A Smart Annotation Method and System Based on Oral CBCT Images

Abstract

This invention discloses an intelligent annotation method and system based on oral CBCT images, belonging to the field of oral medical image processing technology. The method includes: acquiring CBCT voxel data of a target tooth region; performing sliding window analysis on the CBCT voxel data to generate local grayscale statistical features for each sliding window; for each sliding window, constructing a dual-modulation adaptive threshold function containing a primary threshold and a secondary threshold based on the local grayscale statistical features, and jointly judging the CBCT voxel data to form a preliminary tooth region. This solution, through the collaborative work of the primary and secondary thresholds and combined with dynamic adjustment of the judgment conditions based on grayscale distribution similarity, can effectively distinguish grayscale overlapping areas of enamel, dentin, alveolar bone, and other tissues. It solves the problems of fixed threshold segmentation in existing technologies being sensitive to local noise and unable to adapt to changes in grayscale distribution, significantly improving the accuracy and robustness of preliminary tooth region annotation.

Description

Technical Field

[0001] This invention belongs to the field of oral medical image processing technology, specifically relating to an intelligent annotation method and system based on oral CBCT images. Background Technology

[0002] In the field of oral medical imaging, with the development of cone-beam computed tomography (CBCT) technology, oral CBCT images have been widely used in the diagnosis, surgical planning, and restorative design of tooth structures. CBCT images can provide high-resolution three-dimensional information about the tooth structure and surrounding tissues. Compared with traditional two-dimensional X-ray images, they can more accurately reflect the position, shape, and interrelationships of teeth, providing fundamental data support for digital oral healthcare. However, in the process of implementing the technical solutions of the embodiments of this invention, it has been found that the above-mentioned technology has at least the following technical problems:

[0003] In existing technologies, automatic annotation methods for tooth regions are mainly based on threshold segmentation and region growth. Threshold segmentation methods typically classify voxels by setting a fixed threshold based on global or local gray values. Its advantages are simplicity and high computational efficiency. Region growth methods rely on a single growth strategy, but the fixed threshold setting is difficult to adaptively adjust for different tooth sizes, different scanning devices, or local noise conditions, leading to missed or false detections in the initial tooth region. Furthermore, a single growth strategy cannot simultaneously maintain high-confidence regions and explore low-confidence regions, which can easily cause region breakage or false expansion, resulting in low annotation accuracy of the final tooth annotation map. Summary of the Invention

[0004] This invention addresses the problems of fixed threshold segmentation in existing technologies being sensitive to local noise and unable to adapt to changes in grayscale distribution. It provides an intelligent annotation method based on oral CBCT images. By working collaboratively with primary and secondary thresholds and dynamically adjusting the judgment conditions based on grayscale distribution similarity, it can effectively distinguish grayscale overlapping areas of enamel, dentin, alveolar bone, and other tissues. This solves the problems of fixed threshold segmentation in existing technologies being sensitive to local noise and unable to adapt to changes in grayscale distribution, and significantly improves the accuracy and robustness of preliminary annotation of tooth regions.

[0005] According to one aspect of this specification, a smart annotation method based on oral CBCT images is provided, applied to the smart annotation of tooth regions in oral CBCT images, including:

[0006] Sliding window analysis was performed on the CBCT voxel data of the acquired target tooth region to generate local gray-scale statistical features for each sliding window;

[0007] Based on the local gray-scale statistical features of each sliding window, the CBCT voxel data are jointly judged by the constructed dual-modulation adaptive threshold function to form a preliminary tooth region.

[0008] A dual-path growth strategy is applied to the preliminary tooth region to generate a candidate voxel set; the multidimensional confidence vector of the candidate voxel set is calculated and weighted to obtain a growth confidence map; the preliminary tooth region is updated based on the growth confidence map.

[0009] Three-dimensional connectivity analysis was performed on the updated preliminary tooth region, and consistency assessment was conducted by combining local gray-level statistical features and growth confidence maps to generate boundary optimization mapping;

[0010] Based on local gray-scale statistical features, growth confidence map, and boundary optimization mapping, adjacent regions in the updated preliminary tooth region are merged and judged to generate a tooth annotation map.

[0011] Furthermore, local grayscale statistical features of each sliding window are generated, including:

[0012] The CBCT voxel data is divided into sliding windows of variable size, and the local gray-scale statistical features of each sliding window are calculated.

[0013] Furthermore, the initial tooth area is formed, including:

[0014] Based on the local gray-level statistical features of each sliding window, a dual-modulation adaptive threshold function containing a main threshold and a secondary threshold is constructed; wherein, the main threshold is a judgment criterion dynamically generated based on the voxel gray-level mean and variance within the sliding window; and the secondary threshold is an auxiliary judgment criterion generated based on the gray-level distribution pattern.

[0015] The local gray-level statistical features are compared with the predefined typical gray-level statistical feature distribution of tooth enamel to generate a gray-level distribution similarity score, which is then used as the weight of the dual-modulation adaptive threshold function.

[0016] Based on the primary threshold, secondary threshold, and grayscale distribution similarity score, the CBCT voxel data of each sliding window are jointly judged to form a preliminary tooth region.

[0017] Furthermore, after the initial tooth region is formed, the process also includes generating heterogeneity admission weights:

[0018] Based on the local gray-level statistical features of the sliding window corresponding to the CBCT voxel data, the gray-level gradient of adjacent voxels is calculated.

[0019] The similarity score, variance, and gray gradient of the gray-level distribution are weighted to generate heterogeneity admission weights.

[0020] Furthermore, a dual-path growth strategy is performed on the initial tooth region to generate a candidate voxel set, including:

[0021] A dual-path growth strategy, including conservative path growth and exploratory path growth, is executed in parallel on the preliminary tooth region to generate a candidate voxel set.

