieved, providing a coordinate basis for the automated calculation of the major axis angle. The model first accurately locates the defined molar region, then specifically detects two sets of core anatomical points (crown apex and root apex) within the localized region, and finally outputs the standardized image coordinates and detection confidence scores of each key point.

[0016] III. Parameter Calculation Layer

[0017] It includes an anatomical correction module and a quantitative calculation module. At its core, it uses linear algebra tools to automate the correction of key point distortions and the calculation of the major axis angle, thus enhancing the mathematical rigor of the method.

[0018] 1. Anatomical Correction Module

[0019] To address the systematic distortion of the molar region in panoramic radiographs, which is horizontally reduced and vertically enlarged, a diagonal scaling matrix is ​​used to linearly transform the original coordinates. Essentially, this corrects the projection geometric error in reverse through matrix operations.

[0020] (1) Core transformation parameters:

[0021] Horizontal scaling factor k h =1.25 (counters 80% reduction in horizontal distance distortion);

[0022] Vertical scaling factor k v = 0.970 (compensates for 103.1% of the magnification distortion at the vertical distance);

[0023] Two-dimensional scaling transformation matrix

[0024] (2) Matrix representation of coordinate correction: The original crown midpoint and root midpoint are both represented as column vectors.

[0025]

[0026] The corrected coordinate vector is the product of the matrix and the original vector:

[0027]

[0028] 2. Quantitative Calculation Module:

[0029] (1) Tooth long axis fitting: using the corrected crown apex C ′ Root tip R ′ With the endpoints as the major axis, it can be represented as a vector:

[0030]

[0031] Its linear equation can be expressed in vector parametric form:

[0032]

[0033] (2) Calculation of the angle between the major axes of the teeth: Extract the vector of the major axis of the wisdom tooth The major axis vector of the second molar Calculate the included angle using the dot product formula:

[0034]

[0035] If θ > 90°, take the supplementary angle 180° - θ.

[0036] IV. Result Output Layer

[0037] Includes only a results output module, enabling the visualization and structured automated output of major axis angle measurement results:

[0038] 1. Visual annotation: Overlaying key points onto the original image;

[0039] 2. Structured Report: Generates patient de-identified information, image parameters, and measurement results (including angles, key point coordinates, and confidence levels), and supports PDF export for integration with the hospital's electronic medical record system.

[0040] The beneficial effects of adopting the above technical solution are as follows:

[0041] 1. Improve robustness: Through image standardization and anatomical correction, it adapts to images from different devices, achieving a detection success rate of ≥85% in low-quality images, with key point localization error ≤2 pixels;

[0042] 2. Reduced implementation costs: The lightweight model does not require a dataset of millions of images, can be trained using a regular GPU, and the processing time for a single image on a basic computer is ≤10 seconds;

[0043] 3. Meets clinical needs: The angle calculation error is ≤±3°, providing reliable data for the measurement of intermolar angles. The structured report directly supports clinical diagnosis and scientific research statistics. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the working principle of the automated measurement system for the long axis angle of molars according to the present invention; (designated as an abstract drawing).

[0045] Figure 2 This is a visualization diagram of the training results described in this invention;

[0046] Figure 3 This is a diagram illustrating the effect of the training scheme described in this invention. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, so that the technical means, creative features, objectives and effects of the present invention can be easily understood.

[0048] First, using the open-source annotation tool Labelme, the clinical panoramic images were annotated with the regions containing the second molar and wisdom teeth, along with key points, to construct a labeled dataset. Then, a YOLOv8 model was trained to achieve molar region localization and key point detection. The key point coordinates were optimized using an anatomical correction module, and the tooth midline angle was calculated using a quantitative calculation module. Finally, the results output layer provided visualized annotations and a structured report, enabling the system's practical application. The specific steps are as follows:

[0049] Step 1: Data Collection and Filtering

[0050] This step focuses on connecting with subsequent Labelme annotation and YOLOv8 model training. Starting from actual clinical needs, it employs a process of "compliant data collection - multi-dimensional screening - privacy protection" to ensure that the data covers image features under different hardware conditions to adapt to the model's cross-device requirements, while eliminating invalid information to guarantee annotation and training accuracy. Simultaneously, it strictly adheres to medical data compliance requirements. The specific operations are as follows:

[0051] (1) Data size and format requirements

[0052] A total of 271 clinical curved tomographic images were collected, covering commonly used clinical formats and resolutions ranging from 1920*1080 pixels to 3072*1536 pixels, ensuring that the data covered the image features under different hardware conditions and improving the device compatibility of subsequent models.

