Method for quantitatively evaluating wisdom tooth and mandibular nerve canal risk based on spatial distance field

Abstract

The application discloses a wisdom tooth and mandibular nerve tube risk quantitative evaluation method based on a space distance field, and belongs to the technical field of medical images. A CBCT image is trained and inferred based on an nnU-Net neural network, and a semantic segmentation mask containing a mandibular wisdom tooth, a dental pulp of the wisdom tooth and a mandibular nerve tube is output. The mask is subjected to instance separation, and independent instance masks of the structures are obtained. A center line of the mandibular nerve tube mask is extracted and a signed distance field is generated based on the mask. A dynamic local coordinate system is constructed based on the center line, multi-dimensional risk analysis is performed in combination with minimum distance, a dental pulp center position and geometric features extracted from the distance field, a nerve damage probability is calculated through a logistic regression model, is mapped to a clinical risk level, and a structured report is output. The application realizes automation and objectivity of the evaluation process, improves the segmentation and geometric quantitative precision, and provides quantifiable and accurate decision support for clinical treatment.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a method for quantitative assessment of the risk of wisdom teeth and mandibular nerve canal based on spatial distance field. Background Technology

[0002] Wisdom teeth, also known as third molars, are common impacted teeth in humans. Damage to the inferior alveolar canal (IAC) is a major and serious complication during mandibular wisdom tooth extraction, potentially leading to persistent or permanent numbness or pain in the lower lip, chin, and gums, significantly impacting quality of life. Therefore, accurate preoperative assessment of the three-dimensional relationship between the wisdom tooth and the inferior alveolar canal is crucial for developing a surgical plan and mitigating the risk of nerve damage.

[0003] Traditional risk assessment primarily relies on two-dimensional orthopantomograms (OPG), which, while showing approximate adjacent relationships, suffer from issues such as image overlap and magnification distortion, failing to provide precise three-dimensional distance and orientation information, and making it even more difficult to determine whether the neural canal is compressed, perforated, or encased by the tooth root. The widespread adoption of cone-beam computed tomography (CBCT) has made accurate three-dimensional assessment possible. However, existing CBCT-based assessment methods still have significant limitations, mainly in the following aspects: First, the assessment process is highly dependent on the physician's manual operation and subjective experience. Clinically, physicians usually need to manually locate and measure the closest distance between the wisdom tooth and the nerve canal on a CBCT workstation, and visually determine the orientation (e.g., buccal, lingual, apical, etc.). This method is not only time-consuming and laborious, but also has poor repeatability of measurement results, heavily relying on the operator's experience and skill, making standardization and quantification difficult.

[0004] Secondly, existing automated or semi-automated methods have shortcomings in segmentation accuracy and instance separation. Although some studies have attempted to apply deep learning models (such as U-Net and its variants) to perform semantic segmentation of oral structures in CBCT images, these general-purpose models have limited segmentation accuracy for high-risk structures with small volume and subtle contrast differences, such as wisdom teeth and neural tubes. More importantly, semantic segmentation can only distinguish between different categories of structures (such as "neural tube" and "teeth"), but cannot automatically separate multiple instances of the same category (such as left and right neural tubes, or multiple impacted wisdom teeth), which poses an obstacle to subsequent accurate geometric analysis of individual instances.

[0005] Furthermore, the methods for quantifying distance and orientation are rather crude and lack fine characterization of complex anatomical morphology. Existing methods are mostly based on Euclidean distance calculations at voxel centers, and their accuracy is limited by image resolution (voxel level), making it impossible to achieve sub-voxel level precision measurements. At the same time, the determination of directional relationships (buccal / lingual) often relies on a simple global coordinate system or a manually defined reference plane, failing to consider the curvature of the mandibular canal itself and the variability of the impacted wisdom tooth angle, leading to inaccurate orientation determination, especially in complex impacted cases.

[0006] Finally, risk assessment lacks a systematic and interpretable quantitative model. Current assessments are mostly based on simple classifications using a single distance threshold (such as "contact" or "no contact"), or rely on physicians' subjective comprehensive judgment of multiple signs (such as neural tube stenosis, interruption of the linea alba, root apex curvature, etc.). This model lacks an objective, quantitative scoring system that integrates multi-dimensional geometric features (precise distance, orientation, canal wall morphology, degree of encapsulation, etc.), and cannot output a unified, probabilistic risk level, which is detrimental to the construction of clinical decision support systems and the standardized selection of surgical procedures.

