Dental implant detection system integrating multi-angle CBCT projection and deep learning

Abstract

The present invention relates to a dental implant detection system configured to support accurate detection of dental implants using cone beam computed tomography (CBCT) data. More specifically, the present invention relates to a dental implant detection system that integrates multi-angle CBCT projection and deep learning, wherein the system converts three-dimensional CBCT data into multi-angle projection images, automatically detects the location and status of an implant by optimizing image quality and applying deep learning algorithms, thereby minimizing time-related and human errors that may occur in conventional manual diagnostic methods and improving detection accuracy.

Description

Dental implant detection system combining multi-angle CBCT projection and deep learning

[0001] The present invention relates to a dental implant detection system that supports the accurate detection of dental implants by utilizing Cone Beam Computed Tomography (CBCT) data. More specifically, the invention relates to a dental implant detection system that combines multi-angle CBCT projection and deep learning, which converts 3D CBCT data into multi-angle projection images, automatically detects the location and condition of implants by optimizing image quality and applying deep learning algorithms, thereby minimizing temporal and human errors that may occur in conventional manual diagnostic methods and improving detection accuracy.

[0002] A dental implant detection system is a technology that detects the location and condition of implants by analyzing dental medical imaging data. This technology visualizes implants and surrounding structures through precise image analysis, enabling medical professionals to improve diagnostic accuracy and establish plans necessary for implant procedures and post-operative care.

[0003] In existing dental implant detection systems, the primary method used was for medical professionals to identify the location of implants and make diagnoses based on manual analysis and single-angle CBCT images. However, since this method relies heavily on the experience and intuition of medical professionals, variations in diagnostic accuracy could occur. Furthermore, the limitation of observing images only from specific angles resulted in incomplete analysis of anatomical information regarding structures surrounding the implant. Additionally, existing systems lacked quantitative data analysis and insufficient automation of medical image processing, leading to time-consuming processes and a high probability of errors.

[0004] Therefore, there is an urgent need to develop a dental implant detection system that can solve the aforementioned problems, analyze tooth image data to visualize the anatomical structure of teeth more precisely, improve detection accuracy by automatically detecting implants, and thereby systematically and rapidly analyze the location and condition of implants to reduce the burden on medical staff while providing patients with more reliable diagnoses and treatment plans.

[0005] Accordingly, the technical problem of the present invention is conceived from this point, and the objective of the present invention is to provide a dental implant detection system that combines multi-angle CBCT projection and deep learning, which provides data of an implant observed from various viewpoints through multi-angle CBCT projection.

[0006] In addition, the invention provides a dental implant detection system that combines deep learning with multi-angle CBCT projection, which provides projection images from various viewpoints generated by multi-angle CBCT projection.

[0007] In addition, it provides a dental implant detection system that combines multi-angle CBCT projection and deep learning to automate implant detection.

[0008] In addition, it provides a dental implant detection system that combines deep learning with multi-angle CBCT projection, which adjusts the range of density values ​​through various correction techniques and minimizes distortion in the metal virtual image area.

[0009] In addition, it provides a dental implant detection system that combines deep learning with multi-angle CBCT projection, which removes unnecessary noise by applying filtering techniques such as Gaussian smoothing.

[0010] In addition, the invention provides a dental implant detection system that combines multi-angle CBCT projection and deep learning to ensure consistency between multi-angle projected images through voxel resampling and correction of physical characteristics of the original data.

[0011] In addition, it provides a dental implant detection system that combines deep learning with multi-angle CBCT projection to minimize distortion caused by metal artifacts.

[0012] In addition, it provides a dental implant detection system that combines deep learning with multi-angle CBCT projection to optimize image quality.

[0013] In addition, it provides a dental implant detection system that combines deep learning with multi-angle CBCT projection, which provides high-quality dental implant images.

[0014] In addition, the invention provides a dental implant detection system that combines multi-angle CBCT projection and deep learning, utilizing projection images with optimized quality as training data for a deep learning model.

[0015] To realize the objective of the present invention, a dental implant detection system combining multi-angle CBCT projection and deep learning comprises: an input unit that receives input data from the outside into the system; an output unit that outputs result data to the outside of the system; a communication unit that transmits and receives data between the inside and outside of the system; a storage unit that stores data generated by the system; a control unit that controls the operation of the system; and a memory unit that executes various programs and stores data accordingly. In this system, the memory unit comprises: a CBCT (Cone Beam Computed Tomography) data input unit that generates 3D volume data; an MIP generation unit that converts the 3D volume data into a multi-angle 2D projection image; an image quality optimization unit that optimizes the quality of the 2D projection image to improve the accuracy of analysis; a data annotation and preparation unit that annotates implant information in the 2D projection image and prepares deep learning training data; and a deep learning analysis unit that automatically detects implants based on the annotated implant information and a deep learning training model. The present invention provides a dental implant detection system combining multi-angle CBCT projection and deep learning, comprising a performance evaluation unit that quantitatively evaluates the performance of the deep learning model to verify the reliability of the results.

[0016] According to the present invention, according to embodiments of the present invention, by observing the implant from various viewpoints through multi-angle CBCT projection, information loss that may occur at a single angle can be compensated for, and the accuracy of implant detection can be improved by utilizing deep learning-based detection.

[0017] In addition, projected images from various viewpoints generated by multi-angle CBCT projection can increase the diversity of the dataset for training deep learning models, thereby improving the generalization performance of the model.

[0018] In addition, deep learning-based implant detection systems can automate implant detection, reducing manual work for medical staff and allowing them to spend more time on diagnosis and treatment planning.

[0019] In addition, by utilizing multi-angle CBCT projection and deep learning models, variables in the implant detection process can be reduced, and the accuracy and consistency of diagnostic results can be improved.

[0020] In addition, the accuracy of the final projected image can be improved by adjusting the range of density values ​​through various correction techniques and minimizing distortion in the metal virtual image region.

