A multi-scale feature image semantic segmentation method for alveolar bone defect detection

Abstract

The present application relates to the technical field of medical image processing, in particular to a multi-scale feature image semantic segmentation method for alveolar bone defect detection. The present application converts the cone-beam computed tomography tensor containing artifacts to the frequency domain, applies a directional band-pass mask to filter out noise, and performs nonlinear truncation on the attenuation coefficient overlap area in the spatial domain to amplify the gray scale gradient; multi-scale spatial features are extracted using a feature network, and dynamic weighting is performed based on the anisotropic local gradient variance; in the optimization stage, an asymmetric weighted three-dimensional geometric deviation penalty term is introduced to force the decoupling of the connected boundaries, and the segmentation grid is output; finally, the local morphological principal axis of a single tooth is fitted by segmentation, and projection calculation is performed along the principal axis between the structure contour lines, effectively improving the three-dimensional segmentation accuracy and morphological fidelity of low-contrast microstructure.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a multi-scale feature image semantic segmentation method for alveolar bone defect detection. Background Technology

[0002] Oral health is a vital component of overall health, and obtaining information on the structural composition of teeth and surrounding supporting tissues (such as alveolar bone) is crucial for disease diagnosis and surgical simulation. Cone-beam computed tomography (CBCT), with its isotropic voxels and high spatial resolution, has become a core digital examination tool in dentistry. In orthodontic and periodontal treatment, precise segmentation of the tooth and alveolar bone structures, especially the identification of alveolar bone defects induced by inflammation (such as bone fissures and fenestrations), is essential for assessing the safe range of root movement and developing surgical plans.

[0003] Currently, tooth and alveolar bone segmentation in CBCT images primarily employs 3D convolutional neural networks or visual Transformer architectures. However, existing technologies still face significant technical obstacles when processing fine tissue structures with low-contrast boundary features. First, the thickness of thin alveolar bone regions (such as the labial and buccal bone plates) is often less than one millimeter. The downsampling mechanism of traditional neural networks causes the high-frequency spatial signals of these minute bone defects to be submerged by the signals of surrounding normal tissue, resulting in irreversible information loss. This makes the reconstructed segmentation model unable to reproduce key clinical microscopic pathological features. Second, the X-ray quality attenuation coefficients at the junction of tooth roots and alveolar bone are extremely similar, leading to extremely low contrast at tissue boundaries in imaging. Existing regional loss functions do not penalize misclassification of edge pixels sufficiently, easily causing severe pixel-level adhesion and morphological distortion at the tooth-bone interface. Furthermore, medical panoramic data suffers from extreme class imbalance, with background voxels often exceeding 90%. Traditional convolution operators lack effective encoding of long-range topological dependencies of tissue structures, resulting in illogical holes or fragments in the predicted 3D models. Meanwhile, the massive amount of 3D data and the complex, gigantic network architecture place extremely high demands on the video memory and computing power of clinical workstations, limiting the real-time deployment of high-precision algorithms at the grassroots level. In summary, existing technologies struggle to simultaneously achieve fidelity in capturing subtle pathological features and accurately decoupling low-contrast boundaries, necessitating a technical solution that can improve the segmentation accuracy and morphological fidelity of subtle tissue structures in medical volumetric data.

[0004] In summary, improving the segmentation accuracy and morphological fidelity of fine tissue structures with low-contrast boundary features in medical volumetric data is a technical problem that urgently needs to be solved in this field.

[0005] To address this, a multi-scale feature image semantic segmentation method for alveolar bone defect detection is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-scale feature image semantic segmentation method for alveolar bone defect detection. This invention involves converting the artifact-laden cone-beam computed tomography tensor to the frequency domain, applying a directional bandpass mask to filter noise, and performing nonlinear truncation on the overlapping region of attenuation coefficients in the spatial domain to amplify the gray-level gradient. Multi-scale spatial features are extracted using a feature network, and dynamic weighting is performed based on anisotropic local gradient variance. During the optimization stage, an asymmetric weighted three-dimensional geometric deviation penalty term is introduced to force decoupling of adhesive boundaries, outputting a segmentation mesh. Finally, the principal axis of the local morphology of a single tooth is piecewise fitted, and projection calculations are performed along this principal axis between structural contour lines. This invention effectively improves the three-dimensional segmentation accuracy and morphological fidelity of low-contrast fine tissue structures.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A multi-scale feature image semantic segmentation method for alveolar bone defect detection includes: The original cone-beam computed tomography tensor containing metal artifacts is obtained and transformed to the frequency domain; a directional bandpass mask is applied in the frequency domain to attenuate the frequency band vector corresponding to the ray-hardening artifact matrix, while retaining the spatial high-frequency features representing the edge of the alveolar bone cortex, and reconstructed into a purified spatial tensor through inverse transformation; The cleaned spatial tensor is subjected to nonlinear window width and window level truncation, and the gray-level gradient in the overlapping area of ​​the dentin and cortical bone attenuation coefficients is non-uniformly amplified to generate an adaptive voxel tensor. Multi-scale spatial features of the adaptive voxel tensor are extracted through a feature extraction network, and dynamic weights are assigned to the multi-scale spatial features based on the local gradient variance distribution. The output is a multi-label 3D prediction probability tensor containing the tooth target and alveolar bone region, which is then reconstructed into a 3D segmented patch mesh. The cementoenamel boundary and alveolar ridge crest contour lines on the 3D segmented patch mesh are located, and the principal axis of the single tooth shape is fitted to establish a local coordinate system. The projection line between the two contour lines is calculated along the direction of the principal axis of the shape, and the Euclidean distance quantization scalar is output.

[0008] Preferably, a directional bandpass mask is applied in the frequency domain. Specifically, this is achieved by: obtaining a frequency domain energy tensor from the original cone-beam computed tomography tensor via a three-dimensional discrete Fourier transform; extracting the coordinates of the central voxel containing the metallic material and mapping them to the frequency domain as the origin of the radiation source; extracting a high-energy distribution region radiating from the origin of the radiation source within the frequency domain energy tensor, and defining its coordinate index as the ray-hardening artifact matrix; constructing a continuous transition mask of the same size as the frequency domain energy tensor as the directional bandpass mask; using the directional bandpass mask to attenuate the vector magnitudes within the ray-hardening artifact matrix while maintaining the external vector magnitudes unchanged to preserve the spatial abrupt gradient characteristics; and reconstructing and outputting the purified spatial tensor via an inverse Fourier transform.

[0009] Preferably, the adaptive voxel tensor is generated by: extracting the mass decay coefficient values ​​of voxels in the purified spatial tensor; setting the closed interval between the preset first and second decay coefficient boundary values ​​as the overlapping region of the decay coefficients of dentin and cortical bone; performing nonlinear remapping on voxels inside the overlapping region using a logistic growth activation function; introducing a nonlinear asymptotic decay function on voxels outside the overlapping region, and performing continuous numerical compression mapping on both sides based on the boundary values ​​to generate the adaptive voxel tensor with continuous values ​​at the decay boundaries. Preferably, multi-scale spatial features are extracted through a feature extraction network. Specifically, the feature extraction network is a multi-level shifted window three-dimensional self-attention network. By dividing the adaptive voxel tensor into a non-overlapping sequence of three-dimensional voxel blocks, multi-head self-attention calculations are performed at different network depth levels to obtain local perceptual features. After shifting the self-attention window, global dependency features across regions are calculated. Finally, the multi-scale spatial features are synthesized by splicing them together in the channel dimension.