[0053] (2) Data screening criteria

[0054] Invalid images were removed through a two-stage screening process, retaining cases that met the annotation and training requirements. The specific criteria are as follows:

[0055] Primary screening (hardware interference elimination): Exclude images with severe obstruction by the explant, as these images can obscure the anatomical structure of the teeth, making it impossible to accurately locate key points;

[0056] Secondary screening (tooth integrity exclusion): Images with more than 1 / 3 of the teeth being damaged are excluded. Such images show defects such as crown breakage or root fracture of the second molar or wisdom tooth, resulting in the loss of key points.

[0057] (3) Data de-identification processing

[0058] The screened images are subjected to full-dimensional de-identification to eliminate all information that could be associated with the patient's identity:

[0059] 1. Image metadata processing: Read image metadata, delete fields such as patient name, patient ID, date of birth, medical record, and examination date, and replace them with anonymous numeric identifiers;

[0060] 2. Storage Management: The de-identified image labels are stored in sorted order, which facilitates subsequent annotation and dataset partitioning.

[0061] Step 2: Use Labelme to label the filtered images.

[0062] This step focuses on adapting to YOLOv8 model training and ensuring the anatomical accuracy of labeled data. All operations are designed around clinical anatomical rules and model input requirements, as detailed below:

[0063] (1) Preparation before annotation: Pre-setting of annotation system

[0064] A tooth annotation template was pre-built to standardize annotation and mitigate subjective bias, ensuring consistency across all data. The template includes two core annotation categories: bounding box categories and key point labels. After importing the label template into Labelme, users can directly select the preset labels during annotation, eliminating the need for manual input and avoiding spelling errors. See Table 1 for a detailed list of label categories.

[0065] Label Name Actual corresponding part T_pair molars 2M_c Second molar crown apex 3M_c wisdom tooth crown cusp 2M_r Second molar root apex 3M_r wisdom tooth root apex

[0066] Table 1. Correspondence between labels and actual locations on dental X-rays

[0067] (2) Core annotation process: "Segmentation annotation - Key point annotation" cascade execution

[0068] 1. Markings for molars

[0069] Outline drawing logic: Use tangent rectangles for annotation. The rectangle must completely contain the target molar (covering the crown and root), and the four sides of the rectangle must be tangent to the edge of the molar. If the molar is partially obscured by alveolar bone, the rectangle must still be drawn based on the actual complete shape of the molar to ensure complete coverage. The annotation rectangle for horizontally impacted wisdom teeth should clearly define the boundary between the wisdom tooth and the second molar, and try to avoid overlapping of the two rectangles to keep the boundary clear and distinguishable.

[0070] 2. Marking key points

[0071] Load the completed segmentation and annotation file, and only perform point annotations within the tooth area defined by the segmentation mask to avoid background interference. Select the point annotation tool and operate in the order of "crown apex → root apex":

[0072] Cusp: The midpoint of the line connecting the central fossa and the mesial cusp on the occlusal surface of the crown. This is the starting point of the top of the long axis of the tooth anatomy. If the occlusal surface is severely worn, refer to the position at 1 / 2 of the mesial and distal diameter of the crown to ensure accuracy by relying on the symmetry of the tooth anatomy.

[0073] Root apex: For single-rooted teeth, mark the farthest point of the root tip directly; for multi-rooted teeth, mark the geometric center of each root apex; if the root is completely obscured, calculate the location by combining the position of the crown apex and the crown-root ratio.

[0074] Annotation saving: After the single-instance annotation is completed, it is saved as a JSON format file in the images_done folder. The name is consistent with the corresponding image. The file automatically records the contour coordinates and category label.

[0075] (3) Post-annotation data management: quality control and format conversion

[0076] 1. Marking quality spot checks

[0077] Randomly sampled labeled data were blind-reviewed to check the integrity of the segmented contours and the anatomical accuracy of the key points. If issues such as contours overflowing the background or key points deviating from their anatomical positions were found, they should be corrected immediately to ensure data reliability.

[0078] 2. Annotation file format conversion and segmentation

[0079] Format conversion: Batch convert the segmented annotation JSON files to a normalized format (coordinate values ​​divided by image width and height), generating a txt text file, with each line recording the category and key point coordinates.

[0080] Dataset partitioning: The labeled data is divided into training set and test set in an 8:2 ratio to maintain a balance between normal budding and obstructed cases and avoid class imbalance affecting model training. After partitioning, the data is stored according to the "image-label" correspondence for easy model retrieval later.