[0022] The conservative path growth accepts voxels with a heterogeneity admission weight greater than a preset conservative threshold; the exploratory path growth accepts voxels with a difference between the local gray-scale statistical features of the preliminary tooth region and a preset discrimination threshold.

[0023] Further, the multidimensional confidence vector of the candidate voxel set is calculated and weighted to obtain a growth confidence map. The preliminary tooth region is then updated based on the growth confidence map, including:

[0024] The candidate voxel set is judged from two aspects: conservative path growth and exploratory path growth, to obtain the accepted voxels;

[0025] A multidimensional confidence vector is calculated for the accepting voxel, and the multidimensional confidence vectors are weighted and synthesized to obtain the growth confidence value. All growth confidence values ​​are combined to obtain the growth confidence map.

[0026] Based on the growth confidence map and preset upper and lower thresholds, the voxels in the preliminary tooth region are iteratively upgraded and downgraded until convergence, generating an updated preliminary tooth region.

[0027] Furthermore, the candidate voxel set is evaluated from two aspects: conservative path growth and exploratory path growth, to obtain the accepting voxels, including:

[0028] The conservative path growth is judged by calculating the statistical similarity between each candidate voxel in the candidate voxel set and the nearest voxel in the initial tooth region, and combining the gray-scale distribution similarity score to form a conservative path score.

[0029] The determination is made from the perspective of exploration path growth: the exploration path score is calculated based on the lowest gray-level distribution similarity score on the shortest connected path from each candidate voxel in the candidate voxel set to the voxel in the initial tooth region;

[0030] The scores of the conservative path and the exploration path are weighted to obtain a comprehensive score.

[0031] If the overall score is greater than the preset screening threshold, then the corresponding candidate voxel is determined to be an accepted voxel.

[0032] Furthermore, the boundary optimization mapping is generated, including:

[0033] Three-dimensional connectivity analysis was performed on the updated preliminary tooth region to obtain connected candidate regions;

[0034] Statistical analysis was performed on the local gray-level statistical characteristics of voxels within each connected candidate region and the corresponding values ​​of the growth confidence map.

[0035] Based on the statistical analysis results, the connected candidate regions are divided into pure regions and mixed regions;

[0036] Cluster analysis is performed on voxels within the mixed region to generate boundary optimization mappings.

[0037] Further, a tooth annotation map is generated, including:

[0038] Based on each pair of adjacent regions in the updated preliminary tooth region, a multi-criteria merging judgment is performed to obtain a multi-criteria score;

[0039] If the multi-criteria score is greater than a preset merging threshold, the corresponding regions are merged to obtain the final tooth region, which is then used as the tooth annotation map.

[0040] According to one aspect of this specification, an intelligent annotation system based on oral CBCT images is provided, comprising:

[0041] The data acquisition and processing module is used to perform sliding window analysis on the acquired CBCT voxel data of the target tooth region and generate local grayscale statistical features for each sliding window.

[0042] The preliminary tooth region generation module is used to jointly determine the CBCT voxel data based on the local gray-scale statistical features of each sliding window and through the constructed dual-modulation adaptive threshold function to form a preliminary tooth region.

[0043] The preliminary tooth region update module is used to perform a dual-path growth strategy on the preliminary tooth region to generate a candidate voxel set; calculate the multidimensional confidence vector of the candidate voxel set and perform weighted processing to obtain a growth confidence map; and update the preliminary tooth region according to the growth confidence map.

[0044] The boundary optimization mapping module is used to perform three-dimensional connectivity analysis on the updated preliminary tooth region, and combine local gray-level statistical features and growth confidence map for consistency evaluation to generate boundary optimization mapping;

[0045] The tooth annotation map generation module is used to merge adjacent regions in the updated preliminary tooth region based on local gray-scale statistical features, growth confidence map, and boundary optimization mapping to generate a tooth annotation map.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. This invention, through the collaborative operation of primary and secondary thresholds and combined with the dynamic adjustment of gray-level distribution similarity judgment conditions, can effectively distinguish gray-level overlapping areas of enamel, dentin, alveolar bone, and other tissues. It solves the problems of fixed threshold segmentation in the prior art being sensitive to local noise and unable to adapt to changes in gray-level distribution, significantly improving the accuracy and robustness of the initial annotation of tooth regions, reducing the accumulation of errors in the subsequent regional growth stage, and laying the foundation for generating accurate tooth annotation maps.

[0048] 2. This invention employs a dual-path parallel mechanism, utilizing heterogeneity admission weights to ensure the stability of the core region while exploring ambiguous edge regions through statistical feature differences, achieving complementary expansion. This effectively solves the problem of regional breakage or erroneous expansion caused by a single growth strategy. By collaboratively screening candidate voxels through dual paths, the integrity and accuracy of the candidate set are significantly improved, providing a more reliable data foundation for subsequent confidence calculation and region optimization, thereby enhancing the boundary accuracy and structural coherence of the tooth annotation map. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart of an intelligent annotation method based on oral CBCT images provided in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of a module of an intelligent annotation system based on oral CBCT images provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Specifically, such as Figure 1As shown, this embodiment of the invention provides an intelligent annotation method based on oral CBCT images, including: acquiring CBCT voxel data of a target tooth region; performing sliding window analysis on the CBCT voxel data to generate local gray-level statistical features for each sliding window; for each sliding window, constructing a dual-modulation adaptive threshold function containing a main threshold and a secondary threshold based on the local gray-level statistical features, and jointly judging the CBCT voxel data to form a preliminary tooth region; executing a dual-path growth strategy including conservative path growth and exploratory path growth in parallel on the preliminary tooth region to generate a candidate voxel set; calculating the multidimensional confidence vector of the candidate voxel set and weighting it to obtain a growth confidence map, updating the preliminary tooth region according to the growth confidence map; performing three-dimensional connectivity analysis on the updated preliminary tooth region, and combining the local gray-level statistical features and the growth confidence map for consistency evaluation to generate a boundary optimization mapping; and merging adjacent regions in the updated preliminary tooth region according to the boundary optimization mapping, the growth confidence map, and the local gray-level statistical features to generate a tooth annotation map. Among them, sliding window analysis refers to dividing CBCT voxel data into multiple local windows and capturing local gray-scale distribution characteristics by calculating statistics within each window; dual-modulation adaptive threshold function refers to achieving dynamic segmentation determination through a combination of primary and secondary thresholds; dual-path growth strategy refers to simultaneously executing two growth modes: conservative path and exploratory path; and three-dimensional connectivity analysis refers to performing spatial connectivity detection on the updated region and evaluating regional consistency by combining statistical features and confidence levels.