[0010] Preferably, dynamic weight allocation is performed on multi-scale spatial features based on the local gradient variance distribution, outputting a multi-label 3D prediction probability tensor containing tooth targets and alveolar bone regions, and reconstructing it into a 3D segmentation patch mesh. Specifically, this involves: constructing a 3D bounding box centered on each voxel point of the adaptive voxel tensor; calculating the discrete variance of the first-order partial derivatives of the voxels within the bounding box in the three orthogonal spatial dimensions; converting the discrete variance into numerical factors distributed in a closed interval between 0 and 1 using a logistic sigmoid activation function with scaling parameters; downsampling and aligning the numerical factors according to the spatial resolution of the feature extraction network layer to generate a hierarchical weight matrix; performing element-wise multiplication of the hierarchical weight matrix with the corresponding multi-scale spatial features to generate a semantic feature tensor with an adaptive decay distribution, inputting it into the semantic segmentation model, and reconstructing and outputting a multi-label 3D prediction probability tensor.

[0011] Preferably, the generation process of the multi-label 3D prediction probability tensor is optimized by a loss function that includes a boundary-aware penalty term. The specific implementation of the boundary-aware penalty term includes: using a 3D Laplacian discrete second-order partial derivative operator to perform spatial convolution filtering on the multi-label 3D prediction probability tensor and the real label mask tensor respectively, extracting the first derivative extreme points, and forming a first 3D surface boundary point set and a second 3D surface boundary point set respectively; calculating the bidirectional Hausdorff distance between the first 3D surface boundary point set and the second 3D surface boundary point set, and introducing the bidirectional Hausdorff distance as a spatial geometric deviation value into the overall loss function.

[0012] Preferably, before reconstructing and outputting the 3D segmented mesh, a post-processing step is included. Specifically, this includes: converting the average physical width parameter of the periodontal ligament based on healthy human statistics into a standard voxel distance threshold according to the device's spatial resolution; separating the multi-label 3D prediction probability tensor into independent tooth solid masks and cortical bone solid masks; calculating a 3D Euclidean distance field outward from the outer surface of the tooth solid mask, generating a distance transformation tensor; traversing all voxels in the cortical bone solid mask, mapping their spatial coordinates to the distance transformation tensor, and identifying cortical bone classification voxels whose distance values ​​are less than the standard voxel distance threshold; performing connected component thickness analysis on the identified cortical bone classification voxels, and reassigning only artifact adhesion voxels with a thickness dimension not greater than the standard voxel distance threshold and exhibiting a banded distribution as background voxels.

[0013] Preferably, the Euclidean distance quantization scalar is output, specifically implemented by: extracting the three-dimensional curvature extrema points of the neck of the three-dimensional segmented mesh of an independent tooth object and fitting a first spline curve as the cementoenamel boundary contour; extracting the gradient abrupt change point set of the cortical bone segmentation surface and fitting a second spline curve as the alveolar ridge contour; using principal component analysis to extract the covariance matrix eigenvectors of the vertex coordinates of the topological segment of a single tooth target morphology, and using the direction of the first principal component corresponding to the largest eigenvalue as the principal axis of the tooth morphology; constructing a local three-dimensional rectangular coordinate system with the three-dimensional volume centroid of the tooth as the origin and the principal axis of the morphology as the vertical axis; within the three-dimensional rectangular coordinate system, starting from the sampling point of the first spline curve, projecting a line segment along the direction parallel to the vertical axis to the second spline curve, and calculating the absolute value of the physical size of the projected line segment as the Euclidean distance quantization scalar.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a joint feature perception and enhancement mechanism in the frequency and spatial domains by applying a directional bandpass mask in the frequency domain combined with nonlinear window width and window level truncation in the spatial domain. This mechanism can suppress the interference of stripe artifacts in medical volumetric data while specifically amplifying the gray-level differences in the overlapping range of attenuation coefficients of specific tissues. This design improves the signal-to-noise ratio and structural contrast of the target region, effectively preserving and highlighting the boundary features of fine tissue structures before entering the feature extraction network, providing a reliable data foundation for subsequent accurate semantic segmentation.

[0015] 2. This invention proposes a network optimization scheme that integrates dynamic weighting of local gradient variance with a boundary-aware penalty term. In the feature extraction stage, the network is guided to allocate computational resources towards boundary regions with gray-level fluctuations based on the local gradient variance distribution. In the model optimization stage, a penalty term is constructed using spatial geometric feature extraction operators and geometric deviation values ​​to form spatial geometric constraints on three-dimensional physical interfaces. This scheme promotes spatial decoupling of adjacent tissue structures with similar imaging performance, improving the three-dimensional segmentation accuracy and morphological fidelity of fine tissue structures.

[0016] 3. This invention establishes a morphological reconstruction and three-dimensional spatial quantization mechanism driven by prior tissue morphology. By introducing physiological physical width parameters to perform morphological opening operations to separate adherent voxels, and fitting the independent morphological principal axes of teeth based on principal component analysis to establish a local microscopic coordinate system, dense projection calculations along specific morphological directions are achieved. This mechanism successfully transforms voxel-level semantic segmentation results into Euclidean distance quantization scalars, automatically outputting distance parameters that conform to medical measurement standards, reducing measurement errors caused by spatial projection transformation in the global coordinate system, and enhancing the objectivity and accuracy of the quantitative analysis results. Attached Figure Description

[0017] Figure 1 A flowchart of a multi-scale feature image semantic segmentation method for alveolar bone defect detection provided in an embodiment of the present invention; Figure 2 The flowchart of the multi-scale feature extraction and dynamic weight allocation module provided in the embodiments of the present invention is as follows: Figure 3 A flowchart for morphological principal axis quantization distance measurement calculation provided in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figures 1 to 3 This invention provides a multi-scale feature image semantic segmentation method for alveolar bone defect detection, the technical solution of which is as follows: A multi-scale feature image semantic segmentation method for alveolar bone defect detection, characterized by comprising: The original cone-beam computed tomography tensor containing metal artifacts is obtained and transformed to the frequency domain; a directional bandpass mask is applied in the frequency domain to attenuate the frequency band vector corresponding to the ray-hardening artifact matrix, while retaining the spatial high-frequency features representing the edge of the alveolar bone cortex, and reconstructed into a purified spatial tensor through inverse transformation; The cleaned spatial tensor is subjected to nonlinear window width and window level truncation, and the gray-level gradient in the overlapping area of ​​the dentin and cortical bone attenuation coefficients is non-uniformly amplified to generate an adaptive voxel tensor. Multi-scale spatial features of the adaptive voxel tensor are extracted through a feature extraction network, and dynamic weights are assigned to the multi-scale spatial features based on the local gradient variance distribution. The output is a multi-label 3D prediction probability tensor containing the tooth target and alveolar bone region, which is then reconstructed into a 3D segmented patch mesh. The generation process of the multi-label 3D prediction probability tensor is optimized by a loss function that includes a boundary-aware penalty term. The boundary-aware penalty term extracts the interface edges through a spatial geometric feature extraction operator, and uses the spatial geometric deviation between the predicted edge and the real edge as the gradient penalty value. Locate the cementoenamel boundary and alveolar crest contour lines on the three-dimensional segmented mesh, fit the principal axis of a single tooth morphology to establish a local coordinate system, perform projection line calculation between the two contour lines along the principal axis of morphology, and output the Euclidean distance quantization scalar.