[0021] In addition, by applying filtering techniques such as Gaussian smoothing to remove unnecessary noise, errors that may occur during the data analysis and diagnosis process can be reduced.

[0022] In addition, consistency between multi-angle projected images can be ensured through voxel resampling and correction of the physical characteristics of the original data.

[0023] In addition, by minimizing distortion caused by metal artifacts, the accuracy of the diagnosis can be improved even in CBCT data containing implants or metal prostheses.

[0024] In addition, optimizing image quality enhances data reliability, enabling clinicians to receive more accurate information necessary for diagnosis and treatment planning.

[0025] In addition, by providing high-quality dental implant images, medical staff can quickly identify anatomical features and reduce analysis time.

[0026] In addition, when quality-optimized projection images are used as training data for deep learning models, they can improve model performance and enhance the accuracy of detection and prediction.

[0027] FIG. 1 is a block diagram of a dental implant detection system combining multi-angle CBCT projection and deep learning according to one embodiment of the present invention.

[0028] FIG. 2 is a drawing showing CBCT data according to one embodiment of the present invention.

[0029] Figure 3 is a diagram showing the overall operation of a dental implant detection system combining multi-angle CBCT projection and deep learning.

[0030] Figure 4 is a diagram showing the process of generating 3D volume data by processing data in the CBCT data processing unit.

[0031] FIG. 5 is a diagram showing the generation of a projected image in an MIP generation unit according to an embodiment of the present invention.

[0032] FIG. 6 is a diagram showing the process of improving image quality in an image quality optimization unit according to an embodiment of the present invention.

[0033] FIG. 7 is a diagram showing the process of annotating implant information in a data annotation and preparation unit according to an embodiment of the present invention.

[0034] FIG. 8 is a diagram showing the operation of a deep learning analysis unit according to an embodiment of the present invention.

[0035] FIG. 9 is a diagram exemplarily showing performance indicators evaluated by a performance evaluation unit according to one embodiment of the present invention.

[0036] Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

[0037] The present invention is capable of various modifications and may take various forms, and specific embodiments are illustrated in the drawings and described in detail in the text. However, this is not intended to limit the invention to the specific disclosed forms, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention.

[0038] FIG. 1 is a block diagram of a dental implant detection system combining multi-angle CBCT projection and deep learning according to one embodiment of the present invention.

[0039] Referring to FIG. 1, a dental implant detection system (100) combining multi-angle CBCT projection and deep learning is composed of an input unit (110), an output unit (120), a communication unit (130), a storage unit (140), a control unit (150), and a memory unit (160). The input unit (110) inputs external data into the system, the output unit (120) outputs data from the system to the outside, the communication unit (130) communicates data inside and outside the system, the storage unit (140) stores data generated by the system, the control unit (150) controls all operations of the system, and the memory unit (160) stores execution of various programs and data associated therewith.

[0040] Additionally, the memory unit (160) includes a CBCT data processing unit (170), an MIP generation unit (180), an image quality optimization unit (190), a data annotation and preparation unit (200), a deep learning analysis unit (210), and a performance evaluation unit (220). The CBCT data processing unit (170) receives CBCT (Cone Beam Computed Tomography) data and generates 3D volume data, the MIP generation unit (180) converts the 3D volume data into multi-angle 2D projection images, the image quality optimization unit (190) optimizes the quality of the 2D projection images to improve the accuracy of the analysis, the data annotation and preparation unit (200) annotates implant information in the 2D projection images and prepares deep learning training data, the deep learning analysis unit (210) automatically detects implants based on the annotated implant information and the deep learning training model, and the performance evaluation unit (220) quantitatively evaluates the performance of the deep learning training model to verify the reliability of the results.

[0041] FIG. 2 is a drawing showing CBCT data according to one embodiment of the present invention.

[0042] Referring to Fig. 2, CBCT data refers to three-dimensional image data acquired through Cone Beam Computed Tomography technology. CBCT is widely used in medical fields, such as dentistry, diagnosis of the oral cavity and jawbone, planning of implant procedures, and visualization of anatomical structures of the head and neck. Cone Beam Computed Tomography technology uses a cone-shaped X-ray beam to scan specific parts of a patient from various angles and reconstructs them using computer algorithms to generate high-resolution three-dimensional volume data.

[0043] CBCT data is stored in the DICOM (Digital Imaging and Communications in Medicine) format, and slice images are sequentially aligned to form a single continuous 3D data set. This data is expressed in voxel (three-dimensional pixel) units, where each voxel represents a numerical value of the radiation absorption rate in that space. In CBCT data, Hounsfield Unit (HU) values ​​are measured based on differences in tissue density and are used to distinguish anatomical structures such as air, soft tissue, bone, and metal. For example, air is approximately -1000 HU, soft tissue ranges from 0 to 200 HU, bone ranges from 300 to 1000 HU, and implants or metal are expressed as 2500 HU or higher.

[0044] The biggest advantage of CBCT data is that it has a lower radiation exposure compared to conventional CT and can acquire data in a relatively short time. In addition, CBCT provides isotropic resolution, allowing for the maintenance of the same level of image quality from all directions. This enables the visualization and analysis of precise three-dimensional anatomical structures and is particularly useful for determining the location of dental implants, diagnosing periapical lesions, and planning orthodontic treatment.

[0045] Consequently, CBCT data provides accurate and detailed 3D information, serving as an important decision-making tool for medical professionals when formulating diagnoses and treatment plans. Based on this, various technical analyses are possible, such as subsequent processing like Maximum Intensity Projection (MIP), generation of 2D projected images, and training of deep learning-based object detection models. As CBCT data can be utilized not only for simple visualization but also for quantitative analysis, it has established itself as an essential data format in the field of medical imaging.

[0046] Figure 3 is a diagram showing the overall operation of a dental implant detection system combining multi-angle CBCT projection and deep learning.