[0054] Specifically, this embodiment of the invention extracts local grayscale statistical features through a sliding window, providing a data foundation for subsequent threshold calculation. Subsequently, the joint determination based on primary and secondary thresholds can adapt to grayscale variations in different local regions. For example, at the enamel-dentin junction, the secondary threshold can help correct segmentation deviations that may be caused by the primary threshold. In the dual-path growth strategy, the conservative path prioritizes retaining voxels with stable statistical features, while the exploratory path attempts to include candidate voxels with features similar to the initial region. Parallel execution of both avoids the limitations of a single strategy. A growth confidence map is generated through multi-dimensional confidence vector weighting, quantifying the reliability of each candidate voxel. For example, statistical consistency confidence reflects the degree of matching between the voxel and local features, edge fidelity confidence assesses its probability of being located at a structural boundary, and three-dimensional connectivity analysis further optimizes the region boundary, for example, by eliminating isolated noise points through connected component detection. The merging determination integrates adjacent regions based on multi-criteria scores, ultimately generating accurate tooth annotation results.

[0055] Specifically, this invention enhances the adaptability of the segmentation process by dynamically adjusting the threshold through local statistical features. Through parallel growth along two paths, it preserves high-confidence regions while exploring potential targets, effectively balancing the accuracy and completeness of segmentation. Furthermore, three-dimensional connectivity analysis and multi-criteria merging further optimize the continuity and consistency of region boundaries, solving the problems of region breakage or excessive merging in traditional methods. Through the above technical solutions, this invention can reduce missed or false detections caused by fixed thresholds, avoid region breakage or false expansion, and improve the spatial continuity and boundary accuracy of tooth annotation maps, thereby achieving more reliable automatic tooth region annotation in complex oral CBCT images.

[0056] Specifically, the sliding window analysis of CBCT voxel data, generating local grayscale statistical features for each sliding window, includes:

[0057] The CBCT voxel data were divided into sliding windows of variable size, and local gray-level statistical features were calculated for each sliding window. The local gray-level statistical features included local mean, variance, local skewness, and quantiles of the gray-level histogram.

[0058] Among them, the variable-size sliding window refers to the method of dividing the window into sections that dynamically adjusts the window size according to the spatial distribution characteristics of the target tooth region. Specifically, it can be implemented using a multi-level sampling method based on the image pyramid, such as setting the window size to different levels like 3×3×3, 5×5×5, and 7×7×7. The local mean refers to the arithmetic mean of the gray values ​​of all voxels within the sliding window. The variance refers to the dispersion of gray values ​​within the sliding window. The local skewness refers to the asymmetry of the gray value distribution within the sliding window. The quantile of the gray-level histogram refers to the value located at a specific percentage position after the gray values ​​within the sliding window are arranged in ascending order.

[0059] Specifically, in the processing of CBCT voxel data, the image space is traversed through a multi-scale sliding window. The coverage of each window is dynamically adjusted according to the current level. For each window, the gray values ​​of all voxels inside are first extracted, and the local mean and variance are calculated. The mean is used to eliminate noise interference, and the variance is used to identify high-contrast regions. Then, the local skewness is calculated to capture the asymmetric characteristics of the gray-level distribution. For example, the enamel region may show a right-skewed distribution. At the same time, a gray-level histogram is generated and quantiles are extracted. These statistical features together constitute a quantitative description of the local region, providing a multi-dimensional discrimination basis for subsequent threshold segmentation.

[0060] Specifically, embodiments of the present invention adapt to the morphological differences of different tooth sizes through multi-scale windows. For example, small windows capture details of the cusps, while large windows cover the overall structure of the root. At the same time, by combining the combined features of mean, variance, skewness, and quantiles, the local gray-scale distribution characteristics can be described more comprehensively. Through the above technical solution, this application can effectively solve the problem of missed detection of tooth structures caused by fixed window sizes. For example, it avoids blurring of the gap boundaries due to excessively large windows, or failing to cover the entire crown area due to excessively small windows. At the same time, the combined use of multi-dimensional statistical features improves the discrimination ability of local gray-scale features, thereby improving the accuracy of subsequent threshold segmentation.

[0061] Specifically, for each sliding window, a dual-modulation adaptive threshold function containing a main threshold and a secondary threshold is constructed based on local gray-level statistical features, and the CBCT voxel data is jointly judged to form a preliminary tooth region, which specifically includes:

[0062] For each sliding window, a dual-modulation adaptive threshold function containing a main threshold and a secondary threshold is constructed based on local gray-level statistical features, and its expression is as follows:

[0063] (1)

[0064] In the formula, This represents a dual-modulation adaptive threshold function. This represents the score for the similarity of grayscale distributions. Indicates the main threshold. Indicates the secondary threshold. This represents the weighting coefficient; the principal threshold is obtained by mapping the local mean to the variance, and its specific calculation formula is as follows:

[0065] (2)

[0066] In the formula, Represents the local mean. Represents variance. The adjustment coefficient representing the local mean. The adjustment factor representing variance;

[0067] The secondary threshold is generated by the difference between skewness and quantile.