[0081] Step 3: YOLOv8 Model Training

[0082] This step uses the labeled dataset constructed in step two as input. For molar instance segmentation and core anatomical point detection tasks, the YOLOv8-pose model is trained and optimized. The core principle is to combine the characteristics of small sample datasets with the deployment requirements of primary care equipment. Through reasonable architecture selection, parameter configuration, and training strategies, the model ensures both clinical accuracy and practical feasibility. Details are as follows:

[0083] (1) Model initialization and training parameter configuration

[0084] 1. Model architecture selection:

[0085] Based on the YOLOv8-pose architecture, the output dimensions of key points were adjusted, retaining only the coordinate output channels of the two core anatomical points, namely the crown tip and root tip. Redundant detection modules were removed to reduce invalid calculations and improve inference efficiency.

[0086] 2. Determining Training Parameters

[0087] Configure parameters based on dataset size and clinical accuracy requirements:

[0088] Training rounds: set according to the characteristics of small sample datasets to balance the risk of model convergence and overfitting.

[0089] Batch size: Based on standard GPU computing power settings, ensuring no memory overflow while guaranteeing parameter update efficiency.

[0090] Optimizer and learning rate: The Adam optimizer is selected. The initial learning rate is set according to the normal range for small sample training. A decay strategy is used in the later stage of training to avoid loss oscillation at the end of training.

[0091] Loss function: YOLOv8-pose uses a combination of coordinate loss and confidence loss for optimization, focusing on improving the accuracy of key point localization.

[0092] (2) Training process control and data augmentation strategies

[0093] 1. Data Augmentation Solution

[0094] Enhancement techniques were designed to address the characteristics of curved tomographic images, such as small grayscale differences and easy obscuring of teeth:

[0095] Geometric transformation enhancement: Random rotation and horizontal flipping are performed on the training set images within a reasonable range, and the segmentation mask and key point coordinates are adjusted simultaneously to ensure that the enhanced tooth anatomy is not deformed and the key point positions are matched.

[0096] Enhanced grayscale adjustment: The brightness and contrast of the image are finely adjusted to simulate the image characteristics of different X-ray equipment and improve the model's cross-equipment adaptability.

[0097] Occlusion simulation enhancement: Small-scale occlusion blocks are added to some training samples to simulate the occlusion of tooth roots by alveolar bone, thereby improving the robustness of the model to clinical occlusion cases.

[0098] 2. Training monitoring and early stop mechanism

[0099] Monitoring metrics: Focusing on the validation set, we pay attention to the key point localization error and detection confidence to ensure that the monitoring metrics meet the needs of clinical applications.

[0100] Early stopping mechanism: When the validation set metrics do not show significant improvement for several consecutive rounds, early stopping is triggered to preserve the optimal model weights and avoid ineffective training and overfitting.

[0101] (3) Model performance verification

[0102] 1. Core accuracy verification

[0103] Validation was conducted based on the test set to ensure that the key point localization error met clinical quantitative requirements and that the detection confidence reached the preset threshold, thereby reducing the pressure of subsequent anatomical correction.

[0104] 2. Deployment Feasibility Verification

[0105] Validation was conducted on standard equipment at the grassroots level:

[0106] Inference speed: Ensure that the model inference time for a single image meets the needs of rapid clinical diagnosis.

[0107] Cross-device adaptability: The performance stability of the model was verified under images from different resolutions and different devices, with no significant fluctuations in accuracy, and it is suitable for multi-device application scenarios at the grassroots level.

[0108] After successful verification, the two types of models are encapsulated into a standardized inference interface, with the output format adapted to the subsequent parameter calculation layer. This provides tooth region masks, key point coordinates, and confidence scores, providing input data for anatomical correction and quantitative calculation.

[0109] Step 4: Parameter Calculation

[0110] This step takes the tooth region mask and key point coordinates output by the model in step three as input, and completes numerical optimization and quantitative calculation by relying on the anatomical correction module and the quantitative calculation module. The core is to convert the model detection results into clinically usable accurate parameters through tooth anatomical rule constraints and mathematical algorithm transformation.

[0111] First, the anatomical correction module is activated, loading the built-in dental anatomy rule library. This library stores basic parameters such as the direction of the crown-root apex line, which is based on clinical consensus. The model's output crown-root apex coordinates are compared one by one. If the root apex coordinates exceed the terminal region of the root defined by the segmentation mask, or if the crown apex deviates from the central region of the crown covered by the segmentation mask, it is considered a significant deviation from the reasonable range. In this case, the coordinates are adjusted based on the crown apex position of the same tooth, combined with the anatomical proportions in the rule library, to ensure that the corrected key points conform to the anatomical structure of the tooth. After correction, the corrected key point coordinates are output, retaining only the results where the positioning error meets clinical requirements, providing precise anchor points for subsequent quantitative calculations. Then, the quantitative calculation stage begins. First, the anatomical long axis of a single tooth is fitted using the corrected crown-root apex coordinates as endpoints. After fitting, the long axis vectors of the second molar and the molar are extracted. The angle between the two vectors is calculated by the dot product of the vectors, thus obtaining a relatively accurate estimate of the angle between the long axes of the second molar and the wisdom tooth.