[0020] Example 1: This embodiment relates to a method for detecting alveolar bone defects. Weak areas of the alveolar bone (such as bone windows and fractures less than 1 mm thick) are easily lost during cone-beam computed tomography (CBCT) imaging due to downsampling. Furthermore, because the mass attenuation coefficients of the tooth root (dentin and cementum) and the alveolar bone cortex highly overlap in X-ray imaging, coupled with the interference of radiation-induced hardening artifacts from intraoral metal restorations, the bone-dental junction exhibits extremely low contrast and severe pixel-level adhesion.

[0021] As one embodiment of the present invention, refer to Figure 1 A flowchart of a multi-scale feature image semantic segmentation method for alveolar bone defect detection, referring to... Figure 2 The flowchart for the multi-scale feature extraction and dynamic weight allocation module is shown below. Figure 3 Flowchart of morphological principal axis quantization distance measurement calculation.

[0022] Furthermore, a directional bandpass mask is applied in the frequency domain: a three-dimensional discrete Fourier transform is performed on the acquired original cone-beam computed tomography tensor to obtain a frequency domain energy tensor containing an amplitude matrix and a phase matrix; the coordinates of the central voxel containing the metallic material are extracted from the original cone-beam computed tomography tensor, and the central voxel coordinates are mapped to the frequency domain energy tensor as the origin of the radiation source; in the frequency domain energy tensor, a set of spectral energy lines arranged radially around the origin of the radiation source is extracted, and the coordinate index of the set of spectral energy lines is defined as the ray-hardening artifact matrix; a center cutoff frequency filtering matrix is ​​constructed as the directional bandpass mask; the vector amplitudes within the coordinate index range of the ray-hardening artifact matrix are multiplied by an attenuation coefficient in the open interval of zero to one; the vector amplitudes outside the coordinate index range of the ray-hardening artifact matrix remain unchanged, thus completing the noise reduction process.

[0023] Specifically, when processing input data containing metal artifacts, the original three-dimensional voxel tensor is first traversed to locate the three-dimensional physical coordinate center of the high-density metal restoration, and this center point is mapped to the frequency domain space to establish it as the origin of the radiation source. In the frequency domain energy tensor, based on the physical characteristics of broadband dispersion of metal scattering, a continuous wedge-shaped high-energy distribution interval is extracted with the origin of the radiation source as the vertex and diverging along the main direction of ray hardening. The method for determining the main direction of ray hardening is as follows: with the origin of the radiation source as the center, on the central horizontal section of the frequency domain energy tensor (corresponding to the detector plane direction), all angular directions from 0 to 359 degrees are traversed with a step size of 1 degree; along each angular direction, sampling is performed outward from the origin of the radiation source, and the frequency index corresponding to the first drop in amplitude to less than 3 times the median amplitude of the entire frequency domain is recorded as the artifact extension boundary of that direction; continuous angular directions with amplitudes always higher than this threshold are merged, and the segment with the largest coverage angle range is taken, and the angle bisector direction of this segment is the main direction of ray hardening. Centered on the main direction, a predetermined extension angle (15 degrees in this embodiment) is extended to both sides to form a wedge-shaped interval. In the radial dimension, the wedge-shaped interval is extended from the DC component side to the frequency point corresponding to the maximum value in the artifact extension boundary. The set of frequency domain coordinate indices jointly determined by the above-mentioned angle range and radial range is defined as the ray-hardening artifact matrix.

[0024] Subsequently, a Gaussian directional continuous transition mask of the same size as the frequency domain energy tensor is constructed to replace the step-cut mask as the directional bandpass mask. In this mask, the minimum attenuation coefficient value is configured at the position corresponding to the central axis of the distribution matrix, and smoothly increases outwards along the direction away from the central axis according to a Gaussian function, transitioning to a value of 1.0. Element-wise multiplication is performed between this continuous transition mask and the frequency domain energy tensor to smoothly reduce the energy weight of the artifact frequency band, and the data is converted back to the spatial domain through a three-dimensional inverse discrete Fourier transform, outputting the purified tensor.

[0025] This invention adapts to the broadband dispersion physical characteristics of multicolor X-rays scattering within metals by extracting a wedge-shaped continuous frequency domain interval and constructing a Gaussian-oriented continuous transition mask. This smooth transition mechanism mathematically avoids the spatial domain ringing effect caused by hard truncation in the frequency domain of traditional ideal filters, ensuring the smoothness and physical connectivity of the tissue boundaries in the spatial tensor after noise reduction.

[0026] Furthermore, the adaptive voxel tensor is generated, specifically by: extracting the mass decay coefficient values ​​of all voxels in the purified spatial tensor; presetting a first decay coefficient boundary value and a second decay coefficient boundary value, and setting the numerical closed interval between the first decay coefficient boundary value and the second decay coefficient boundary value as the overlapping region of the dentin and cortical decay coefficients; using a logistic growth activation function to perform nonlinear remapping on voxels located inside the overlapping region of the dentin and cortical decay coefficients; for voxels located outside the overlapping region, introducing a nonlinear asymptotic decay function, and performing continuous numerical compression mapping on both sides based on the first decay coefficient boundary value and the second decay coefficient boundary value, so that the global voxel tensor maintains numerical continuity at the boundaries, thereby generating the adaptive voxel tensor.

[0027] Specifically, since dentin and the surrounding cortical bone have similar physical densities, based on medical statistics, a first attenuation coefficient boundary value of 400 and a second attenuation coefficient boundary value of 1500 are set to form a closed core interval for overlapping attenuation coefficients in periodontal tissues. For target voxels whose values ​​fall within this closed interval, a logistic growth function is introduced for nonlinear mapping. This function has a large derivative value in the central region of the closed interval, which can non-uniformly stretch adjacent voxels with slight numerical differences in the mapping space. Simultaneously, for background air, soft tissue, and high-density enamel voxels whose values ​​fall outside this closed interval in the purified spatial tensor, a nonlinear asymptotic attenuation function is introduced. Using the first and second attenuation coefficient boundary values ​​as the starting point, continuous numerical compression mapping is performed towards both sides, ensuring that the global voxel tensor maintains numerical continuity at the attenuation boundaries. This mapping mechanism avoids the artificial step pseudo-gradient created by hard thresholding in the spatial domain, and nonlinearly amplifies the real density change in the tooth-bone interface region, thus transforming it into an adaptive voxel tensor with continuous physical edges and high contrast features, providing reliable preprocessing data for subsequent network extraction of spatial gradient features.

[0028] This invention, by pre-setting a specific physical attenuation range and constructing numerically continuous nonlinear stretching and progressive compression mapping mechanisms inside and outside the range, specifically amplifies the weak gray-level gradient at the junction of dentin and cortical bone. This design overcomes the inherent contrast limitations of imaging equipment, not only alleviating tissue boundary adhesion caused by low contrast at the data level, but also eliminating pseudo-boundary interference caused by artificial truncation. This ensures that subsequent local gradient variance calculations can accurately focus on the real pathological defect boundaries, enhancing the distinguishability and physical realism of the target features.