[0068] The local gray-level statistical features are compared with the predefined typical gray-level statistical feature distribution of tooth enamel to generate a gray-level distribution similarity score. The specific calculation formula is as follows:

[0069] (3)

[0070] In the formula, This represents the total number of gray levels in the histogram. Indicates the number of elements in the current sliding window. grayscale histogram values, Indicates typical tooth enamel. The grayscale histogram values ​​of each level are used as weights for the dual-modulation adaptive threshold function.

[0071] For each sliding window, CBCT voxel data are jointly judged based on the main threshold, secondary threshold, and gray-level distribution similarity score to form a preliminary tooth region. The main threshold is a judgment criterion dynamically generated based on the mean and variance of voxel gray levels within the sliding window; the secondary threshold is an auxiliary judgment criterion generated based on the gray-level distribution morphology; and the gray-level distribution similarity score is a quantitative index generated by comparing the probability distribution differences between the current window's gray-level statistical characteristics and typical enamel characteristics.

[0072] Specifically, in the sliding window analysis stage, after the local gray-level statistical features are extracted, the primary threshold is generated by inputting the mean and variance into a preset mapping function, and the secondary threshold is generated by the difference between skewness and quantiles. Simultaneously, the gray-level histogram of the current window is compared with a pre-stored typical enamel gray-level distribution template to obtain a score between 0 and 1. This score is used to weight and fuse the primary and secondary thresholds; for example, when the score is high, the primary threshold dominates; when the score is low, the weight of the secondary threshold increases. Finally, a joint decision function is used to perform binary classification on each voxel within the window, marking voxels that meet the primary or secondary threshold conditions as candidate tooth regions, forming the preliminary segmentation result.

[0073] Specifically, the embodiments of the present invention, through the collaborative operation of primary and secondary thresholds and the dynamic adjustment of the judgment conditions based on gray-level distribution similarity, can effectively distinguish gray-level overlapping areas of enamel, dentin, alveolar bone, and other tissues. Through the above technical solution, this application solves the problems of fixed threshold segmentation in the prior art being sensitive to local noise and unable to adapt to changes in gray-level distribution, significantly improving the accuracy and robustness of the initial annotation of tooth regions. Through the dual-threshold joint judgment mechanism, stable segmentation performance can be maintained in high-noise or uneven gray-level regions, reducing the accumulation of errors in the subsequent region growth stage, and laying the foundation for generating accurate tooth annotation maps.

[0074] Specifically, after the initial tooth region is formed, the process also includes generating heterogeneity admission weights, which specifically include:

[0075] The local gray-level statistical features of the corresponding sliding window of the CBCT voxel data are read, and the gray-level gradient of adjacent voxels is calculated. The specific calculation formula is as follows:

[0076] (4)

[0077] In the formula, Voxel representation and adjacent voxels grayscale gradient, Voxel representation grayscale value, Indicates adjacent voxels grayscale value;

[0078] The similarity score of gray-level distribution, variance, and gray-level gradient are weighted to generate heterogeneous admission weights. Local gray-level statistical features refer to the statistical attributes of voxel gray-level values ​​within a sliding window, which can be calculated using local mean, variance, skewness, and quantiles. Gray-level gradient refers to the rate of gray-level change between adjacent voxels. Weighting involves linearly combining multiple feature parameters according to preset weight coefficients.

[0079] Specifically, when generating heterogeneity admission weights, the gray-scale distribution characteristics of the current region are first obtained based on the local gray-scale statistical features of the sliding window. For example, local variance can be used to measure the dispersion of gray-scale distribution. Then, by calculating the gray-scale gradient between adjacent voxels, possible structural boundaries or noise interference regions are identified. Furthermore, the gray-scale distribution similarity score is weighted and fused with local variance and gray-scale gradient. For example, the gray-scale distribution similarity score can reflect the degree of matching between the current region and a typical enamel region, variance can characterize the stability of gray-scale distribution, and gray-scale gradient can indicate the degree of drastic structural changes. By pre-setting weight coefficients, the three are weighted and summed to obtain the heterogeneity admission weight value.

[0080] Specifically, this invention integrates three features—grayscale distribution similarity score, variance, and gradient—to dynamically assess the admission probability of voxels in heterogeneous regions. Through the above technical solution, this invention can effectively distinguish between real tooth regions and noise interference regions. By comprehensively considering the stability of grayscale distribution and the intensity of gradient changes, it suppresses erroneous growth caused by local grayscale mutations. At the same time, the dynamic calculation of heterogeneity admission weights provides a differentiated judgment basis for subsequent dual-path growth strategies, thereby improving the integrity and boundary accuracy of tooth annotation.

[0081] Specifically, a dual-path growth strategy, encompassing both conservative and exploratory path growth, is executed in parallel on the initial tooth region to generate a candidate voxel set, including:

[0082] A dual-path growth strategy, comprising conservative path growth and exploratory path growth, is executed in parallel on the preliminary tooth region to generate a candidate voxel set. Conservative path growth accepts voxels with a heterogeneity admission weight greater than a preset conservative threshold; exploratory path growth accepts voxels whose difference from the local gray-level statistical features of the preliminary tooth region is less than a preset discrimination threshold. Conservative path growth refers to a stability expansion strategy based on heterogeneity admission weights; exploratory path growth refers to an expansion strategy based on statistical feature similarity; and heterogeneity admission weights are dynamic weight parameters generated by combining gray-level distribution similarity, local variance, and gray-level gradient.