[0007] In summary, there is an urgent need for an automated, highly accurate, and quantifiable method to assess the positional relationship and risk of damage between the wisdom tooth and the mandibular nerve canal. This method would overcome the subjectivity of manual operation, improve the accuracy of segmentation and measurement, achieve objective judgment of complex positional relationships, and ultimately provide accurate and reliable structured decision-making basis for clinical practice by integrating multi-dimensional features to construct an objective risk prediction model. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for quantitative assessment of the risk of wisdom teeth and mandibular nerve canal based on spatial distance field.

[0009] The objective of this invention is achieved through the following technical solution: a method for quantitatively assessing the risk of wisdom teeth and mandibular nerve canals based on spatial distance fields, comprising the following steps: S1. Based on the nnU-Net neural network, train and infer CBCT images, and output a semantic segmentation mask containing high-risk areas of mandibular wisdom teeth, mandibular wisdom pulp and mandibular nerve canal. S2. Separate instances from the semantic segmentation mask to obtain instance masks for the mandibular nerve canal, mandibular wisdom tooth, and mandibular wisdom tooth pulp. S3. Based on the instance mask of the mandibular nerve canal, extract the center line of the mandibular nerve canal and generate the corresponding symbolic distance field SDF; based on the instance mask of the mandibular wisdom tooth and the pulp of the mandibular wisdom tooth, calculate the minimum distance from its voxel to the symbolic distance field SDF of the mandibular nerve canal to determine the contact type. S4. Based on the mandibular canal centerline, a dynamic local coordinate system is constructed. Combining the minimum distance, the position of the pulp centroid of the mandibular wisdom tooth, and the geometric features extracted from the symbolic distance field SDF of the mandibular nerve canal, a multi-dimensional risk analysis is performed. The probability of nerve damage is calculated based on a regularized risk scoring model and mapped to a clinical risk level, outputting a structured risk assessment report.

[0010] Preferably, step S1 further includes the following sub-steps: Obtain the CBCT dataset containing annotation information, filter and remap the labels in the CBCT dataset, retain the labels corresponding to the mandibular wisdom tooth, mandibular wisdom tooth pulp and mandibular nerve canal, and convert them into continuous integers; Next, the CBCT dataset is preprocessed, which includes extracting dataset fingerprints to automatically configure network parameters, cropping images based on non-zero bounding boxes, resampling all images to the median voxel spacing, and normalizing image intensity. An nnU-Net neural network is constructed and trained. The nnU-Net neural network includes an input layer, an encoding downsampling layer, a decoding upsampling layer, and an output layer. The loss function of the nnU-Net neural network is a weighted average of the Dice loss and the cross-entropy loss. The nnU-Net neural network is trained using the AdamW optimizer. The preprocessed CBCT image is input into the trained nnU-Net neural network, which outputs a semantic segmentation mask for high-risk regions, used for subsequent instance separation and risk quantification.

[0011] Preferably, step S2 further includes the following sub-steps: Median filtering is applied to the semantic segmentation mask to remove salt-and-pepper noise; An opening operation is performed on the semantic segmentation mask after median filtering to smooth the boundaries and remove pinholes and isolated noise points; The semantic segmentation mask after the opening operation is repaired with a hole filling algorithm to fix internal defects. It is based on flood filling to traverse voxels to ensure the compactness of the entity's interior. Next, a three-dimensional connected component analysis is performed to separate independent instance components; Based on a preset anatomical adaptive voxel threshold, filter noisy components in the instance components; The filtered instance components are renumbered and mapped to the standard tooth positions in the FDI tooth position representation.

[0012] Preferably, the extraction of the mandibular canal centerline employs a 3D skeletonization algorithm to refine the binary mask of the mandibular nerve canal, and extracts the longest path as the main centerline based on the Dijkstra algorithm; the signed distance field SDF of the mandibular nerve canal is generated by subtracting the internal distance field from the external distance field, wherein the value of the external distance field is positive and the value of the internal distance field is negative.

[0013] Preferably, the criteria for determining the contact type are as follows: if the minimum distance is greater than or equal to 2mm, it is separation; if the minimum distance is greater than or equal to 0mm and less than 2mm, it is contact; if the minimum distance is less than 0mm, it is intrusion.