[0029] Furthermore, multi-scale spatial features are extracted through a feature extraction network. Specifically, the feature extraction network is configured as a multi-level shift-window three-dimensional self-attention network; the adaptive voxel tensor is divided into a sequence of non-overlapping three-dimensional voxel blocks of fixed size; and multi-head self-attention computation is performed at the first network depth level to obtain local perceptual features. In the specific implementation, the initial channel dimension of the multi-level shift window 3D self-attention network is set to 96, and the network depth includes four levels. In each level, the number of attention heads for multi-head self-attention computation is set to 3, 6, 12, and 24, respectively; the multi-head self-attention computation is based on a 3D non-overlapping voxel block sequence with a preset dimension, performing serialized linear projection and feature reconstruction.

[0030] At the second network depth level, the three-dimensional self-attention window is translated by a preset voxel unit distance along the first, second, and third orthogonal axes, respectively, to calculate the global dependency features; the local perceptual features and the global dependency features are then concatenated along the channel dimension to synthesize the multi-scale spatial features.

[0031] The feature extraction network requires supervised training before actual deployment. The real label tensor used for training is obtained by spatially registering Micro-CT (high-resolution micro-computed tomography) images with standard cone-beam computed tomography images from the same machine. The high contrast of Micro-CT is utilized to extract precise masks of fine cortical bone defects, which are then downsampled and aligned to the standard cone-beam computed tomography resolution as a training benchmark. This guides the network to learn the distribution patterns of fine boundaries that are difficult to annotate manually at low resolution.

[0032] The specific process for constructing the training data is as follows: First, Micro-CT scans (voxel resolution no less than 20 micrometers isotropic) and clinical CBCT scans (voxel resolution consistent with the target application device, such as 200 micrometers isotropic) are performed sequentially on the same extracted tooth. Before the two scans, three zirconia microspheres with a diameter of approximately 1 mm are adhered to the crown surface as reference markers. Based on the spatial coordinates of the three reference markers, a point set matching algorithm is used to calculate the rigid body transformation matrix to complete the initial rigid registration. Subsequently, using mutual information as a similarity measure, a regularized B-spline variable registration algorithm is used to locally refine the rigid registration results to compensate for the minor deformations caused by the loss of soft tissue in the extracted tooth. On the Micro-CT images, a bone cortex mask is semi-automatically segmented by an oral radiologist using a bone cortex density threshold, and the mask is manually reviewed and confirmed as the true label. After applying Gaussian pre-filtering (σ = downsampling factor / 2) to the true label at Micro-CT resolution, trilinear interpolation is used to downsample to CBCT resolution to obtain the true label tensor.

[0033] Specifically, the input adaptive voxel tensor is spatially segmented into non-overlapping 3D physical voxel blocks of size 4×4×4, and flattened into high-dimensional sequential feature vectors through linear projection. At the first deep layer of the network, a multi-head self-attention mechanism within a fixed local window is used to extract the morphological dependencies between adjacent voxel blocks, forming a local perceptual feature map. After completing the local perceptual feature extraction at the first deep layer, the system introduces a 3D block fusion layer to perform spatial downsampling on the feature map. Specifically, adjacent 2×2×2 voxel blocks are stitched together along the channel dimension, and the channel dimension is adjusted to a preset size through linear mapping, thereby reducing the spatial resolution of the feature map to half its original value along the first, second, and third orthogonal axes.

[0034] Subsequently, the downsampled features are input into the second network depth level. The original local computation window is shifted by two physical voxels along the X, Y, and Z axes respectively, establishing a new cross-regional computation window. Multi-head self-attention computation is then performed again within the shifted window, enabling voxel blocks belonging to different local windows to generate cross-boundary association computations, thereby capturing the global topological dependency features of the overall jawbone morphology. Finally, the two levels of attention features are cascaded and concatenated along the channel dimension to output complete multi-scale spatial features.

[0035] This invention employs a multi-level shift-window 3D self-attention network structure, establishing long-distance interrelationships between voxels in different spatial regions through discrete voxel block serialization and spatial window translation mechanisms. This architecture enhances the model's ability to recognize the continuity of the complex overall topological structure of the jawbone and reduces the occurrence of irregular tissue discontinuities in the prediction model.

[0036] Furthermore, based on the local gradient variance distribution, dynamic weight allocation is performed on the multi-scale spatial features to output a multi-label 3D prediction probability tensor containing the tooth target and alveolar bone region, and reconstructed into a 3D segmented patch mesh. The specific implementation includes: constructing a 3D cube bounding box with a side length of a preset number of voxels, centered on each voxel point in the adaptive voxel tensor; calculating the first-order partial derivatives of all voxel points inside the 3D cube bounding box in the three spatial dimensions of the first orthogonal axis, the second orthogonal axis, and the third orthogonal axis, and calculating the discrete variance of all the first-order partial derivatives. After calculating the discrete variance values ​​of all the first-order partial derivatives, the maximum and minimum variance values ​​within the global bounding box are first extracted, and the range normalization operation is performed on the variance tensor. Then, the normalized variance values ​​are input into a logistic sigmoid activation function with a scaling parameter. Specifically, the mathematical expression of the logistic sigmoid activation function is as follows: The independent variable x represents the normalized variance value, the dependent variable f(x) represents the continuous numerical factor of the output, and the scaling parameter k is set to a constant between 10.0 and 15.0 (preferably 12.0) based on the prior gray-level distribution of the alveolar bone image. The setting of this parameter range is based on balancing the extreme saturation of attention weights with the smooth propagation of network gradients: when k < 10.0, the mapping value of the high gradient variance region (such as x=1) deviates significantly from 1.0, which cannot produce sufficient attention amplification effect on the boundary of fine bone defects, and is very easy to introduce attention interference from low-frequency noise; when k > 15.0, the function excessively approaches the step function near the center point, causing the effective gradient to disappear during backpropagation, hindering the model's continuous learning of the blurred transition region of the alveolar bone. By constraining the scaling parameter to a closed interval of 10.0 to 15.0, this scheme nonlinearly maps the variance response to continuous numerical factors distributed within a closed interval of zero to one. This ensures that fluctuations in small gradient variance can be instantly amplified to high weights close to 1, while preserving mathematical continuity and differentiability, effectively driving stable optimization in deep networks. This scaling parameter nonlinearly maps the variance response to continuous numerical factors distributed within a closed interval of zero to one. This mapping method ensures that boundary regions with high gradient variance receive high weights approaching one, while homogeneous regions within low-variance structures receive low weights approaching zero.

[0037] The numerical factors are downsampled and aligned according to the spatial resolution of different levels in the feature extraction network to generate corresponding level weight matrices; the level weight matrices are used as channel attention weight parameters and element-wise multiplied with the multi-scale spatial features of the corresponding level.

[0038] Specifically, to accurately capture the boundaries of fine tissues, a 3D cube-shaped local perceptual bounding box with a side length of 3 physical voxels is constructed, centered on each voxel in the input tensor. Inside this bounding box, the discrete first-order gradients of all voxels along the X, Y, and Z axes are calculated, and the variance of these first-order gradients is statistically analyzed. The variance values ​​are mapped to numerical factors between 0 and 1 using a normalized exponential function. Since the 3D spatial resolution of the feature map decreases progressively at different depths of the feature extraction network, a 3D average pooling operation is simultaneously applied to downsample and align the numerical factors accordingly. The aligned numerical factor matrix is ​​directly used as the channel attention weight parameter matrix, and element-wise multiplied with the multi-scale spatial feature tensor of the corresponding resolution is performed to amplify the feature response values ​​of high-variance regions, outputting a semantic feature tensor with an adaptive decay distribution and spatial edge enhancement properties. Subsequently, the semantic feature tensor is input into the subsequent decoding network layer of the semantic segmentation model for multi-label category confidence reconstruction. Finally, the multi-label three-dimensional prediction probability tensor containing the tooth target and alveolar bone region is output and reconstructed into a three-dimensional segmentation mesh using a surface reconstruction algorithm.