[0083] Specifically, when executing the dual-path growth strategy in parallel, the conservative path growth first selects voxels that meet the conservative threshold based on the heterogeneity admission weight. At the same time, the exploratory path growth runs independently. By calculating the difference in statistical features between the candidate voxels and the preliminary tooth region, for example, if the absolute difference in the gray-scale mean is less than 0.3, it is determined as a potential expansion region and added to the candidate set. The candidate voxel sets generated by the two paths are merged through set operations to form a comprehensive candidate set covering the high-confidence region and the potential expansion region.

[0084] This invention employs a dual-path parallel mechanism, utilizing heterogeneity admission weights to ensure the stability of the core region while exploring ambiguous edge regions through statistical feature differences, achieving complementary expansion. Through the above technical solution, this application effectively solves the problem of regional fragmentation or erroneous expansion caused by a single growth strategy. By collaboratively screening candidate voxels through dual paths, the integrity and accuracy of the candidate set are significantly improved, providing a more reliable data foundation for subsequent confidence calculation and region optimization, thereby improving the boundary accuracy and structural coherence of the tooth annotation map.

[0085] Specifically, the multidimensional confidence vector of the candidate voxel set is calculated and weighted to obtain a growth confidence map. The initial tooth region is then updated based on this growth confidence map, including:

[0086] The candidate voxel set is judged from two aspects: conservative path growth and exploratory path growth, to obtain the accepted voxels.

[0087] Calculate multidimensional confidence vectors for the acceptance voxels; the multidimensional confidence vectors include statistical consistency confidence, structural connectivity confidence, and edge fidelity confidence;

[0088] The growth confidence value is obtained by weighted synthesis of the multidimensional confidence vectors, and the growth confidence map is obtained by combining all the growth confidence values.

[0089] Based on the growth confidence map and preset upper and lower thresholds, the voxels in the preliminary tooth region are iteratively upgraded and downgraded until convergence, generating an updated preliminary tooth region.

[0090] Among them, statistical consistency confidence refers to the degree of matching between candidate voxels and preliminary tooth regions in terms of gray-level statistical features. Specifically, it can be achieved by calculating the similarity difference between candidate voxels and adjacent regions in terms of gray-level mean and variance.

[0091] Structural connectivity confidence refers to the connectivity strength between candidate voxels and the preliminary tooth region in three-dimensional space, which can be achieved by shortest path analysis or by counting the number of connected voxels in the neighborhood.

[0092] Marginal fidelity confidence refers to the probability that a candidate voxel is located at the boundary of a tooth region, which can be achieved by gradient magnitude calculation or edge response function evaluation.

[0093] Weighted composition refers to the linear combination of multidimensional confidence vectors by assigning different weights to them;

[0094] Iterative upgrading and downgrading refers to dynamically adjusting the voxel state based on the comparison between the growth confidence value and the threshold. Specifically, it can be achieved by upgrading the voxel to a defined region when the confidence value is higher than the upper threshold and downgrading it to an excluded region when the confidence value is lower than the lower threshold.

[0095] Specifically, the candidate voxel set is obtained through two strategies: conservative path growth and exploratory path growth. The conservative path focuses on preserving high-confidence regions, while the exploratory path focuses on expanding potential regions. When performing multidimensional confidence calculations on candidate voxels, statistical consistency confidence is generated by comparing the gray mean and variance differences between the candidate voxel and its neighboring regions. Structural connectivity confidence is generated by analyzing the connectivity strength on the shortest path from the candidate voxel to the nearest determined region. Edge fidelity confidence is generated by the response of gradient operators or edge detection filters. During the weighted synthesis process, the weights can be dynamically adjusted according to the gray distribution characteristics of different regions. During the iterative upgrading and downgrading process, the upper and lower thresholds can be set to dynamic ranges to ensure that each iteration only adjusts voxels with critical confidence levels until the state of all voxels no longer changes.

[0096] This scheme, through comprehensive evaluation of multi-dimensional confidence vectors and combined with dynamic weight adjustment and iterative state update mechanisms, not only retains the precise control of high-confidence regions by conservative paths but also utilizes exploratory paths to expand potential regions. Through the above technical solutions, this application effectively solves the problem of region fragmentation or misexpansion caused by a single growth strategy in the prior art. By using multi-dimensional confidence evaluation and dynamic iteration mechanisms, it can accurately distinguish between tooth regions and non-target tissues under complex grayscale distributions and noise interference, improving the continuity and boundary clarity of the annotation results.

[0097] Specifically, the candidate voxel set is judged from two aspects: conservative path growth and exploratory path growth, and the accepted voxels include:

[0098] Conservative path growth determination: For each candidate voxel in the candidate voxel set, the statistical similarity between it and the nearest voxel in the initial tooth region is calculated. The specific calculation formula is as follows:

[0099] (5)

[0100] In the formula, Indicates candidate voxels The nearest voxel in the initial tooth region Statistical similarity Indicates candidate voxels The average gray value of the neighborhood. Indicates the nearest voxel in the initial tooth region The average gray value of the neighborhood. Indicates candidate voxels The standard deviation of gray level in the neighborhood. Indicates the nearest voxel in the initial tooth region The standard deviation of gray level in the neighborhood. The normalization parameter representing the difference between the means. The normalized parameter representing the difference in standard deviation is combined with the gray distribution similarity score to form a conservative path score.

[0101] Exploration path growth determination: For each candidate voxel in the candidate voxel set, the exploration path score is calculated based on the lowest gray-level distribution similarity score on the shortest connected path from the voxel in the initial tooth region to that candidate voxel. The specific calculation formula is as follows:

[0102] (6)

[0103] In the formula, Indicates candidate voxels The exploration path score, This represents any voxel within the initial tooth region. This indicates the initial voxel assembly of the tooth region. Representing a path A certain voxel on, Voxel representation The local gray-level feature vector, Voxel representation The local gray-level feature vector, Voxel representation and voxels The grayscale distribution similarity score.