[0014] Preferably, the dynamic local coordinate system uses the tangent vector of the closest point to the target voxel on the mandibular canal centerline as the anterior-posterior direction, and performs a cross product between the centerline tangent vector and the occlusal plane normal vector to obtain the buccal-lingual direction vector; by calculating the projection of the target voxel's vector relative to the nearest point onto the buccal-lingual direction vector, it is determined whether the target voxel is located on the buccal or lingual side; by comparing the height of the pulp centroid of the mandibular wisdom tooth with the mandibular canal centerline, and introducing a preset apical correction distance, it is further assisted in determining whether the target voxel is located on the apical or root side.

[0015] Preferably, the multi-dimensional risk analysis includes: determining the cortical continuity of the mandibular nerve canal wall based on the SDF gradient modulus; calculating the length of the perpendicular line from the target voxel to the center of the mandibular nerve canal through the canal wall using the ray projection method, and comparing it with the local cross-sectional radius to determine whether there is apical encasement; and determining whether there is canal stenosis by calculating the equivalent radius of the local cross-section of the mandibular nerve canal.

[0016] Preferably, the rule-based risk scoring model is a logistic regression model, and the structured risk assessment report includes at least the following information: minimum distance, contact type, directional relationship, detection results of high-risk signs, probability of nerve damage, and corresponding clinical risk level.

[0017] The beneficial effects of this invention are: 1) By integrating a specialized-trained nnU-Net segmentation model, automated instance separation and post-processing, and geometric quantization analysis based on symbolic distance field (SDF), an end-to-end automated closed loop was constructed from raw CBCT image input to structured risk report output. This avoids the high dependence of traditional methods on physicians' manual sketching, measurement, and subjective judgment, freeing physicians from tedious and repetitive manual operations. It not only significantly shortens the assessment time but also eliminates the fluctuations in results caused by differences in operator experience and subjectivity, ensuring the objectivity, repeatability, and standardization of assessment results.

[0018] 2) This invention first employs the nnU-Net model, fine-tuned for high-risk areas of the oral cavity (mandibular wisdom teeth, pulp, and neural tube), to achieve high-precision semantic segmentation. Then, through a series of morphological post-processing steps such as median filtering, opening operations, and hole filling, combined with anatomically based adaptive voxel thresholding and 3D connected component analysis, it effectively separates independent instances such as the left and right neural tubes, multiple wisdom teeth, and their pulps, while filtering out image noise and segmentation artifacts. This strategy effectively solves the problems of poor segmentation of small and complex structures and multi-instance confusion in general segmentation models, laying a clean and reliable data foundation for subsequent accurate geometric quantization.

[0019] 3) A dynamic local coordinate system based on the neural canal centerline can adapt to the curvature of the neural canal. Combined with apical correction assisted by the pulp centroid, it can accurately and automatically determine the buccal / lingual and apical / lateral orientation of the wisdom tooth and the neural canal. This method overcomes the coarseness and inaccuracy of orientation determination based on a global coordinate system or manual reference plane, and is particularly suitable for the assessment of various complex impacted conditions.

[0020] 4) This invention goes beyond simple distance measurement. It integrates minimum distance, precise location, and multiple high-risk geometric features derived from the SDF (such as cortical continuity of the canal wall, apical occlusion ratio, and degree of canal stenosis) to construct a quantitative risk model based on logistic regression. This model transforms objective features supported by various medical literature into a unified probability value for nerve injury, mapping it to "high," "medium," and "low" clinical risk levels. This systematic and quantitative assessment provides surgeons with rich and interpretable decision support far exceeding traditional qualitative descriptions or single-threshold judgments. It helps in developing more individualized and risk-controlled surgical plans (such as choosing conventional tooth extraction, crown resection, or staged extraction), thereby maximizing the reduction of nerve injury risk preoperatively. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0023] See Figure 1This invention provides a technical solution: a method for quantitatively assessing the risk of wisdom teeth and mandibular nerve canal based on spatial distance field, comprising the following steps: S1. Based on the nnU-Net neural network, train and infer CBCT images, and output a semantic segmentation mask containing high-risk areas of mandibular wisdom teeth, mandibular wisdom pulp and mandibular nerve canal. S2. Separate instances from the semantic segmentation mask to obtain instance masks for the mandibular nerve canal, mandibular wisdom tooth, and mandibular wisdom tooth pulp. S3. Based on the instance mask of the mandibular nerve canal, extract the center line of the mandibular nerve canal and generate the corresponding symbolic distance field SDF; based on the instance mask of the mandibular wisdom tooth and the pulp of the mandibular wisdom tooth, calculate the minimum distance from its voxel to the symbolic distance field SDF of the mandibular nerve canal to determine the contact type. S4. Based on the mandibular canal centerline, a dynamic local coordinate system is constructed. Combining the minimum distance, the position of the pulp centroid of the mandibular wisdom tooth, and the geometric features extracted from the symbolic distance field SDF of the mandibular nerve canal, a multi-dimensional risk analysis is performed. The probability of nerve damage is calculated based on a regularized risk scoring model and mapped to a clinical risk level, outputting a structured risk assessment report.