[0039] This invention utilizes the discrete variance of local spatial gradients to construct a weight matrix and performs precise spatial resolution alignment calculations between the weights and a multi-scale feature network. Element-wise multiplication establishes a data link from feature enhancement to multi-label classification probability reconstruction. This mechanism guides the network to allocate optimization resources to boundary regions with frequent grayscale fluctuations during backpropagation, mitigating the loss of high-frequency spatial details caused by continuous downsampling and ensuring the geometric fidelity of the subsequently output multi-label 3D prediction probability tensor.

[0040] The specific implementation of dynamic weight allocation based on local gradient variance distribution further includes: replacing the three-dimensional cube bounding box with an anisotropic cuboid bounding box whose major axis is distributed along the longitudinal crown root preset direction of the global cone-beam computed tomography tensor; calculating the discrete variance values ​​of all voxel points inside the anisotropic cuboid bounding box in the plane direction perpendicular to the longitudinal crown root preset direction, and using these values ​​to generate the hierarchical weight matrix to output the semantic feature tensor with adaptive decay distribution.

[0041] Specifically, since alveolar bone fissures primarily manifest as physical defects perpendicular to the long axis of the tooth, a pre-defined global longitudinal axis (i.e., the Z-axis, typically parallel to the general growth direction of the patient's teeth) in the CBCT scan image is used as a geometric guide. The sampling window is adjusted to an anisotropic cuboid with its long axis parallel to the global longitudinal axis. Within this asymmetric bounding box, the variance of the first-order partial derivative in the cross-sectional direction perpendicular to the longitudinal plane is calculated. This geometric reshaping of the sampling space shields a large amount of interference from normal physiological textures distributed longitudinally along the root surface, allowing numerical factors to more sensitively reflect physical gaps in the alveolar bone less than 0.2 mm wide, generating a more discernible hierarchical weight matrix.

[0042] This technique transforms the sampling bounding box from an isotropic to an anisotropic structure, effectively filtering out irrelevant gray-level fluctuation noise parallel to the root surface. This ensures that the gradient variance calculation is absolutely focused on the bone defect gap perpendicular to the root surface, improving the network's recognition weight for extremely thin bone plate defects on the labial and buccal sides. It also solves the problem of detection sensitivity of fine deterioration edges under complex texture interference.

[0043] Furthermore, the generation process of the multi-label 3D prediction probability tensor is optimized through a loss function that includes a boundary-aware penalty term. The specific implementation of the boundary-aware penalty term includes: setting the spatial geometric feature extraction operator as a 3D Laplacian discrete second-order partial derivative operator; performing spatial convolution filtering on the multi-label 3D prediction probability tensor using the 3D Laplacian discrete second-order partial derivative operator to extract first-order derivative extrema points to form a first 3D surface boundary point set; performing spatial convolution filtering on the mask tensor corresponding to the real labels using the 3D Laplacian discrete second-order partial derivative operator to extract first-order derivative extrema points to form a second 3D surface boundary point set; calculating the bidirectional Hausdorff distance between the first 3D surface boundary point set and the second 3D surface boundary point set, and introducing the bidirectional Hausdorff distance as the spatial geometric deviation value into the overall loss function.

[0044] Before introducing spatial geometric bias values, a 3D Dice similarity coefficient loss is used as the basic segmentation loss function to evaluate the volumetric overlap between the predicted probability tensor and the ground truth label. The final overall loss function is constructed by dynamically weighting and summing the basic segmentation loss function and the spatial geometric bias values. The weight coefficients of the basic segmentation loss function are set as follows: The penalty weighting coefficient for the spatial geometric deviation value is ,in The value was set to 0.1 in the early stage of model training and increased linearly to 0.3 as the number of iterations increased. This was to ensure that the network prioritized fitting the macroscopic volume overlap in the early stage of training and forced optimization of the spatial position deviation of the microscopic edges in the later stage of training.

[0045] Specifically, during the iterative optimization phase of the neural network parameters, an independent boundary supervision module targeting the geometric properties of three-dimensional space was set up. A three-dimensional Laplacian discrete operator convolution kernel with fixed weight parameters was invoked. This kernel was set as a 3×3×3 26-neighborhood spatial matrix, with the central voxel weight set to 26 and the weights of the surrounding 26 spatial voxels set to -1. The selection criteria for the 26-neighborhood weight setting of the three-dimensional Laplacian discrete operator (central weight 26, each of the 26 neighbors assigned a weight of -1) are as follows: In the segmentation task of the alveolar bone-dentin interface, since CBCT voxels have isotropic resolution and the root surface may exhibit boundary features in various spatial directions, using an equally weighted 26-neighborhood Laplacian kernel can produce non-zero second-order derivative responses in all 13 independent spatial direction pairs (6 face directions + 6 edge directions + 1 angular direction × 4 = essentially 26 neighbors covering all directions), ensuring that the operator has equal detection sensitivity for the bone-dentin boundary in each orientation. To mitigate the response attenuation caused by the large distance in the corner neighborhood, the corner neighborhood weights can be adjusted to... The edge neighborhood weights are adjusted to The neighborhood weights remain at -1, and the center weight is adjusted to the sum of the absolute values ​​of all the neighborhood weights mentioned above.

[0046] By locating points in a spatial coordinate system where the discrete second-order partial derivative is zero and the first-order derivative reaches an extremum, the surface boundary coordinates of the predicted results in 3D space are extracted, forming a first 3D surface boundary point set. Simultaneously, the same operator is used to extract the manually labeled real 3D mask tensor, forming a second 3D surface boundary point set. Then, the minimum Euclidean distance from each point in the first set to the second set is calculated, and the maximum value is extracted as the forward deviation error; conversely, the minimum distance from the second set to the first set is calculated, and the maximum value is extracted as the backward deviation error. The larger of the forward and backward errors is selected as the bidirectional Hausdorff distance, and this is included as the spatial geometric deviation value in the total loss function.

[0047] This invention extracts discrete boundary features using the Laplacian operator and combines it with bidirectional Hausdorff distance, introducing spatial geometric constraint variables independent of volume overlap during the model optimization stage. This scheme increases the gradient penalty for misclassified voxels at the edges, promoting morphological decoupling of the low-contrast bone-dental physical interface.

[0048] Before incorporating the bidirectional Hausdorff distance into the overall loss function, an asymmetric penalty adjustment step is included: identifying the forward deviation vector shifted from the first three-dimensional surface boundary point set to the interior of the second three-dimensional surface boundary point set, and the backward deviation vector shifted to the exterior; assigning a first penalty coefficient to the forward deviation vector and a second penalty coefficient to the backward deviation vector, wherein the value of the second penalty coefficient is greater than that of the first penalty coefficient.