[0104] The conservative path score and the exploratory path score are weighted to obtain a comprehensive score. The comprehensive score is then used to determine if it exceeds a preset screening threshold; if so, the corresponding candidate voxel is accepted. The conservative path score is a weighted value obtained by combining the statistical similarity and gray-level distribution similarity scores between the candidate voxel and the initial tooth region voxels. Specifically, this can be achieved by calculating the difference in gray-level mean and variance ratio, combined with the overlap of the gray-level distribution histogram within the sliding window. The exploratory path score is the minimum gray-level distribution similarity score based on the shortest connected path between the candidate voxel and the initial tooth region. This can be achieved using a 3D adjacency traversal algorithm and a minimum path score extraction algorithm. The comprehensive score is a value obtained by linearly combining the conservative path score and the exploratory path score with preset weights. This can be achieved using empirical weights or dynamically adjusted coefficients based on training data. The screening threshold is the critical value used to determine whether a candidate voxel is accepted. This threshold can be obtained through statistical analysis of the score distribution of correctly labeled voxels in historical data.

[0105] Specifically, candidate voxels are first subjected to dual-path determination. In the conservative path determination, the difference in mean grayscale value and variance ratio between the candidate voxel and the nearest preliminary region voxel are calculated. Combined with the similarity score between the grayscale histogram within the sliding window and the typical enamel distribution, a conservative path score is generated using a weighted formula. In the exploratory path determination, a three-dimensional connected path from the candidate voxel to the preliminary region is constructed. The grayscale distribution similarity score of all voxels on the path is extracted, and the minimum value is taken as the exploratory path score. Subsequently, the conservative score and the exploratory score are weighted and summed according to a preset ratio. If the result exceeds the screening threshold, the candidate voxel is marked as an acceptable voxel. The entire process is implemented through a parallel computing framework to ensure determination efficiency.

[0106] This solution employs a dual-path collaborative decision-making mechanism to maintain the stability of high-confidence regions in the conservative path and to uncover the rationality of potential connected regions in the exploratory path. The weighted combination of the two approaches forms a more reliable decision-making basis, effectively overcoming the limitations of a single strategy. Through the above technical solution, this application solves the problem of region breakage or erroneous expansion caused by a single growth strategy in the prior art. By balancing conservatism and exploratoryness through the dual-path collaborative decision-making mechanism, it accurately identifies low-contrast regions with potential connectivity while maintaining the integrity of high-confidence regions, significantly improving the boundary accuracy and regional continuity of tooth annotation.

[0107] Specifically, a three-dimensional connectivity analysis is performed on the updated preliminary tooth region, and a consistency assessment is conducted by combining local gray-level statistical features and growth confidence maps. The generated boundary optimization mapping specifically includes:

[0108] Three-dimensional connectivity analysis was performed on the updated preliminary tooth region to obtain connected candidate regions;

[0109] Statistical analysis was performed on the local gray-level statistical characteristics of voxels within each connected candidate region and the corresponding values ​​of the growth confidence map.

[0110] Based on the statistical analysis results, the connected candidate regions are divided into pure regions and mixed regions;

[0111] Cluster analysis is performed on voxels within the mixed region to generate boundary optimization mappings.

[0112] Among them, three-dimensional connectivity analysis refers to identifying interconnected sets of voxels through three-dimensional spatial neighborhood relationships, which can be implemented using region growing algorithms or graph theory-based connected component labeling algorithms; local gray-level statistical features refer to statistical quantities such as gray-level mean, variance, skewness, and quantiles extracted from the sliding window; growth confidence map refers to a voxel-level confidence distribution map generated by weighted multidimensional confidence vectors; statistical analysis refers to calculating the mean, variance, and distribution pattern of the local gray-level statistical features and growth confidence values ​​of all voxels in the connected candidate region; pure region refers to a region whose local gray-level statistical features highly match the typical enamel distribution and whose growth confidence values ​​are stable; mixed region refers to a region with significant gray-level heterogeneity or growth confidence fluctuations; cluster analysis refers to grouping voxels in the mixed region based on multidimensional data of local gray-level statistical features and growth confidence values.

[0113] Specifically, in the three-dimensional connectivity analysis stage, firstly, by traversing the three-dimensional neighborhood relationships of all voxels, voxels that are spatially adjacent and belong to the same preliminary tooth region are grouped into connected candidate regions. Subsequently, for each connected candidate region, the distribution characteristics of the local gray-level mean, variance, and growth confidence value of all voxels within it are calculated and compared with the preset typical statistical intervals of enamel. Regions with gray-level mean within the typical interval and variance below the threshold are identified as pure regions; regions with gray-level distribution deviations or excessively high variances are identified as mixed regions. For voxels within mixed regions, their local gray-level statistical features and growth confidence values ​​are further extracted to form multi-dimensional feature vectors. These vectors are then divided into different subclasses using a clustering algorithm. Boundary optimization mappings are generated based on the feature differences between subclasses, thereby accurately distinguishing the transition regions between teeth and adjacent tissues.

[0114] This solution, through three-dimensional connectivity analysis combined with multi-dimensional feature clustering, can accurately identify the distribution of heterogeneous voxels within mixed regions, avoiding misjudgments caused by local noise or grayscale gradations. Statistical analysis further subdivides the region into pure and mixed types, allowing for targeted optimization of processing strategies and significantly improving boundary localization accuracy. Through the above technical solutions, this application effectively addresses the boundary ambiguity problem in the processing of complex three-dimensional structures using existing methods. Cluster analysis accurately separates different tissue types within mixed regions, reducing misjudgments between teeth and periodontal tissues. Simultaneously, the pure region determination mechanism based on statistical analysis preserves the structural integrity of high-confidence regions, avoiding over-segmentation and thus generating more accurate tooth annotation results.