[0112] Step 5: Output Results

[0113] This step focuses on the quantitative calculation results from step four, outputting the results through visual annotation and structured reporting, balancing the needs of clinicians for intuitive viewing with the requirements for electronic data management. The visual annotation process involves layering segmentation masks and corrected key points onto the original curved tomographic image: the segmentation masks use semi-transparent color blocks for differentiation; key points are distinguished by different colored circular markers, with different colors corresponding to the coronal apex and root apex, and the diameter is adaptively adjusted according to the image resolution to ensure clear visibility; after annotation, a fused image of the original image and the annotated layers is generated, supporting zooming in and out, facilitating physician verification of the match between key anatomical locations and measurement results.

[0114] The structured report generation process automatically outputs standardized documents covering patient de-identification information, imaging parameters, measurement results, and correction records. Patient de-identification information retains only anonymous digital identifiers and anonymized examination dates. Imaging parameters include the type of acquisition device, original resolution, and standardized resolution. Measurement results include the angle between the midline of the second molar and the wisdom tooth, the coordinates of corrected key points, and the confidence level of the test. This process strictly complies with clinical data recording standards, ensuring that the report provides doctors with accurate diagnostic references while seamlessly integrating into existing clinical workflows.

Claims

1. A method for automatically measuring the included angle of the long axes of a molar pair, characterized by comprising: obtaining a panoramic image of a patient's oral cavity; standardizing the panoramic image to obtain a standardized image and a coordinate mapping table of an adapted model; processing the standardized image through a target recognition model to locate the molar pair region and detect the core anatomical key points of the second molar and wisdom tooth, and output the standardized image coordinates and detection confidence of the key points; anatomically correcting the key point coordinates to correct the projection geometry error; based on the corrected key point coordinates, fitting the long axis vectors of the second molar and wisdom tooth, and calculating the included angle of the long axes of the two through vector dot product; outputting the measurement result and visual annotation of the included angle and a structured report.

2. The method of claim 1, wherein, The image standardization process specifically includes: using an adaptive threshold denoising algorithm to suppress X-ray scattering artifacts; enhancing the gray difference between teeth and jaw bones through contrast limited adaptive histogram equalization; scaling the image to a preset resolution to establish the coordinate mapping relationship before and after scaling.

3. The method of claim 1, wherein, The target recognition model is based on the YOLOv8-pose architecture, and the core anatomical key points include the second molar crown tip point, the second molar root tip point, the wisdom tooth crown tip point, and the wisdom tooth root tip point, and the key point detection is only performed within the segmentation mask limit range of the molar pair region.

4. The method of claim 1, wherein, The anatomical correction employs a linear transformation with a diagonal scaling matrix, a horizontal scaling factor k h = 1.25, a vertical scaling factor k v = 0.970, and a transformation matrix The corrected coordinates are the product of the transformation matrix and the original keypoint coordinate vector.

5. The method of claim 1, wherein, The long axis angle calculation specifically comprises: taking the corrected crown tip coordinate and the root tip coordinate as end points, respectively fitting a second molar long axis vector and the wisdom tooth long axis vector The angle is calculated by a dot product formula If θ>90°, the supplementary angle 180°-θ is taken as the final angle value.

6. The method of claim 1, wherein, The visual annotation is to superimpose the key points on the original image, and the structured report contains patient de-identification information, image parameters, measurement results, and key point coordinates and confidence, and supports PDF export and hospital electronic medical record system docking. An automatic measurement system for the angle between the long axes of the molars, for performing the method according to any one of claims 1-6, characterized in that it comprises: comprising: an image acquisition module for obtaining the original panoramic image of a patient's oral cavity, supporting data interface docking with different types of dental image devices; an image preprocessing module in communication with the image acquisition module for denoising, enhancing, and scaling the original image to output a standardized image and a coordinate mapping table; a target recognition module in communication with the image preprocessing module for locating the molar pair region and detecting the core anatomical key points through the YOLOv8-pose architecture model, and outputting the key point coordinates and confidence; a parameter calculation module in communication with the target recognition module, including an anatomical correction unit and a quantitative calculation unit, for correcting the key point coordinates and calculating the included angle of the long axes; a result output module in communication with the parameter calculation module for automatically outputting visual annotations and structured reports.

Images