[0049] Specifically, when calculating the bidirectional Hausdorff distance, the direction of the displacement vector of the predicted boundary relative to the actual boundary is analyzed. If the predicted cortical bone margin shifts inward toward the dentin (indicating over-prediction of bone volume), it is identified as a forward bias vector; if the predicted boundary shifts outward (indicating missed diagnosis of bone defects), it is identified as a backward bias vector. A second penalty coefficient of 1.5 and a first penalty coefficient of 1.0 are set. The ratio of the second penalty coefficient of 1.5 to the first penalty coefficient of 1.0 is determined based on the following clinical evidence: According to the statistics of clinical adverse events reported in the orthodontic literature regarding missed diagnosis of orthodontic-induced root resorption and bone fenestration [specific literature can be supplemented], the severity of orthodontic complications caused by missed diagnosis of bone defects is approximately 1.5 to 2.0 times that of overdiagnosis. This embodiment uses the lower limit of this range, 1.5, to ensure the sensitivity of bone defect detection while avoiding a significant increase in the false positive rate of segmentation caused by over-penalization. This ratio can be adjusted according to specific clinical application scenarios (such as pre-orthodontic assessment, periodontitis progression monitoring). When backward bias is detected, the loss function incurs a gradient penalty of 1.5 times. This asymmetric weighting forces the model to prioritize the detection integrity of bone tissue boundaries when faced with blurred boundaries of extremely low contrast, rather than blindly seeking to maximize mathematical overlap.

[0050] This technique embeds an asymmetric penalty mechanism, consistent with clinical diagnostic logic, into the loss function. By artificially increasing the mathematical cost of missed diagnoses of bone defects (backward bias), it forces the model to maintain higher fidelity at low-contrast boundaries. This design reduces the false negative rate caused by imaging blur without sacrificing overall segmentation accuracy, making the segmentation results more aligned with clinical safety evaluation benchmarks.

[0051] Furthermore, before reconstructing and outputting the 3D segmented mesh, a post-processing step is included. Specifically, this includes: obtaining the average physical width parameter of the periodontal ligament based on statistical analysis of oral physiology in healthy individuals; converting the average physical width parameter of the periodontal ligament into a standard voxel distance threshold based on the isotropic resolution setting parameters of the original cone-beam computed tomography (CBCT) device; separating the multi-label 3D prediction probability tensor into independent tooth solid masks and cortical bone solid masks; and calculating a 3D Euclidean distance field outward from the outer surface of the tooth solid mask to generate a distance transformation containing spatial distance gradient information. Tensor; Traverse all voxels in the cortical bone solid mask, map their spatial coordinates to the distance transformation tensor, and identify cortical bone classification voxels whose distance values ​​are less than the standard voxel distance threshold; Subsequently, perform connected component thickness analysis on the identified cortical bone classification voxels, and only force the reassignment of artifact adhesion voxels with thickness dimensions less than or equal to the standard voxel distance threshold and distributed in a band to background voxels, while retaining connected bone bodies with thicknesses greater than the threshold, thereby completing the directional removal of abnormal adhesion voxels that do not meet the physical distance to the tooth root surface without damaging the morphology of healthy cortical bone.

[0052] Specifically, before converting the probability tensor into a 3D solid model, an objective tissue morphology physiological constant is introduced for spatial morphology correction. The average physical width parameter of the periodontal ligament is converted into a standard voxel distance threshold. The nominal periodontal ligament width is 0.20 mm, calculated as the arithmetic mean of the lower limit (0.15 mm) and upper limit (0.25 mm) of the average physical width of the periodontal ligament. This value is divided by the device's spatial resolution (0.20 mm / voxel in this embodiment), yielding a quotient of 1.00. This quotient is then rounded down (using the floor function) and incremented by 1, resulting in a standard voxel distance threshold of 2 voxel units. This rounding down and incrementing strategy ensures that the threshold is never less than the number of voxels corresponding to the lower limit of the periodontal ligament's physical width, conservatively preventing healthy periodontal spaces from being misdiagnosed as adhesions. For a device with a spatial resolution of 0.25 mm / voxel, the calculation is floor(0.20 / 0.25) + 1 = floor(0.80) + 1 = 1 voxel.

[0053] Subsequently, the multi-label 3D prediction probability tensor is separated into independent tooth solid masks and cortical bone solid masks. Using the outer surface of the tooth solid mask as the starting point, Euclidean distances are calculated to the surrounding 3D space, generating a distance transformation tensor containing spatial distance gradient information. Next, all voxels in the cortical bone solid mask are traversed, and their spatial coordinates are precisely mapped to the distance transformation tensor, identifying the set of cortical bone voxels whose distance values ​​are less than the standard voxel distance threshold.

[0054] Finally, a joint audit based on prediction confidence and topological connectivity is performed on the set of voxels to be processed from the bone cortex: the system extracts the classification confidence values ​​of the voxels in the multi-label 3D prediction probability tensor; only voxels with classification confidence values ​​lower than a preset safety threshold, and which are in a free state in the topological structure or cross the periodontal ligament space only in the form of bridges with the width of a single voxel, are judged as artifact adhesions caused by partial volume effects and are forcibly reassigned as background voxels. For thin bone plate voxels that maintain planar continuity with the surrounding bone cortex in the topology, even if their local physical thickness is less than the standard voxel distance threshold, they are retained.

[0055] This invention utilizes a three-dimensional Euclidean distance field to reconstruct the boundary isolation logic between multi-labeled entities. Using independently extracted tooth surfaces as the computational benchmark, it performs voxel-level directional intervention only within the threshold range of the periodontal ligament space, avoiding irreversible damage to the extremely thin labial and buccal bone plates. This design successfully severs abnormal adhesions at the tooth-bone interface while effectively ensuring the physical fidelity of the three-dimensional morphology of the fine cortical bone, ensuring that the final reconstructed three-dimensional entity model conforms to objective biological topological distribution principles.

[0056] Furthermore, the Euclidean distance quantization scalar is output. Specifically, this is achieved by: extracting independent tooth objects after removing conglomerates based on the periodontal ligament width parameter, and calculating the three-dimensional curvature extrema points of the cervical region transitioning between the crown and root in the three-dimensional segmented mesh, forming a first spline curve as the cementoenamel boundary contour; extracting the set of gradient abrupt change points on the segmented surface belonging to the cortical bone classification from the multi-labeled three-dimensional prediction probability tensor, and fitting them to form a second spline curve as the alveolar ridge crest contour; and using principal component analysis to extract the vertex coordinates contained within the target morphological topology segment of the independent tooth object. The eigenvectors of the covariance matrix formed by the data set; the direction of the first principal component eigenvector corresponding to the largest eigenvalue is determined as the principal axis of the single tooth morphology; a three-dimensional rectangular coordinate system is constructed with the three-dimensional volume centroid of the single tooth as the origin and the principal axis of the single tooth morphology as the longitudinal rectangular coordinate axis; in the three-dimensional rectangular coordinate system, starting from the discrete sampling point on the first spline curve, a projection line segment is drawn along the direction parallel to the longitudinal rectangular coordinate axis to the second spline curve, and the absolute value of the physical size of the projection line segment is calculated as the Euclidean distance quantization scalar.

[0057] Specifically, after removing the adhesion voxels, independent surface models of each individual tooth are isolated. The extreme points of the crown-root transition region are located by calculating the spatial curvature of the mesh surface, and a closed first spline curve is fitted. Simultaneously, the abrupt contours of the cortical bone surface in the prediction tensor are extracted and fitted with a second spline curve. To eliminate physical projection errors caused by patient head positioning, the coordinate data set of all three-dimensional vertices on the surface of the current independent tooth model is extracted, the covariance matrix of this set is calculated, and principal component analysis is performed. The direction of the first principal component with the largest variance eigenvalue is defined as the morphological principal axis of the tooth. The three-dimensional volume geometric centroid of the independent tooth is calculated as the spatial origin, and a local vertical axis is established with the morphological principal axis, completing the construction of a three-dimensional rectangular coordinate system. Finally, in this local coordinate system, a ray projection is initiated from the uniformly distributed sampling points on the first spline curve along the local vertical axis, and the absolute value of the physical dimension between the intersection of the projection line and the second spline curve is calculated. This distance scalar is output for pathological staging reference.