[0024] In this embodiment, focusing on high-risk clinical structures (third molar, pulp, and IAC), a three-pronged strategy of "specialized model + geometric quantification + risk scoring" is adopted: nnU-Net v2 is used for specialized fine-tuning to focus on high-risk areas, improving segmentation accuracy and computational efficiency; SDF is introduced in conjunction with a dynamic centerline coordinate system and pulp-assisted judgment to achieve sub-voxel-level distance and direction quantification; a rule-based risk scoring model is constructed based on quantified indicators, forming an end-to-end automated prediction closed loop. This provides an automated and interpretable solution to the problems of existing methods relying on manual measurement, ambiguous direction judgment, and subjective risk assessment, demonstrating significant clinical translational potential. Through automated processes, end-to-end processing is achieved from cone-beam computed tomography (CBCT) image input to structured risk assessment report output, providing objective, accurate, and quantifiable preoperative risk prediction for clinicians.

[0025] In some embodiments, S1 further includes the following sub-steps: Obtain the CBCT dataset containing annotation information, filter and remap the labels in the CBCT dataset, retain the labels corresponding to the mandibular wisdom tooth, mandibular wisdom tooth pulp and mandibular nerve canal, and convert them into continuous integers; Next, the CBCT dataset is preprocessed, which includes extracting dataset fingerprints to automatically configure network parameters, cropping images based on non-zero bounding boxes, resampling all images to the median voxel spacing, and normalizing image intensity. An nnU-Net neural network is constructed and trained. The nnU-Net neural network includes an input layer, an encoding downsampling layer, a decoding upsampling layer, and an output layer. The loss function of the nnU-Net neural network is a weighted average of the Dice loss and the cross-entropy loss. The nnU-Net neural network is trained using the AdamW optimizer. The preprocessed CBCT image is input into the trained nnU-Net neural network, which outputs a semantic segmentation mask for high-risk regions, used for subsequent instance separation and risk quantification.

[0026] In this embodiment, S1, specialized semantic segmentation based on nnU-Net neural network: The goal of this step is to segment key high-risk anatomical structures from the original CBCT image with high precision, namely the mandibular third molar (wisdom tooth), its internal pulp, and the mandibular nerve canal (IAC). S1 can be divided into the following sub-steps: S11. Data Preparation and Preprocessing: First, obtain an open-source oral CBCT dataset (e.g., the ToothFairy3 dataset, containing 532 CBCT cases). Filter and remap the dataset labels, retaining only labels related to the left / right mandibular wisdom teeth (FDI numbers 38 / 48), their corresponding pulp (138 / 148), and the left / right mandibular nerve canals (3 / 4), and mapping them to consecutive integer labels (e.g., 0-6). This step primarily aims to focus the model on learning high-risk regions and shield it from interference from irrelevant background structures. Next, perform standardization preprocessing on the dataset, including: Data fingerprint extraction: Automatically calculates the image size, voxel spacing distribution, foreground intensity statistics, etc. in the dataset, which are used for subsequent automatic configuration of network parameters.

[0027] Image cropping: Based on non-zero bounding boxes, all images are cropped to the smallest region containing only the foreground label (i.e., the high-risk structure mentioned above), so as to significantly reduce the computational load and focus on key regions.

[0028] Resampling: All images are resampled to the central voxel spacing in the dataset. Third-order spline interpolation is used for the image data, and nearest-neighbor interpolation is used for the segmentation labels to ensure spatial consistency. For data with excessive anisotropy (such as a spacing on one axis being more than 3 times that of other axes), an independent Z-axis resampling strategy is used to avoid interpolation artifacts.

[0029] Intensity normalization: For CT modalities, a truncation normalization based on foreground intensity statistics is used (e.g., truncating intensity values ​​from the 0.5% to 99.5th percentile), followed by z-score normalization, with background voxels set to zero. This enhances the model's robustness to intensity differences across different scanning devices and protocols.