[0047] Referring to FIG. 3, a dental implant detection system (100) combining multi-angle CBCT projection and deep learning is a system that effectively detects implants by precisely analyzing three-dimensional anatomical information based on CBCT (Cone Beam Computed Tomography) data. First, original CBCT data is input, and a Maximum Intensity Projection (MIP) technique is applied to generate two-dimensional projection images from multiple angles. In this process, multi-angle projection images are acquired by considering various angles of the CBCT data (such as the front and 45 degrees to the left and right), and Hounsfield Unit (HU) correction and image quality optimization techniques are applied in parallel so that the anatomical structure and characteristics of the implant can be clearly expressed. As a result, a high-quality projection image with improved visual clarity of the implant is prepared.

[0048] The generated multi-angle projection images undergo a data annotation step and are utilized as training data. Using an annotation tool (LabelIMG), the location of an implant is defined as a bounding box, and this location information is stored in a structured data format along with the class label (implant). This annotated data is used to train a deep learning-based object detection model. Subsequently, a mainstream object detection model, such as Faster R-CNN, is applied to learn and detect implant objects. The model receives multi-angle projection images as input to detect the location of implants and optimizes detection accuracy through iterative training and validation. To complement this, Three-Fold Cross-Validation is performed to strengthen the model's reliability and prevent overfitting, thereby improving generalization performance.

[0049] Finally, the trained model utilizes test data to derive implant detection results. In this process, performance metrics such as Precision, Recall, and F1 score are measured by comparing the bounding boxes predicted by the model with the actual annotated data. Achieving a high F1 score demonstrates that the model is operating effectively, reducing false positives and eliminating false negatives. The final detection results are output through a visualization tool, enabling the identification and diagnosis of implant locations.

[0050] Figure 4 is a diagram showing the process of generating 3D volume data by processing data in the CBCT data processing unit.

[0051] Referring to FIG. 4, the CBCT data processing unit (170) plays a key role in the initial stage of the present invention and reads Cone Beam Computed Tomography (CBCT) data and converts the data into 3D volume data that can be subsequently processed and analyzed. The CBCT data processing unit (170) provides the basis for operations to utilize CBCT data and generates highly reliable data by precisely processing the input data.

[0052] First, the CBCT data processing unit (170) reads data in the DICOM (Digital Imaging and Communications in Medicine) format and generates 3D volume data. To do this, it utilizes the Python-based SimpleITK library (an open-source framework for medical image processing and analysis) and 3D Slicer software, and in this process, it extracts metadata such as the original size, voxel spacing, and origin position of the CBCT image. The above information ensures the accuracy of projection image generation and analysis in subsequent data processing steps. For example, the voxel spacing determines the resolution of the data, and the origin position serves as a reference for data alignment in a 3D coordinate system.

[0053] The above process is explained in more detail as follows. The process of reading data in DICOM (Digital Imaging and Communications in Medicine) format to generate 3D volume data is a crucial first step in performing precise medical image analysis using CBCT data. This process is performed using the Python-based SimpleITK library and 3D Slicer software, focusing on preserving the original information of the CBCT data without loss and providing accurate and reliable data for subsequent analysis and visualization. This process is essential to preserve anatomical information while converting CBCT data into a three-dimensional structure, and to ensure the accuracy and consistency of analysis in subsequent steps.

[0054] DICOM data is a standard format for medical imaging data that contains multiple slice images and includes important metadata such as patient information, image resolution, scan area size, and scan intervals. The SimpleITK library is a tool capable of reading and processing this data, automating the process of integrating slice images into a series of aligned 3D data. The first step in reading a DICOM file through SimpleITK is to load the corresponding DICOM series and align them in the correct order. This alignment process prevents data distortion issues that may occur if scan slices are not placed in the correct order.

[0055] After reading the DICOM data, 3D volume data is generated based on each cross-sectional image. This 3D volume data represents the CBCT scan area as a data structure composed of voxels and is designed to maintain spatial resolution and coordinate system information. In this step, important metadata such as voxel spacing, origin, and data size are extracted. Voxel spacing indicates the size occupied by each voxel in actual space and is a key factor in determining data resolution. For example, if the voxel spacing is not set accurately, aspect ratio distortion may occur in the reconstructed image, which can reduce the reliability of the data analysis process. The origin defines the reference point for data alignment in the 3D coordinate system and is essential information for maintaining consistency during alignment and comparison operations with other image data. For example, this metadata is extracted from the "Pixel Spacing," "Slice Thickness," and "Image Position (Patient)" tags within the DICOM file.

[0056] SimpleITK provides various functions to manipulate and analyze the 3D volume data generated in the above process. Using Python code, you can check key information such as voxel spacing, origin location, and data size, and utilize this to perform tasks such as aligning or resizing the data. For example, the `volume.GetSpacing()` function returns voxel spacing information, and the `volume.GetOrigin()` function checks the origin of the data. Such processes are performed to ensure the accuracy and quality of the data.

[0057] The generated 3D volume data is imported into 3D Slicer software for visualization and further verification. 3D Slicer is a tool for visualizing, analyzing, and processing medical imaging data, allowing the data to be separated into axial, coronal, and sagittal planes for observation from various angles. This visualization provides an opportunity to evaluate data precision and, if necessary, correct incorrect metadata or coordinate system errors. Additionally, 3D Slicer allows for detailed data examination and enables clear observation of clinically important structures, such as nerve canals, bone structures, and surrounding tissues.

[0058] The CBCT data processing unit (170) reads data in the DICOM (Digital Imaging and Communications in Medicine) format to generate 3D volume data, and then utilizes the metadata extracted from the above process for subsequent steps of generating and analyzing projection images. The voxel spacing is used to maintain physical size during the projection image generation (MIP) process and to accurately represent anatomical structures in each projection direction. The origin position maintains consistent alignment between multi-angle projection images and enables precise coordinate-based analysis.