[0115] Specifically, based on boundary optimization mapping, growth confidence map, and local gray-level statistical features, adjacent regions in the updated preliminary tooth region are merged and judged to generate a tooth annotation map, which specifically includes:

[0116] For each pair of adjacent regions in the updated preliminary tooth region, a merging determination is made based on multiple criteria;

[0117] The multi-criteria score is calculated by using boundary optimization mapping, growth confidence map, and local gray-scale statistical features to obtain statistical compatibility score, structural connectivity score, and boundary inhibition score. The specific calculation formula is as follows:

[0118] (7)

[0119] In the formula, Indicates adjacent regions and adjacent areas Multi-criteria scoring This represents the statistical compatibility score. The weighting factors represent the statistical compatibility scores. The structural connectivity score is represented by... The weighting factor represents the structural connectivity score. Indicates the boundary suppression score. The weighting factor represents the boundary suppression score.

[0120] If the multi-criteria score exceeds a preset merging threshold, the corresponding regions are merged to obtain the final tooth region, which is then used as the tooth annotation map. The statistical compatibility score is a quantitative indicator generated by comparing the consistency of local gray-scale statistical feature distributions of adjacent regions; the structural connectivity score is a path reliability indicator generated based on the growth confidence map and 3D connectivity analysis results; the boundary suppression score is a suppression factor generated using voxel classification results in boundary optimization mapping; and the multi-criteria score is a comprehensive score generated by a weighted linear combination of the statistical compatibility score, structural connectivity score, and boundary suppression score.

[0121] In the merging decision-making process, local gray-scale statistical features of adjacent regions are first extracted, such as mean, variance, and quantiles, and statistical compatibility scores are calculated to determine whether the gray-scale distributions match. Subsequently, based on the growth confidence map and the results of three-dimensional connectivity analysis, confidence values ​​on the shortest connected paths between regions are extracted to generate structural connectivity scores to assess path reliability. At the same time, based on the clustering results of mixed regions in the boundary optimization mapping, boundary suppression scores are calculated to reflect the ambiguity of the boundary regions. Finally, the three scores are weighted and summed according to preset weights. If the result exceeds the merging threshold, region merging is performed to ensure the accuracy of the merging decision.

[0122] This solution effectively avoids the problem of incorrect or missed merging caused by the limitations of a single criterion by integrating the evaluation of three dimensions: statistical compatibility, structural connectivity, and boundary suppression. Through the above technical solution, this application can significantly improve the accuracy of tooth region merging determination and reduce the phenomenon of over-merging or under-merging caused by a single criterion. Through a multi-dimensional evaluation mechanism, while ensuring the consistency of grayscale features, it strengthens the structural connectivity reliability and boundary clarity constraints. The final generated tooth annotation map has higher annotation accuracy and regional integrity in complex boundary regions and noisy interference scenarios.

[0123] The implementation of the various embodiments of the present invention is based on programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide an intelligent annotation system based on oral CBCT images, which is used to execute an intelligent annotation method based on oral CBCT images from the above method embodiments.

[0124] like Figure 2As shown, the system includes: a data acquisition and processing module, used to perform sliding window analysis on the acquired CBCT voxel data of the target tooth region to generate local gray-level statistical features for each sliding window; a preliminary tooth region generation module, used to jointly determine the CBCT voxel data based on the local gray-level statistical features of each sliding window using a constructed dual-modulation adaptive threshold function to form a preliminary tooth region; a preliminary tooth region update module, used to execute a dual-path growth strategy on the preliminary tooth region to generate a candidate voxel set; calculate the multidimensional confidence vector of the candidate voxel set and perform weighted processing to obtain a growth confidence map, and update the preliminary tooth region according to the growth confidence map; a boundary optimization mapping module, used to perform three-dimensional connectivity analysis on the updated preliminary tooth region, and combine the local gray-level statistical features and the growth confidence map for consistency evaluation to generate a boundary optimization mapping; and a tooth annotation map generation module, used to merge adjacent regions in the updated preliminary tooth region according to the local gray-level statistical features, the growth confidence map, and the boundary optimization mapping to generate a tooth annotation map.

[0125] The intelligent annotation system based on oral CBCT images provided in this invention addresses the problems of fixed threshold segmentation in existing technologies being sensitive to local noise and unable to adapt to changes in grayscale distribution. By employing several modules and working collaboratively with primary and secondary thresholds, combined with dynamic adjustment of judgment conditions based on grayscale distribution similarity, the system can effectively distinguish grayscale overlapping areas of enamel, dentin, alveolar bone, and other tissues. This solves the problems of fixed threshold segmentation in existing technologies being sensitive to local noise and unable to adapt to changes in grayscale distribution, and significantly improves the accuracy and robustness of preliminary annotation of tooth regions.

[0126] Finally, it should be noted that the above specific embodiments are merely illustrative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and many variations are possible. Any simple modifications, equivalent changes, and alterations made to the above specific embodiments based on the technical essence of the present invention should be considered within the protection scope of the present invention.