[0030] After preprocessing, the data is stored in .npz format for efficient loading.

[0031] S12. Neural Network Construction and Training: Construct and train a specialized nnU-Net v2 network. This network employs a classic encoder-decoder architecture with skip connections. Both the encoder and decoder consist of multiple residual blocks, each containing a 3x3x3 convolutional layer, instance normalization, and a Leaky ReLU activation function (slope of 0.01). Downsampling is achieved through max pooling with a stride of 2, and upsampling is achieved through transposed convolutions with a stride of 2. The network's output layer uses 1x1x1 convolutions to convert the number of feature map channels into the number of segmentation classes, and generates a probability map for each voxel belonging to each class using the Softmax function. The loss function is a weighted average of Dice loss and cross-entropy loss, for example: L = 0.5 * L_Dice + 0.5 * L_CE, to simultaneously optimize region overlap and pixel-level classification accuracy. The optimizer is AdamW, with an initial learning rate of 3e-4, and a multinomial learning rate scheduling strategy for dynamic decay. During training, 80% of the data is used for training and 20% for validation, with an early stopping strategy (training stops if there is no performance improvement after 10 consecutive validation cycles). After training, the preprocessed CBCT images are input into the trained model to output a semantic segmentation mask containing the three types of high-risk structures mentioned above.

[0032] In some embodiments, S2 further includes the following sub-steps: Median filtering is applied to the semantic segmentation mask to remove salt-and-pepper noise; An opening operation is performed on the semantic segmentation mask after median filtering to smooth the boundaries and remove pinholes and isolated noise points; The semantic segmentation mask after the opening operation is repaired with a hole filling algorithm to fix internal defects. It is based on flood filling to traverse voxels to ensure the compactness of the entity's interior. Next, a three-dimensional connected component analysis is performed to separate independent instance components; Based on a preset anatomical adaptive voxel threshold, filter noisy components in the instance components; The filtered instance components are renumbered and mapped to the standard tooth positions in the FDI tooth position representation.

[0033] In this embodiment, the separation of instances of S2, mandibular wisdom teeth, dental pulp, and mandibular nerve canal: Since the semantic segmentation mask output by S1 cannot distinguish different instances of the same structure (such as the left and right mandibular nerve canals), this step is mainly used for post-processing of the mask to separate independent instance objects. S2 can be divided into the following sub-steps: S21. Morphological Post-processing: First, apply mean filtering (e.g., 3x3x3 neighborhood) to the semantic segmentation mask to remove salt-and-pepper noise from the image. Next, perform opening operations (erosion followed by dilation) to smooth structural boundaries, remove small holes and isolated noise points, while maintaining the integrity of the main shape. Then, use a hole-filling algorithm to fill the holes inside the structure, ensuring the accuracy of subsequent volume calculations and connected component analysis.

[0034] S22. 3D Connectivity Analysis and Instance Labeling: For the post-processed binary mask, 26-neighborhood connectivity analysis (e.g., using the `cc3d.connected_components_3d` function) is used to label all connected voxel regions, thereby separating independent instance components. Anatomically adaptive voxel thresholds are set for filtering based on prior knowledge of the volume of different anatomical structures. For example, the volume of a wisdom tooth instance can be set to be greater than 1000 voxels, the volume of a mandibular nerve canal instance can be greater than 400 voxels, and the volume of a dental pulp instance can be greater than 300 voxels. Components with volumes smaller than the corresponding thresholds are considered noise and filtered out. Finally, a unique identifier is assigned to each retained instance component and mapped to the International Dental Federation (FDI) standard tooth position numbering system. For example, the left mandibular nerve canal is labeled "3", and the right mandibular wisdom tooth is labeled "48", forming clean, instance-level mask data for subsequent independent geometric analysis.

[0035] In some embodiments, the extraction of the mandibular canal centerline employs a 3D skeletonization algorithm to refine the binary mask of the mandibular nerve canal, and extracts the longest path as the main centerline based on the Dijkstra algorithm; the signed distance field SDF of the mandibular nerve canal is generated by subtracting the internal distance field from the external distance field, wherein the value of the external distance field is positive and the value of the internal distance field is negative.

[0036] In some embodiments, the criteria for determining the contact type are as follows: if the minimum distance is greater than or equal to 2 mm, it is separation; if the minimum distance is greater than or equal to 0 mm and less than 2 mm, it is contact; if the minimum distance is less than 0 mm, it is intrusion.