[0059] Additionally, the CBCT data processing unit (170) constructs 3D volume data and converts the input data into an analyzable format through linear interpolation and array transformation processes. In particular, before converting the volume data into a projected image, the density values ​​of each voxel (the smallest unit of CBCT volume data) are systematically organized to accurately reflect the 3D structure. This process is important for preserving the anatomical information of the CBCT data and minimizing distortion or loss of the original data. Linear interpolation is a process that ensures the continuity of the 3D volume data by generating intermediate data based on the spacing (slice thickness) between cross-sectional images. In this process, the density value of each voxel is calculated based on the density values ​​of two adjacent slices. For example, when the density values ​​of the two slices are V1 and V2, respectively, the density value at the interpolation location is V x It is calculated as =V1+(V2-V1)*d. This interpolation process smoothly connects data between slices and ensures the accuracy and consistency of 3D volume data. Additionally, the array transformation process serves to systematically organize data by converting 2D cross-sectional images into a 3D array structure. In this process, density values ​​are mapped to each voxel according to x, y, and z coordinates, enabling efficient storage and access of data in computer memory. The array size and voxel resolution are set based on DICOM metadata (e.g., "Pixel Spacing", "Slice Thickness"). Through slice alignment and coordinate mapping, the 3D volume data accurately reflects the anatomical structure of the CBCT, and consistency is maintained during the analysis process through data normalization.

[0060] Additionally, the CBCT data processing unit (170) includes basic filtering functions to minimize noise and defects that commonly occur in medical image data. It detects irregular pixels or damaged areas included in the original data and corrects them to maintain data quality.

[0061] The above process largely consists of detecting irregular pixels, identifying damaged regions, data correction, and verifying results. First, techniques such as threshold-based filtering, continuity analysis, and structural pattern analysis are used to detect irregular pixels and damaged regions. Pixel values ​​are considered outliers if they fall outside the normal range or differ excessively from surrounding pixel values. Subsequently, connectivity with neighboring pixels is evaluated to determine the extent and form of damage, and if outliers form clusters, they are identified as damaged regions. By analyzing whether specific patterns (e.g., striped noise, incomplete linear data) appear in the above steps, a basis is provided for selecting an appropriate correction method. Once damaged data is identified, various correction methods such as median filtering, linear interpolation, and non-local mean filters are applied. Median filtering is effective for simple noise removal, while linear interpolation restores damaged pixel values ​​based on neighboring pixel values ​​while maintaining data continuity.

[0062] When more sophisticated restoration is required, Gaussian smoothing or AI-based restoration algorithms are applied, while boundary-preserving algorithms are used to restore structural features such as bones. These correction processes are designed to preserve the original structure as much as possible while restoring the reliability of the damaged data. The accuracy of the corrected data is evaluated through multiple stages of verification. Visual inspection confirms that damage has been properly repaired, and statistical evaluation verifies that the corrected pixel values ​​fall within the normal range. Additionally, data continuity is analyzed to ensure that no additional distortion occurred during the correction process.

[0063] Finally, the corrected data undergoes normalization and multi-resolution transformation processes and is stored in a state suitable for subsequent analysis. This process preserves both the original and corrected data, enabling restoration operations if necessary. In conclusion, the process of detecting and correcting irregular pixels and damaged areas improves data quality and enhances the reliability of the analysis.

[0064] Additionally, the CBCT data processing unit (170) stores necessary metadata and parameters based on 3D volume data in a structured format. Generally, information is stored using a standardized format such as JSON, XML, or a relational database. This serves as a database that can be referenced by subsequent algorithms when performing projection image generation and analysis tasks. The above information is essential for multi-angle projection and deep learning-based analysis.

[0065] In conclusion, the CBCT data processing unit (170) is responsible for the basic operations of the present invention and includes the entire process from data input to 3D volume data generation and quality assurance. The CBCT data processing unit (170) processes CBCT data stably and reliably, providing a foundation that enables accurate implant detection in subsequent steps.

[0066] FIG. 5 is a diagram showing the generation of a projected image in an MIP generation unit according to an embodiment of the present invention.

[0067] Referring to FIG. 5, the MIP generation unit (180) converts CBCT data into 2D projected images from multiple angles to generate input data suitable for a deep learning-based implant detection model. The MIP generation unit (180) generates high-quality projected images by calculating maximum density values ​​in each projection direction based on 3D volume data, which plays an important role in significantly improving the learning and detection accuracy of the subsequent object detection model.

[0068] The MIP generation unit (180) first receives 3D volume data generated by the CBCT data processing unit (170). The data is an array of density values ​​composed of voxels, representing anatomical structures in 3D space. The MIP generation unit (180) iterates through the volume data according to the projection direction and calculates the maximum density value along the voxel path in each direction. For example, for frontal projection, the voxel values ​​are compared based on the Z-axis, and for left-right 45° projection, the same calculation is performed after rotational transformation. In the above process, the MIP is mathematically defined as follows.

[0069] MIP(x,y) = max{V(x,y,z)lz [Zmin, Zmax]}

[0070] Here, V(x,y,z) represents the density value at the point (x,y,z) of the volume data, and Zmin and Zmax define the range of the projection direction.

[0071] The MIP generation unit (180) utilizes correction techniques, such as linear interpolation, to preserve the accuracy of the original data when generating a projected image. During the projection process, data continuity is maintained, and background values ​​are set to 0 to minimize the influence of noise. Additionally, connectivity between voxel density values ​​is strengthened so that the projected pixels can more accurately reflect the actual anatomical structure.

[0072] The generated projected images are output at various angles, such as a front view, a 45° left view, and a 45° right view, providing high-quality images with a resolution of 640×640. This contributes to securing diversity in training data by obtaining multi-angle data from a single CBCT dataset. In addition, by reflecting features from various angles, it supports deep learning models in learning the complex 3D structure of implants.

[0073] FIG. 6 is a diagram showing the process of improving image quality in an image quality optimization unit according to an embodiment of the present invention.