[0058] This invention establishes a local Cartesian coordinate system based on principal component analysis after separating independent target objects, converting 3D contour features into Euclidean projection line segment parameters in a standard coordinate system. This mechanism eliminates projection bias caused by spatial attitude tilt in the global coordinate system, achieving high-precision conversion from discrete voxel segmentation to objective medical measurement parameters.

[0059] The specific implementation method for fitting the principal axis of a single tooth morphology includes: dividing the three-dimensional segmented mesh of a single tooth into a cervical topological segment, a middle topological segment, and a root topological segment along the crown-root direction; performing principal component analysis on the three topological segments respectively to extract local feature vectors; and selecting the first principal component feature vector corresponding to the cervical topological segment as the principal axis of the single tooth morphology to perform projection line calculation.

[0060] Specifically, for the common root curvature (malformed central cusp or physiological curvature) in the anterior region, after extracting the patch mesh, the longitudinal geometric center axis of the bounding box of the current independent tooth model is first extracted. The tooth model is then divided into three topological segments—cervical, middle, and root—along a direction perpendicular to this longitudinal geometric center axis. Only the vertex coordinate set of the cervical third region is extracted for covariance matrix calculation and PCA eigenvector extraction. Since the cervical region is usually less affected by the overall root curvature and is the area where alveolar bone resorption occurs first, the morphological principal axis established by this local first principal component direction can more accurately represent the insertion vector of the clinical probe. Subsequently, projection line calculation is performed only along the longitudinal axis of this local coordinate system to obtain the physical distance from the cementoenamel junction to the alveolar ridge crest.

[0061] This invention solves the coordinate system drift problem in complex morphological and deformed scenarios by employing a segmented multi-local principal axis fitting mechanism. This method ensures that the path for quantized distance measurement always conforms to the morphological and structural characteristics of periodontal tissues, eliminating geometric projection errors caused by root curvature.

[0062] Example 2: This embodiment reveals the core solution of the present invention for processing extremely thin bone plate defects (such as bone fissures and fenestrations with a thickness of less than 0.5 mm) in oral CBCT images. Addressing the shortcomings of existing technologies, such as the inability to effectively balance noise reduction and boundary fidelity, as well as segmentation adhesions caused by overlapping physical densities of the teeth and bones, this embodiment establishes a high-precision, end-to-end technical implementation path from the underlying physical signal processing to the quantitative output of clinical morphology.

[0063] As one embodiment of the present invention, refer to Figure 1 A flowchart of a multi-scale feature image semantic segmentation method for alveolar bone defect detection, referring to... Figure 2 The flowchart for the multi-scale feature extraction and dynamic weight allocation module is shown below. Figure 3 Flowchart of morphological principal axis quantization distance measurement calculation.

[0064] First, in the physical signal preprocessing stage, this embodiment employs a joint enhancement mechanism based on a frequency-domain directional bandpass mask and a spatial-domain nonlinear mapping. Specifically, a three-dimensional discrete Fourier transform is performed on the original CBCT three-dimensional voxel tensor. In the frequency domain space, the origin of the radiation source is established based on the central voxel coordinates of the metal material, and a set of high-energy spectral lines arranged radially is identified, i.e., the ray-hardening artifact matrix. At this point, an anisotropic directional bandpass mask is applied. By multiplying the frequency band vector corresponding to this distribution matrix by an attenuation coefficient, the stripe noise generated by the metal artifacts is filtered out in a directional manner. Simultaneously, the spatial high-frequency components outside the range of this matrix are strictly kept unattenuated, thereby ensuring that the high-frequency signal representing the edge of the thin alveolar cortex is not destroyed when reconstructed into a clean spatial tensor by the inverse transform. Subsequently, considering the physical characteristic of the high overlap of dentin and cortical bone HU values, a closed interval of dynamic attenuation coefficient based on the patient's bone mineral density statistics (e.g., 400 HU to 1500 HU) is set, and a logistic growth activation function is applied to perform nonlinear mapping on voxels within this interval. This operation artificially stretches the grayscale gradient of the overlapping area at the numerical level, transforming the extremely low-contrast tissue boundaries into highly dynamic features that the network can recognize.

[0065] Secondly, in the multi-scale feature perception stage, this embodiment introduces an anisotropic gradient variance weighting mechanism with tissue morphology orientation. The feature extraction network is constructed as a multi-level shifted window 3D self-attention architecture. While extracting local perceptual features and global dependent features, an anisotropic cuboid local perceptual bounding box is constructed based on the adaptive voxel tensor. Unlike conventional isotropic operations, the major axis of this cuboid bounding box is parallel to the principal axis of the initially fitted tooth morphology, and the discrete variance of the first-order partial derivative is specifically calculated in the direction perpendicular to the plane of the tooth's major axis (i.e., the horizontal section). Since the physical characteristics of alveolar bone defects are mainly manifested as vertical bone interruptions, this anisotropic sampling method can effectively filter out physiological texture noise parallel to the root surface, enabling the calculated hierarchical weight matrix to accurately anchor the physical gaps of thin bone plates. This forces the network to concentrate optimization resources on the highly challenging boundary voxels, thereby achieving high fidelity of the microstructure in the deep feature encoding stage.

[0066] Furthermore, in the model optimization and loss constraint stage, this embodiment implements a geometric boundary decoupling strategy based on asymmetric weighting of clinical morphological risk. A three-dimensional Laplacian discrete second-order partial derivative operator is used to perform convolution filtering on the predicted probability tensor and the true label, extracting two sets of three-dimensional surface boundary point sets and calculating the bidirectional Hausdorff distance. During this process, this embodiment creatively introduces an asymmetric penalty factor: for missed diagnoses of bone defects caused by outward shift of the predicted boundary, a significantly higher penalty weight coefficient than that for inward shift of the predicted boundary is assigned. This mathematical constraint based on clinical diagnostic risk forces the network to generate a clear spatial decoupling dynamic at the dental-bone physical interface, ensuring high sensitivity capture of extremely small bone fenestration areas without sacrificing overall accuracy, and solving the problem of predicted boundary wobbling and blurring in low-contrast regions caused by conventional algorithms.

[0067] Finally, in the quantitative output and clinical translation stage, this embodiment constructs a segmented multi-local principal axis Euclidean projection quantization system. After performing morphological opening operations to remove confluent voxels and reconstructing the patch mesh, for the common root curvature physiological morphology phenomenon in anterior teeth, the single tooth model is divided into three independent topological segments: cervical, middle, and root, and principal component analysis is performed on each segment. The direction of the first principal component feature vector of the cervical topological segment is forcibly extracted as the morphological principal axis of the tooth, thereby establishing a local microscopic rectangular coordinate system. Subsequently, the first spline curve of the cementoenamel junction and the second spline curve of the alveolar ridge are located, and point-to-point dense projection calculation between the two curves is performed along the longitudinal direction of the local morphological principal axis.