[0037] In this embodiment, S3, mandibular nerve canal centerline extraction and hybrid geometry modeling, is mainly used to construct an accurate centerline representation and continuous spatial distance field for each mandibular nerve canal, providing a foundation for subsequent distance calculations. S3 can be divided into the following steps: S31. Centerline Extraction: For each binary mask of the mandibular nerve canal obtained in step S2, a 3D skeletonization algorithm (such as skimage.morphology.skeletonize_3d) is used to refine it, resulting in an initial skeleton with a single-voxel width. Then, the skeleton points are converted into a graph structure, with nodes as skeleton voxels and edge weights equal to the physical Euclidean distance between nodes. Next, Dijkstra's algorithm is used to find the longest path in the graph, which is taken as the main centerline of the mandibular nerve canal. Short pseudo-branches are automatically pruned to ensure that the centerline accurately reflects the main trunk direction of the nerve canal.

[0038] S32. Symbolic Distance Field (SDF) Generation: The symbolic distance field of the mandibular nerve canal mask is calculated. The value of the SDF at any point in space is defined as the shortest distance from that point to the mask surface. It is generally stipulated that the distance is positive when the point is outside the mask, negative when it is inside, and zero when it is on the surface. It can be obtained by calculating the distance transformation field (dist_out) outside the mask and the distance transformation field (dist_in) inside the mask, and then subtracting them: SDF = dist_out - dist_in. The SDF provides continuous distance information with sub-voxel precision.

[0039] S33. Distance Calculation and Contact Type Determination: For each mandibular wisdom tooth and its pulp mask separated in step S2, traverse all voxels and use SDF to quickly query the distance from each voxel to the nearest mandibular nerve canal surface. Take the minimum value among all distances as the minimum distance from the wisdom tooth / pulp to the corresponding nerve canal. Based on this minimum distance, define the contact type: If the minimum distance is ≥ 2.0 mm, it is judged as "separated".

[0040] If 0 mm ≤ minimum distance < 2.0 mm, it is determined as "contact".

[0041] If the minimum distance is less than 0 mm, it is judged as "intrusion" (i.e., the wisdom tooth protease is located inside the neural canal).

[0042] In some embodiments, the dynamic local coordinate system uses the tangent vector of the closest point to the target voxel on the mandibular canal centerline as the anterior-posterior direction, and performs a cross product of the centerline tangent vector and the occlusal plane normal vector to obtain the buccal-lingual direction vector; by calculating the projection of the target voxel's vector relative to the nearest point onto the buccal-lingual direction vector, it is determined whether the target voxel is located on the buccal or lingual side; by comparing the height of the pulp centroid of the mandibular wisdom tooth with the mandibular canal centerline, and introducing a preset apical correction distance, it is assisted in determining whether the target voxel is located on the apical or root side.

[0043] In some embodiments, the multidimensional risk analysis includes: determining the cortical continuity of the mandibular nerve canal wall based on the SDF gradient modulus; calculating the length of the perpendicular line from the target voxel to the center of the mandibular nerve canal through the canal wall using the ray projection method, and comparing it with the local cross-sectional radius to determine whether there is apical encasement; and determining whether there is canal stenosis by calculating the equivalent radius of the local cross-section of the mandibular nerve canal.

[0044] In some embodiments, the rule-based risk scoring model is a logistic regression model, and the structured risk assessment report includes at least the following information: minimum distance, contact type, directional relationship, detection results of high-risk signs, probability of nerve damage, and corresponding clinical risk level.

[0045] In this embodiment, S4, the construction of an anatomical morphology coordinate system and risk quantification, comprehensively utilizes geometric information to construct a local coordinate system to determine orientation, extract multidimensional risk features, and synthesize the final risk assessment result. S4 can be divided into the following steps: S41. Construction of Dynamic Local Coordinate System and Orientation Determination: Buccal-lingual orientation determination: For any voxel to be evaluated on the wisdom tooth, find the nearest point on the central line of the mandibular nerve canal. Calculate the tangent vector (approximately anterior-posterior direction) of the central line at that point. Perform a cross product of this tangent vector and a global occlusal plane normal vector (e.g., [0, 0, 1]) to obtain the local buccal-lingual direction vector at that point. Calculate the vector from the nearest point on the central line to the wisdom tooth voxel and project it onto this buccal-lingual direction vector. Based on the sign of the projection value, the location of the voxel relative to the nerve canal—whether buccal or lingual—can be automatically and accurately determined.