[0074] Referring to FIG. 6, the image quality optimization unit (190) ensures the reliability and accuracy of the final image by performing image quality optimization techniques in parallel during the projection process. For example, it improves contrast by adjusting the display range of density values ​​through Hounsfield Unit (HU) correction, and minimizes distortion caused by artifacts by performing voxel resampling and edge sharpening processing in the metal virtual image area. In addition, it adjusts the overall contrast ratio by applying adaptive histogram equalization and removes unnecessary noise using filtering techniques such as Gaussian smoothing.

[0075] The image quality optimization unit (190) corrects CBCT data using a Hounsfield Unit (HU) correction technique. HU is a unit that measures the radiation absorption rate of tissues and includes a range from air (-1000 HU) to metal (>2500 HU). The correction process is performed in the following manner. The density window width and window level adjustment process sets the display range of density values ​​to be suitable for specific tissues (e.g., soft tissue, bone). For example, to emphasize bone structures, a range including high HU values ​​is selected. The image contrast optimization process ensures that HU correction clarifies the boundaries of tissues in low-contrast areas and clearly reveals important anatomical features.

[0076] Metal artifacts in CBCT data occur around metal implants or prostheses and can cause distortion in the projected image. The image quality optimization unit (190) applies the following techniques to minimize these artifacts. The voxel resampling technique corrects distortion by recalculating voxel values ​​around the metal using linear interpolation or advanced interpolation algorithms. Through this, anatomical structures around the metal are reproduced more accurately. The edge sharpening processing technique improves sharpness by applying an edge detection algorithm to solve the problem of blurred boundaries around the metal. The above process enables more accurate representation of important anatomical structures.

[0077] The image quality optimization unit (190) optimizes the contrast ratio of the entire image using Adaptive Histogram Equalization (AHE). Unlike general histogram equalization, AHE divides the image into small blocks and adjusts the contrast ratio independently for each area. The above technique provides the following advantages: local contrast enhancement, which adjusts images with locally unbalanced brightness and contrast so that details in each area are better revealed; and uniform visual quality, which ensures that brightness is evenly distributed throughout the image so that problems such as certain areas becoming excessively bright or dark are prevented.

[0078] The image quality optimization unit (190) removes unnecessary noise contained in the image using filtering techniques such as Gaussian smoothing. The filtering process is performed as follows: High-frequency noise is removed through low-frequency filtering, and the smoothness of the image is improved (noise reduction). Gaussian smoothing preserves important boundaries and structures during the noise removal process, preventing the data from being oversimplified (maintaining continuity), and removes unnecessary point noise or small artifacts to increase the visual clarity of the image (improving visual quality). In particular, an adaptive thresholding technique is applied to reduce noise that may occur in high-density areas or at boundaries. The technique removes noise and preserves the signal by setting a dynamic thresholding value suitable for each area based on the average density value and standard deviation of local areas within the image.

[0079] In conclusion, the image quality optimization unit (190) performs the role of generating high-quality data optimized for deep learning model training from CBCT data converted into 2D images from multiple angles.

[0080] FIG. 7 is a diagram showing the process of annotating implant information in a data annotation and preparation unit according to an embodiment of the present invention.

[0081] Referring to FIG. 7, the data annotation and preparation unit (200) performs the role of annotating object information on a projection image generated based on CBCT data and preparing a dataset suitable for training a deep learning model. The data annotation and preparation unit (200) structures the data, converts it so that it can be used in the training and verification stages, and provides a foundation that ensures the accuracy and reliability of the deep learning model.

[0082] The data annotation and preparation unit (200) adds location information of the implant to each projected image using an annotation tool such as LabelImg. This process is performed manually and uses a bounding box to clearly mark the area containing the implant. The annotation data includes coordinate information (x, y, width, height) and class label information of the implant, and this information plays an essential role in the model learning implant detection. The annotation work is performed by experts to ensure the reliability of the data and is strictly managed to ensure that each image is accurately labeled.

[0083] The data annotation and preparation unit (200) designates the area where the implant is located in each projected image as a bounding box. This process is performed manually, and the area where the implant is located is defined by pixel coordinates within the image. For example, if the implant starts at point (150, 200) from the top-left of the screen, has a width of 50 pixels, and a height of 80 pixels, the bounding box is stored with coordinates containing the above information.

[0084] The annotation data includes class labels indicating the type of object along with the aforementioned coordinate information. For example, a dental implant may be designated with the class label 'Implant'. This annotation data helps the model go beyond simply learning the location of objects to identify what a specific object is. To enhance operational accuracy, this process is handled by experts, and errors are minimized through a double verification procedure to ensure data reliability. The final data is organized in a format that includes bounding box coordinates and class labels for each image. For example, if an image with the filename 'xray_image_001.png' has bounding box information of (150, 200, 50, 80) and a class label of 'Implant', this data is used to train an object detection model.

[0085] The annotated data is converted into the COCO (Common Objects in Context) format. The COCO format is a standard data structure widely used for training object detection models, storing metadata for each image and object. The format is optimized to allow the model to quickly read and process object information during training, and systematically links CBCT projection images with implant location information.

[0086] COCO format data is stored as a JSON file and consists of `images`, which stores image metadata; `annotations`, which contains object information; and `categories`, which defines the class names and IDs of objects. For example, in the case of CBCT projection images, the filename, dimensions (width and height), and image ID are stored in the `images` section. Subsequently, the bounding box coordinates (x, y, width, height) of objects detected in each image and their corresponding class IDs are recorded in the `annotations` section. Finally, object class name and ID mapping information, such as 'implant', is defined in the `categories` section. This allows for the systematic linking of each image with object information.

[0087] The data annotation and preparation unit (200) also includes data augmentation operations. Data augmentation is a process of expanding a dataset by transforming given data, which helps the model perform more generalized learning in various situations. For example, by applying techniques such as random horizontal flipping, brightness adjustment, and contrast change, data is generated that can learn implant features in various environments. The augmentation operation increases the diversity of the data and enables the model to maintain high performance even under noise or abnormal conditions.

[0088] Additionally, the data annotation and preparation unit (200) divides the data set into a training set and a testing set. Generally, the data is split in a 2:1 ratio, the training data is used for the model to learn object detection, and the testing data is used to evaluate the model's performance. The splitting process is designed to prevent duplication and to ensure that no data bias occurs during the training and testing processes.