Claims

1. A smart annotation method based on oral CBCT images, applied to the smart annotation of tooth regions in oral CBCT images, characterized in that, include: Sliding window analysis was performed on the CBCT voxel data of the acquired target tooth region to generate local gray-scale statistical features for each sliding window; Generate local grayscale statistical features for each sliding window, including: The CBCT voxel data is divided into sliding windows of variable size, and the local gray-scale statistical features of each sliding window are calculated. Based on the local grayscale statistical features of each sliding window, the CBCT voxel data are jointly judged using a constructed dual-modulation adaptive threshold function to form a preliminary tooth region; the preliminary tooth region includes: Based on the local gray-level statistical features of each sliding window, a dual-modulation adaptive threshold function containing a main threshold and a secondary threshold is constructed; wherein, the main threshold is a judgment criterion dynamically generated based on the voxel gray-level mean and variance within the sliding window; and the secondary threshold is an auxiliary judgment criterion generated based on the gray-level distribution pattern. The local gray-level statistical features are compared with the predefined typical gray-level statistical feature distribution of tooth enamel to generate a gray-level distribution similarity score, which is then used as the weight of the dual-modulation adaptive threshold function. Based on the primary threshold, secondary threshold, and grayscale distribution similarity score, the CBCT voxel data of each sliding window are jointly evaluated to form a preliminary tooth region. After forming the preliminary tooth region, the process also includes generating heterogeneity admission weights. Based on the local gray-level statistical features of the sliding window corresponding to the CBCT voxel data, the gray-level gradient of adjacent voxels is calculated. The similarity score, variance, and gray gradient of the gray-level distribution are weighted to generate heterogeneous admission weights. A dual-path growth strategy is applied to the preliminary tooth region to generate a candidate voxel set; the multidimensional confidence vector of the candidate voxel set is calculated and weighted to obtain a growth confidence map; the preliminary tooth region is updated based on the growth confidence map; the dual-path growth strategy is applied to the preliminary tooth region to generate the candidate voxel set, including: A dual-path growth strategy, including conservative path growth and exploratory path growth, is executed in parallel on the preliminary tooth region to generate a candidate voxel set. The conservative path growth accepts voxels with a heterogeneity admission weight greater than a preset conservative threshold; the exploratory path growth accepts voxels with a difference between the local gray-scale statistical features of the preliminary tooth region and a preset discrimination threshold. Three-dimensional connectivity analysis was performed on the updated preliminary tooth region, and consistency assessment was conducted by combining local gray-level statistical features and growth confidence maps to generate boundary optimization mapping; Based on local gray-scale statistical features, growth confidence map, and boundary optimization mapping, adjacent regions in the updated preliminary tooth region are merged and judged to generate a tooth annotation map.

2. The intelligent annotation method based on oral CBCT images according to claim 1, characterized in that, The multidimensional confidence vector of the candidate voxel set is calculated and weighted to obtain a growth confidence map. The preliminary tooth region is then updated based on the growth confidence map, including: The candidate voxel set is judged from two aspects: conservative path growth and exploratory path growth, to obtain the accepted voxels; A multidimensional confidence vector is calculated for the accepting voxel, and the multidimensional confidence vectors are weighted and synthesized to obtain the growth confidence value. All growth confidence values ​​are combined to obtain the growth confidence map. Based on the growth confidence map and preset upper and lower thresholds, the voxels in the preliminary tooth region are iteratively upgraded and downgraded until convergence, generating an updated preliminary tooth region.

3. The intelligent annotation method based on oral CBCT images according to claim 2, characterized in that, The candidate voxel set is evaluated based on both conservative path growth and exploratory path growth to obtain the accepted voxels, including: The conservative path growth is judged by calculating the statistical similarity between each candidate voxel in the candidate voxel set and the nearest voxel in the initial tooth region, and combining the gray-scale distribution similarity score to form a conservative path score. The determination is made from the perspective of exploration path growth: the exploration path score is calculated based on the lowest gray-level distribution similarity score on the shortest connected path from each candidate voxel in the candidate voxel set to the voxel in the initial tooth region; The scores of the conservative path and the exploration path are weighted to obtain a comprehensive score. If the overall score is greater than the preset screening threshold, then the corresponding candidate voxel is determined to be an accepted voxel.

4. The intelligent annotation method based on oral CBCT images according to claim 1, characterized in that, Generate boundary optimization mappings, including: Three-dimensional connectivity analysis was performed on the updated preliminary tooth region to obtain connected candidate regions; Statistical analysis was performed on the local gray-level statistical characteristics of voxels within each connected candidate region and the corresponding values ​​of the growth confidence map. Based on the statistical analysis results, the connected candidate regions are divided into pure regions and mixed regions; Cluster analysis is performed on voxels within the mixed region to generate boundary optimization mappings.

5. The intelligent annotation method based on oral CBCT images according to claim 1, characterized in that, Generate a tooth annotation map, including: Based on each pair of adjacent regions in the updated preliminary tooth region, a multi-criteria merging judgment is performed to obtain a multi-criteria score; If the multi-criteria score is greater than a preset merging threshold, the corresponding regions are merged to obtain the final tooth region, which is then used as the tooth annotation map.

6. An intelligent annotation system based on oral CBCT images, characterized in that, To implement the intelligent annotation method based on oral CBCT images as described in any one of claims 1-5, the method includes: The data acquisition and processing module is used to perform sliding window analysis on the acquired CBCT voxel data of the target tooth region and generate local grayscale statistical features for each sliding window. The preliminary tooth region generation module is used to jointly determine the CBCT voxel data based on the local gray-scale statistical features of each sliding window and through the constructed dual-modulation adaptive threshold function to form a preliminary tooth region. The preliminary tooth region update module is used to perform a dual-path growth strategy on the preliminary tooth region to generate a candidate voxel set; calculate the multidimensional confidence vector of the candidate voxel set and perform weighted processing to obtain a growth confidence map; and update the preliminary tooth region according to the growth confidence map. The boundary optimization mapping module is used to perform three-dimensional connectivity analysis on the updated preliminary tooth region, and combine local gray-level statistical features and growth confidence map for consistency evaluation to generate boundary optimization mapping; The tooth annotation map generation module is used to merge adjacent regions in the updated preliminary tooth region based on local gray-scale statistical features, growth confidence map, and boundary optimization mapping to generate a tooth annotation map.

Images