[0046] Apical / lateral determination: Calculate the centroid coordinates of the pulp inside the mandibular wisdom tooth. By comparing the height of this centroid with the center line of the mandibular nerve canal in the vertical direction (e.g., along the long axis of the jawbone), and introducing a preset apical correction distance based on anatomy (e.g., 0.70 mm), it is possible to help determine whether a specific part of the wisdom tooth (e.g., the root apex) is closer to the apical or lateral side of the nerve canal. This solves the problem that in complex impacted cases, it is difficult to accurately determine the apical relationship based solely on the geometric center.

[0047] S42. High-risk geometric feature extraction: Automated extraction of multiple geometric features related to nerve injury risk from the SDF and mask of the mandibular nerve canal: Cortical interruption detection: Calculate the gradient of the SDF in the region near the surface of the neural canal. An abnormally increased gradient modulus (e.g., exceeding the threshold of 0.8) may suggest thinning or interruption of the cortical bone of the canal wall at that location, which is one of the high-risk signs.

[0048] Periapical wrapping analysis: For a wisdom tooth voxel that is "invasive" or "closely contacting," a ray is emitted from the voxel towards its foot on the central line of the mandibular nerve canal. The length of this ray traveling within the nerve canal (internal length) is calculated by sampling the SDF value. This internal length is compared to the equivalent radius of the local cross-section of the nerve canal at that location. If the ratio of internal length to canal diameter exceeds a threshold (e.g., >0.8), it suggests a risk of apical curvature wrapping around the nerve canal.

[0049] Nerve stenosis detection: Calculate the cross-sectional area at multiple locations along the central line of the mandibular nerve canal (e.g., using the section calculation function of the VTK library). Calculate the local radius using the formula equivalent radius = sqrt(area / π). If the local radius is less than a safe threshold (e.g., 1.0 mm), nerve canal stenosis is identified at that location, making the nerve more susceptible to compression injury.

[0050] S43. Standardized Risk Scoring Model and Report Generation: Integrating all the above quantitative indicators, an interpretable risk scoring model based on logistic regression is constructed. First, a baseline log odds (base_log_odds) is set, corresponding to the overall average nerve injury rate of mandibular wisdom tooth extraction (e.g., based on literature, set to -3.47, corresponding to a base probability of approximately 3%). Then, an odds ratio (OR) based on clinical research evidence is assigned to each risk feature, and it is converted into a logarithmic weight. For example: Intrusion / Penetration (SDF<0): log(OR)=log(15.0) Extremely close contact (SDF < 0.5 mm): log(OR) = log(12.0) Lingual position: log(OR) = log(7.21) Cortical interruption in the vessel wall: log(OR) = log(4.5) Root tip wrapping: log(OR) = log(2.8) Narrow pipe: log(OR) = log(2.0) The baseline log-odds is summed with the log-weighted values ​​of all relevant features to obtain the total log-odds. Finally, the predicted probability P of nerve damage resulting from the wisdom tooth extraction is calculated using the logistic function P = 1 / (1 + exp(-total_log_odds)). This probability value is then mapped to a clinical risk level, for example: Low risk (P<10%): Can be removed using standard procedures.

[0051] Medium risk (10% ≤ P<30%): Detailed preoperative CBCT assessment is required, and careful operation is necessary during the procedure.

[0052] High risk (P ≥ 30%): It is strongly recommended to consider alternatives, such as coronectomy, and to obtain strong preoperative informed consent.

[0053] Ultimately, the system automatically generates a structured risk assessment report (such as CSV or JSON format), including case ID, minimum distance, contact type, location (buccal / lingual, apical / lateral), status of various high-risk signs, predicted probability of nerve damage, and corresponding clinical risk level and recommendations. This report can be directly integrated into the hospital information system to assist clinical decision-making.