[0089] Finally, the data annotation and preparation unit (200) stores the generated data set in a form suitable for the deep learning model training pipeline and, if necessary, attaches additional metadata to increase the efficiency of model training. The above process maintains the quality and consistency of the data and supports the deep learning model in producing stable and reliable results.

[0090] In conclusion, the data annotation and preparation unit (200) accurately annotates implant object information and converts it into a dataset suitable for training a deep learning model to optimize the performance of the system. The data annotation and preparation unit (200) systematically manages a series of tasks ranging from annotation work to data augmentation and dataset splitting, and plays a role in laying the foundation of the data so that the deep learning analysis unit (210) can subsequently perform reliable learning and prediction.

[0091] FIG. 8 is a diagram showing the operation of a deep learning analysis unit according to an embodiment of the present invention.

[0092] Referring to FIG. 8, the deep learning analysis unit (210) is a core component of the present invention that automatically detects dental implants by analyzing multi-angle projection images generated based on CBCT data. The deep learning analysis unit (210) utilizes a Faster R-CNN model, a deep learning algorithm widely used in the field of object detection, to accurately detect the location of the implant and output it as structured data. The above process focuses on maximizing the learning and prediction performance of the model by deeply analyzing the characteristics of the data.

[0093] The deep learning analysis unit (210) first extracts deep features of the input projection image through a feature extraction network. In this step, a hierarchical feature map of the image is generated using a pre-trained convolutional neural network (CNN). This process is carried out through multiple layers of convolution and pooling operations, and each layer learns progressively complex features of the image. As a result, the generated feature map captures unique structural and visual information of the input image and is used as basic data for subsequent detection tasks.

[0094] After the feature map is generated, the Region Proposal Network (RPN) creates candidate regions likely to contain implants based on the feature map. The RPN uses a sliding window method to explore Regions of Interest (ROIs) on the feature map and predicts two pieces of information for each region. These two pieces of information are a binary classification score indicating whether the region contains an implant, and regression offset values ​​used to adjust the position of the bounding box. This step efficiently identifies candidate regions and eliminates unnecessary regions, thereby increasing the accuracy of the analysis.

[0095] Candidate regions generated by the region proposal network are passed to the object detection network. The network converts candidate regions of different sizes into fixed-size tensors through a Region of Interest Pooling layer. Subsequently, a fully connected layer classifies whether each ROI is an implant and determines the final bounding box location. In this step, a multi-task loss function is used to derive the final result of the object detection task. Cross-entropy is used for the classification loss, and Smooth L1 loss is used for the regression loss to optimize the model's performance.

[0096] Model training is performed based on the PyTorch architecture and enhances the model's generalization performance through data augmentation and triple cross-validation techniques. Data augmentation provides various input conditions through random horizontal flipping, brightness and contrast adjustments, etc., while triple cross-validation evaluates the reliability of the model by repeatedly splitting the dataset into training and validation sets. The above training process ensures that the model maintains consistent performance even in diverse environments.

[0097] The performance of the deep learning analysis unit (210) is verified through various evaluation indicators. In the experiment of the present invention, a precision of 0.96, a recall of 0.99, and an F1 score of 0.97 were achieved, and the average precision (AP) of object detection was recorded as 0.97 at an IoU (Intersection over Union) threshold of 0.5 and 0.738 at 0.75. The above results demonstrate that the deep learning analysis unit (210) provides very high accuracy and reliability in detecting implants in CBCT projection images.

[0098] In conclusion, the deep learning analysis unit (210) plays a central role in the dental implant detection system (100) that combines multi-angle CBCT projection and deep learning, and provides precise analysis results through three stages of feature extraction, region proposal, and object detection.

[0099] To explain Figure 8 in more detail, the attached technology shows the structure of an implant detection model such as Faster R-CNN and explains the entire process of detecting the location and class of implants by processing input images from a dataset step by step. This structure consists of a Dataset, Extractor, RPN (Region Proposal Network), ROI Head, and Bounding Box (BBox), and each step is organically connected to perform high-precision implant detection.

[0100] First, the dataset provides source images used for model training and validation, as well as annotation data containing object location and label information. The input image data (img) is used as input for the model, while the object labels and bounding box information (label) are utilized as ground truth during training. This data lays the foundation for the implant detection model.

[0101] Next, the Extractor performs the role of extracting features from the input image. In this step, a pre-trained Convolutional Neural Network (CNN) is typically used to process the input image and generate a feature map containing features ranging from low-level (lines, shapes, etc.) to high-level (complex patterns, forms). This feature map provides the visual information necessary for implant detection and is utilized in subsequent steps to predict the location and class of the implants.

[0102] The Region Proposal Network (RPN) proposes Regions of Interest (ROIs), which are areas likely to contain objects, based on the generated feature maps. The RPN generates small regions (Anchor Boxes) using a sliding window method and performs binary classification to determine the presence of implants and regression to predict bounding box coordinates for each region. Through this process, it selects candidate regions (ROIs) with a high probability of containing implants and provides bounding box information (BBox).

[0103] Subsequently, the ROI Head performs final implant classification and bounding box adjustment based on the candidate regions (ROIs) proposed by the RPN. First, the candidate regions are transformed into a fixed size using techniques such as ROI Pooling or ROI Align. Then, the ROI Head predicts the implant class for the transformed regions and further refines the bounding boxes. At this stage, the class label of the finally detected implant and the precise bounding box coordinates are output.

[0104] A BBox (Bounding Box) is a set of coordinate values ​​(x, y, width, height) representing the location of an object, used to visually represent the results of implant detection. The RPN and ROI Head predict and adjust the location of the implant based on this bounding box information to derive the final result.

[0105] In conclusion, the structure consists of providing input data from the dataset to accurately detect implants in input images, extracting features through the Extractor, proposing candidate regions in the RPN, and performing final classification and correction through the ROI Head. This stepwise approach enables high-accuracy implant detection.