[0068] This mechanism eliminates global projection errors caused by patient head positioning or physiological curvature of tooth roots. The output Euclidean distance quantization scalar can truly reflect the linear physical depth of the clinical periodontal probe along the insertion direction of the periodontal pocket. Through the deep coupling of the above physical signal governance, morphological guidance weighting, risk perception optimization and local coordinate quantization, this scheme finally establishes a highly robust automated evaluation system.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-scale feature image semantic segmentation method for alveolar bone defect detection, characterized in that, include: The original cone-beam computed tomography (CBCT) tensor containing metal artifacts was obtained and transformed to the frequency domain. A directional bandpass mask was applied in the frequency domain to attenuate the frequency band vectors corresponding to the ray-hardening artifact matrix, while preserving the spatial high-frequency features representing the edge of the alveolar cortex. This was then reconstructed into a purified spatial tensor via inverse transformation. Specifically, the original CBCT tensor was subjected to a three-dimensional discrete Fourier transform to obtain the frequency domain energy tensor. The coordinates of the central voxel containing the metal material were extracted and mapped to the frequency domain as the origin of the radiation source. From the frequency domain energy tensor, a high-energy distribution range radiating outwards from the origin of the radiation source was extracted. The coordinate index is defined as the ray-hardening artifact matrix; a Gaussian directional continuous transition mask of the same size as the frequency domain energy tensor is constructed as a directional bandpass mask; in the directional bandpass mask, the minimum attenuation coefficient value is configured at the position corresponding to the central axis of the ray-hardening artifact matrix, and smoothly increases outwards along the direction away from the central axis according to the Gaussian function law to the value 1.0; the directional bandpass mask is used to attenuate the vector magnitude inside the ray-hardening artifact matrix, while keeping the external vector magnitude unchanged to preserve the spatial abrupt gradient characteristics; the purified spatial tensor is reconstructed by inverse Fourier transform. The purified spatial tensor is truncated using a nonlinear window width and window level, and the gray-level gradient of the overlapping region of the dentin and cortical bone attenuation coefficients is non-uniformly amplified to generate an adaptive voxel tensor. Specifically, the mass attenuation coefficient values ​​of voxels in the purified spatial tensor are extracted, and the closed interval between the preset first and second attenuation coefficient boundary values ​​is set as the overlapping region of the dentin and cortical bone attenuation coefficients. The logistic growth activation function is used to perform nonlinear remapping on the voxels inside the overlapping region. A nonlinear asymptotic attenuation function is introduced to the voxels outside the overlapping region, and continuous numerical compression mapping is performed on both sides based on the boundary values ​​to generate an adaptive voxel tensor with continuous values ​​at the attenuation boundary. Multi-scale spatial features of the adaptive voxel tensor are extracted through a feature extraction network, and dynamic weights are assigned to the multi-scale spatial features based on the local gradient variance distribution. This outputs a multi-label 3D prediction probability tensor containing the tooth target and alveolar bone region, which is then reconstructed into a 3D segmented patch mesh. The generation process of the multi-label 3D prediction probability tensor is optimized using a loss function that includes a boundary-aware penalty term. Specifically, a 3D Laplacian discrete second-order partial derivative operator is used to perform spatial convolution filtering on both the multi-label 3D prediction probability tensor and the real label mask tensor, extracting the first-order derivative extrema points to form the first and second 3D surface boundary point sets, respectively. The bidirectional Hausdorff distance between the first and second 3D surface boundary point sets is calculated and introduced into the overall loss function as a spatial geometric deviation value. The cementoenamel boundary and alveolar ridge contour lines on the 3D segmented patch mesh are located, and the principal axis of a single tooth morphology is fitted to establish a local coordinate system. Projection line calculations are performed along the principal axis direction between the two contour lines, outputting a Euclidean distance quantized scalar.

2. The multi-scale feature image semantic segmentation method for alveolar bone defect detection according to claim 1, characterized in that, Multi-scale spatial features are extracted using a feature extraction network. Specifically, the feature extraction network is a multi-level shifted window 3D self-attention network. The adaptive voxel tensor is divided into a non-overlapping sequence of 3D voxel blocks. Multi-head self-attention calculations are performed at different network depth levels to obtain local perceptual features. After shifting the self-attention window, global dependency features across regions are calculated. Finally, the multi-scale spatial features are synthesized by splicing them together in the channel dimension.

3. The multi-scale feature image semantic segmentation method for alveolar bone defect detection according to claim 1, characterized in that, Based on the local gradient variance distribution, dynamic weight allocation is performed on multi-scale spatial features to output a multi-label 3D prediction probability tensor containing tooth targets and alveolar bone regions, which is then reconstructed into a 3D segmentation mesh. Specifically, this involves: constructing a 3D bounding box centered on each voxel point of the adaptive voxel tensor; calculating the discrete variance of the first-order partial derivatives of the voxels within the bounding box in the three orthogonal spatial dimensions; converting the discrete variance into numerical factors distributed in a closed interval between 0 and 1 using a logistic sigmoid activation function with scaling parameters; downsampling and aligning the numerical factors according to the spatial resolution of the feature extraction network layer to generate a hierarchical weight matrix; performing element-wise multiplication of the hierarchical weight matrix with the corresponding multi-scale spatial features to generate a semantic feature tensor with an adaptive decay distribution, which is then input into the semantic segmentation model to reconstruct and output a multi-label 3D prediction probability tensor.

4. The multi-scale feature image semantic segmentation method for alveolar bone defect detection according to claim 3, characterized in that, Before reconstructing and outputting the 3D segmented mesh, a post-processing step is included. Specifically, this involves: converting the average physical width parameter of the periodontal ligament based on healthy human statistics into a standard voxel distance threshold according to the device's spatial resolution; separating the multi-label 3D prediction probability tensor into independent tooth solid masks and cortical bone solid masks; calculating a 3D Euclidean distance field outward from the outer surface of the tooth solid mask, generating a distance transformation tensor; traversing all voxels in the cortical bone solid mask, mapping the spatial coordinates of the voxels to the distance transformation tensor, and identifying cortical bone classification voxels whose distance values ​​are less than the standard voxel distance threshold; performing connected component thickness analysis on the identified cortical bone classification voxels, and reassigning only artifact adhesion voxels with a thickness dimension not greater than the standard voxel distance threshold and exhibiting a banded distribution as background voxels.

5. The multi-scale feature image semantic segmentation method for alveolar bone defect detection according to claim 4, characterized in that, The specific implementation of outputting the Euclidean distance quantization scalar includes: extracting the extreme points of the three-dimensional curvature of the neck of the three-dimensional segmented mesh of an independent tooth object and fitting a first spline curve as the cementoenamel boundary contour; extracting the set of gradient abrupt change points on the cortical bone segmentation surface and fitting a second spline curve as the alveolar ridge contour; using principal component analysis to extract the eigenvectors of the covariance matrix of the vertex coordinates of the topological segment of the target tooth shape, and using the direction of the first principal component corresponding to the largest eigenvalue as the principal axis of the tooth shape; constructing a local three-dimensional rectangular coordinate system with the centroid of the three-dimensional volume of the tooth as the origin and the principal axis of the shape as the vertical axis; and within the three-dimensional rectangular coordinate system, starting from the sampling point of the first spline curve, projecting a line segment along the direction parallel to the vertical axis to the second spline curve, and calculating the absolute value of the physical size of the projected line segment as the Euclidean distance quantization scalar.

Images