[0054] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for quantitatively assessing the risk of wisdom teeth and mandibular nerve canal based on spatial distance field, characterized in that: Includes the following steps: S1. Based on the nnU-Net neural network, train and infer CBCT images, and output a semantic segmentation mask containing high-risk areas of mandibular wisdom teeth, mandibular wisdom pulp and mandibular nerve canal. S2. Separate instances from the semantic segmentation mask to obtain instance masks for the mandibular nerve canal, mandibular wisdom tooth, and mandibular wisdom tooth pulp. S3. Based on the instance mask of the mandibular nerve canal, extract the center line of the mandibular nerve canal and generate the corresponding symbolic distance field SDF; based on the instance mask of the mandibular wisdom tooth and the pulp of the mandibular wisdom tooth, calculate the minimum distance from its voxel to the symbolic distance field SDF of the mandibular nerve canal to determine the contact type. S4. Based on the mandibular canal centerline, a dynamic local coordinate system is constructed. Combining the minimum distance, the position of the pulp centroid of the mandibular wisdom tooth, and the geometric features extracted from the symbolic distance field SDF of the mandibular nerve canal, a multi-dimensional risk analysis is performed. The probability of nerve damage is calculated based on a regularized risk scoring model and mapped to a clinical risk level, outputting a structured risk assessment report.

2. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: S1 further includes the following sub-steps: Obtain the CBCT dataset containing annotation information, filter and remap the labels in the CBCT dataset, retain the labels corresponding to the mandibular wisdom tooth, mandibular wisdom tooth pulp and mandibular nerve canal, and convert them into continuous integers; Next, the CBCT dataset is preprocessed, which includes extracting dataset fingerprints to automatically configure network parameters, cropping images based on non-zero bounding boxes, resampling all images to the median voxel spacing, and normalizing image intensity. An nnU-Net neural network is constructed and trained. The nnU-Net neural network includes an input layer, an encoding downsampling layer, a decoding upsampling layer, and an output layer. The loss function of the nnU-Net neural network is a weighted average of the Dice loss and the cross-entropy loss. The nnU-Net neural network is trained using the AdamW optimizer. The preprocessed CBCT image is input into the trained nnU-Net neural network, which outputs a semantic segmentation mask for high-risk regions, used for subsequent instance separation and risk quantification.

3. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: The S2 step further includes the following sub-steps: Median filtering is applied to the semantic segmentation mask to remove salt-and-pepper noise; An opening operation is performed on the semantic segmentation mask after median filtering to smooth the boundaries and remove pinholes and isolated noise points; The semantic segmentation mask after the opening operation is repaired with a hole filling algorithm to fix internal defects. It is based on flood filling to traverse voxels to ensure the compactness of the entity's interior. Next, a three-dimensional connected component analysis is performed to separate independent instance components; Based on a preset anatomical adaptive voxel threshold, filter noisy components in the instance components; The filtered instance components are renumbered and mapped to the standard tooth positions in the FDI tooth position representation.

4. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: The extraction of the mandibular canal centerline employs a 3D skeletonization algorithm to refine the binary mask of the mandibular nerve canal, and extracts the longest path as the main centerline based on the Dijkstra algorithm; the signed distance field SDF of the mandibular nerve canal is generated by subtracting the internal distance field from the external distance field, where the value of the external distance field is positive and the value of the internal distance field is negative.

5. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: The criteria for determining the contact type are as follows: if the minimum distance is greater than or equal to 2mm, it is separation; if the minimum distance is greater than or equal to 0mm and less than 2mm, it is contact; if the minimum distance is less than 0mm, it is intrusion.

6. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: The dynamic local coordinate system uses the tangent vector of the closest point to the target voxel on the mandibular canal centerline as the anterior-posterior direction. The buccal-lingual direction vector is obtained by cross-product of the centerline tangent vector and the occlusal plane normal vector. The target voxel is located on the buccal or lingual side by calculating the projection of the vector of the target voxel relative to the nearest point onto the buccal-lingual direction vector. The height of the pulp centroid of the mandibular wisdom tooth and the mandibular canal centerline is compared, and a preset apical correction distance is introduced to help determine whether the target voxel is located on the apical or root side.

7. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: The multidimensional risk analysis includes: determining the cortical continuity of the mandibular nerve canal wall based on the SDF gradient modulus; calculating the length of the perpendicular line from the target voxel to the center of the mandibular nerve canal through the canal wall using the ray projection method, and comparing it with the local cross-sectional radius to determine whether there is apical encasement; and determining whether there is canal stenosis by calculating the equivalent radius of the local cross-section of the mandibular nerve canal.

8. The method for quantitative risk assessment of wisdom teeth and mandibular nerve canal based on spatial distance field according to claim 1, characterized in that: The rule-based risk scoring model is a logistic regression model, and the structured risk assessment report includes at least the following information: minimum distance, contact type, directional relationship, detection results of high-risk signs, probability of nerve damage, and corresponding clinical risk level.

Images