[0106] FIG. 9 is a diagram exemplarily showing performance indicators evaluated by a performance evaluation unit according to one embodiment of the present invention.

[0107] Referring to FIG. 9, the performance evaluation unit (220) is a key component that verifies the accuracy and reliability of the results generated by the deep learning-based implant detection model. The performance evaluation unit (220) quantitatively measures the implant detection performance of the model, evaluates the overall efficiency of the system, and provides feedback for optimization. The performance evaluation unit (220) uses standardized evaluation indicators such as Precision, Recall, F1 score, and Average Precision (AP) to objectively verify the quality of the model.

[0108] First, the performance evaluation unit (220) measures precision and recall to evaluate how accurately the model has detected implants. Precision represents the ratio of implants that are actually correctly detected among those detected by the model and reflects the ability to reduce unnecessary false positives. On the other hand, recall represents the ratio of implants that are actually correctly detected by the model and evaluates the performance of minimizing false negatives.

[0109] Precision represents the proportion of implants detected by a model that are actually correctly identified, and it evaluates the ability to reduce unnecessary false positives. Models with high precision have high reliability in detection results because detected objects are more likely to be actual implants. For example, low precision can lead to problems such as misidentifying normal tissue as an implant, posing a significant risk of causing confusion during the diagnostic process. Therefore, precision is an essential metric for measuring how much a model can reduce false alarms.

[0110] Recall represents the proportion of actual implants correctly detected by the model and evaluates the model's performance in minimizing false negatives. Models with high recall are highly likely to detect actual implants without missing them, ensuring that clinically important information is not omitted. Conversely, models with low recall are prone to missing implant detections, posing a risk of missing critical information during the diagnostic process. In this regard, recall plays a crucial role in comprehensively understanding the detection performance of a model.

[0111] Precision and recall have a mutually complementary relationship, and the performance evaluation unit (220) evaluates the performance of the model by considering both indicators simultaneously. If precision is high and recall is low, the model has high reliability of detection results but may miss actual implants. Conversely, if recall is high and precision is low, the model attempts to detect all implants, but unnecessary false positives may increase. Therefore, the performance evaluation unit evaluates the balance between the two indicators to diagnose the overall performance of the model.

[0112] Meanwhile, in the performance test of the present invention, the precision was 0.96 and the recall was 0.99, recording very high values, which shows that the model achieved high accuracy and comprehensiveness simultaneously.

[0113] Next, the F1 score is the harmonic mean of precision and recall, and is used as an indicator to comprehensively evaluate the overall performance of the model. The F1 score emphasizes the balance between precision and recall; in the present invention, a high value of 0.97 was recorded, demonstrating that the model operates effectively without unnecessary detections or omissions. By emphasizing the balance between the two metrics, this score comprehensively measures how reliable the objects detected by the model are, while simultaneously ensuring that actual existing objects are detected without being missed. In particular, the F1 score provides an objective criterion for evaluating model performance without being biased toward either precision or recall.

[0114] The F1 score is defined by the following formula: F1 = 2 * ((Precision * Recall) / (Precision + Recall)). Here, precision represents the proportion of objects detected by the model that were actually correctly identified, while recall represents the proportion of actual objects that the model correctly detected. Since the harmonic mean emphasizes the balance between the two metrics, the F1 score decreases significantly if one metric is low. For example, if precision is 1.0 (100%) but recall is 0.5 (50%), the F1 score is only 0.67 (67%). This indicates an imbalance in model performance and clearly shows areas that require improvement.

[0115] In the present invention, the model recorded a high F1 score of 0.97. This means that precision and recall are balanced at a high level. For example, when precision is 96% (96% of detected implants are accurate) and recall is 99% (99% of actual implants are detected), the F1 score reaches 0.97. These results demonstrate that the model possesses the performance capability to detect important objects without missing them (high recall) while minimizing unnecessary detections (false positives).

[0116] Average Precision (AP) is used as a key metric to evaluate the detection results of a model based on a threshold. AP is calculated based on the Intersection over Union (IoU) threshold, which represents the degree of overlap between the detected bounding box and the actual bounding box. In this invention, AP@0.5 (AP at a threshold of 0.5) and AP@0.75 (AP at a threshold of 0.75) were measured, achieving values ​​of 0.97 and 0.738, respectively. These results demonstrate that the model maintains high accuracy under various overlap conditions. In particular, the performance at AP@0.75 proves that the model operates stably even under stricter conditions.

[0117] The performance evaluation unit (220) also ensures the reliability of the model evaluation through three-fold cross-validation. By dividing the dataset into three parts and using them alternately for training and validation, data bias is prevented and the generalization of results is strengthened. This validation method increases the likelihood that the model will perform the same way on new data and prevents the problem of overfitting.

[0118] The above validation method aims to prevent data bias and enhance the generalization of results by dividing the entire data into three parts and using them alternately for training and validation. This ensures that the model can achieve the same level of performance on new data and prevents the problem of overfitting.

[0119] Triple cross-validation divides the data into three parts, such as D1, D2, and D3, and alternately uses each as training data and validation data. For example, in the first step, D1 and D2 are used as training data and D3 is used as validation data; in the second step, D2 and D3 are used as training data and D1 is used as validation data. In the final step, D3 and D1 are used as training data and D2 is used as validation data. After going through this process three times, the evaluation metrics calculated in each step are averaged to evaluate the final model performance.

[0120] This method ensures that all data is used in equal proportions for training and validation, providing evaluation results that are not biased toward a specific dataset. Additionally, by alternating between training and validation data, it prevents overfitting, where the model becomes excessively optimized for specific data patterns. Consequently, triple cross-validation helps the model achieve consistent performance across diverse data and enhances the reliability of evaluation results.

[0121] Finally, the performance evaluation unit (220) analyzes the evaluation results and provides feedback for optimizing the deep learning model and the entire system. Performance limitations or problems discovered during the evaluation process are used to improve the model's structure, data processing method, or learning algorithm. For example, if detection failures under specific conditions are repeated, the performance evaluation unit (220) may suggest data augmentation or additional learning steps to resolve the conditions.

[0122] Although the invention has been described with reference to the above embodiments, those skilled in the art will understand that various modifications and changes can be made to the invention without departing from the spirit and scope of the invention as described in the following claims.

[0123] [Explanation of the symbol]

[0124] 100: Dental Implant Detection System Combining Multi-angle CBCT Projection and Deep Learning

[0125] 110: Input section

[0126] 120: Output section

[0127] 130: Communications Department

[0128] 140: Storage section

[0129] 150: Control unit

[0130] 160: Memory section

[0131] 170: CBCT Data Processing Unit

[0132] 180: MIP generation section

[0133] 190: Image Quality Optimization Section

[0134] 200: Data Annotations and Preparation

[0135] 210: Deep Learning Analysis Department

[0136] 220: Performance Evaluation Department

Claims

1. An input unit that receives input data from the outside into the system; An output unit that outputs result data to the outside of the system; A communication unit that transmits and receives data between the inside and outside of the system; A storage unit that stores data generated by the system; A dental implant detection system combining multi-angle CBCT projection and deep learning, comprising: a control unit for controlling the operation of the system; and a memory unit for executing various programs and storing data based thereon, The above memory unit is a CBCT (Cone Beam Computed Tomography) data processing unit that receives CBCT data and generates 3D volume data; A MIP generation unit that converts the above 3D volume data into multi-angle 2D projected images; An image quality optimization unit that improves the accuracy of analysis by optimizing the quality of the above 2D projected image; A data annotation and preparation unit that annotates implant information in the above 2D projection image and prepares deep learning training data; A dental implant detection system combining multi-angle CBCT projection and deep learning, comprising: a deep learning analysis unit that automatically detects implants based on the above-mentioned annotated implant information and a deep learning model; and a performance evaluation unit that quantitatively evaluates the performance of the deep learning model to verify the reliability of the results.

2. In Paragraph 1, The above CBCT data processing unit is a dental implant detection system that combines multi-angle CBCT projection and deep learning, utilizing a Python-based SimpleITK library and 3D Slicer software in the process of receiving the above CBCT (Cone Beam Computed Tomography) data and generating 3D volume data.

3. In Paragraph 1, A dental implant detection system combining multi-angle CBCT projection and deep learning, wherein the above-mentioned CBCT data processing unit extracts voxel intervals and origin positions from DICOM (Digital Imaging and Communications in Medicine) data and determines the resolution of the data and the data alignment criteria of the 3D coordinate system in a subsequent data processing step.

4. In Paragraph 3, A dental implant detection system combining deep learning with multi-angle CBCT projection, wherein the above DICOM (Digital Imaging and Communications in Medicine) data includes multiple slice images and metadata information regarding patient information, image resolution, scan area size, and scan interval.

5. In Paragraph 1, The above CBCT data processing unit is a dental implant detection system combining multi-angle CBCT projection and deep learning that improves the continuity of 3D volume data by reading DICOM (Digital Imaging and Communications in Medicine) data and performing linear interpolation to correct slice thickness imbalance.

6. In Paragraph 1, A dental implant detection system combining deep learning and multi-angle CBCT projection, in which the above CBCT data processing unit systematically organizes the density values ​​of each voxel before converting 3D volume data into a 2D projected image to accurately reflect the 3D structure.

7. In Paragraph 6, A dental implant detection system combining multi-angle CBCT projection and deep learning, wherein the density value of each of the above voxels is calculated based on the density values ​​of two adjacent slices.

8. In Paragraph 1, A dental implant detection system combining multi-angle CBCT projection and deep learning, wherein the MIP generation unit calculates a maximum density value in each projection direction based on the 3D volume data to generate the 2D projection image.

9. In Paragraph 1, The above image quality optimization unit is a dental implant detection system that combines multi-angle CBCT projection and deep learning to improve contrast by adjusting the display range of density values ​​through Hounsfield Unit (HU) correction.

10. In Paragraph 1, The above image quality optimization unit is a dental implant detection system combining multi-angle CBCT projection and deep learning that minimizes distortion caused by artifacts by performing voxel resampling and edge sharpening processing in the metal virtual image area.

11. In Paragraph 1, The above image quality optimization unit is a dental implant detection system combining multi-angle CBCT projection and deep learning that adjusts the overall contrast ratio by applying adaptive histogram equalization.

12. In Paragraph 1, The above image quality optimization unit is a dental implant detection system that combines multi-angle CBCT projection and deep learning to remove noise using Gaussian smoothing filtering technology.

13. In Paragraph 1, The above data annotation and preparation unit is a dental implant detection system combining multi-angle CBCT projection and deep learning that adds positional information of implants to each projected 2D image using the LabelImg annotation tool.

14. In Paragraph 13, A dental implant detection system combining deep learning and multi-angle CBCT projection, wherein the position information of the above implant includes the coordinate information (x, y, width, height) and class label information of the implant.

15. In Paragraph 13, The above data annotation and preparation unit is a dental implant detection system combining multi-angle CBCT projection and deep learning, which also includes a data augmentation operation that transforms data to expand the dataset.

16. In Paragraph 1, The above-mentioned deep learning analysis unit is a dental implant detection system combining multi-angle CBCT projection and deep learning that detects the location of an implant using a Faster R-CNN model and outputs it as structured data.

17. In Paragraph 16, The above-mentioned deep learning analysis unit performs model training based on the PyTorch architecture and is a dental implant detection system combining multi-angle CBCT projection and deep learning that improves the generalization performance of the model through data augmentation and triple cross-validation techniques.

18. In Paragraph 1, A dental implant detection system combining multi-angle CBCT projection and deep learning, wherein the above-mentioned performance evaluation unit quantitatively measures the implant detection performance of the model, evaluates efficiency, and provides feedback thereon.

Images