CBCT tooth segmentation method and system based on anatomical perception cascade network

Abstract

The invention discloses a CBCT tooth segmentation method and system based on an anatomical perception cascade network, and the method comprises the steps: a first stage, carrying out the simplified dichotomy segmentation based on an original CBCT image and a coarse segmentation network taking 3D U-Net as a trunk, outputting maxillary and mandibular tooth probability graphs, and taking the maxillary and mandibular tooth probability graphs as prior information to guide the generation of an SDM; in the second stage, the original CBCT image and the calibrated maxillary tooth probability graph and the calibrated mandibular tooth probability graph are spliced together, a formed multi-channel input tensor is input into a fine segmentation network, the fine segmentation network takes Residual U-Net as a trunk, and an improved AGBR module and an improved SDMAA module are integrated in an encoder-decoder architecture of the fine segmentation network; and the decoder fuses all refined and re-calibrated feature maps, and upsamples and reconstructs 42 types of instance segmentation results with correct topology and clear boundaries. According to the method, the problem that in the prior art, when the inherent and local boundary fuzzy defect in CBCT is overcome, an effective pertinence mechanism is lacked is solved, and precise and robust 42-class instance segmentation can be achieved.

Description

Technical Field

[0001] This invention relates to the field of oral imaging technology, and in particular to a CBCT (Cone Beam Computed Tomography) tooth segmentation method and system based on an anatomical perception cascade network. Background Technology

[0002] With the rapid development of digital dentistry, CBCT, with its isotropic high-resolution three-dimensional imaging capabilities, has become an indispensable core technology in oral and maxillofacial surgery. CBCT can clearly show the complex spatial relationships between key structures such as impacted wisdom teeth, adjacent teeth, and the inferior alveolar nerve canal, providing an anatomical basis that traditional two-dimensional panoramic radiographs cannot match for precise preoperative measurements and risk assessments, thereby helping doctors to develop better surgical path planning strategies.

[0003] Despite the rich 3D information provided by CBCT, it still has the following inherent limitations in automatic tooth segmentation tasks: Because medical staff require patients to close their mouths and bite down on the occlusal plate when collecting dental data, the CBCT images collected clinically are in a natural occlusal state. In a natural occlusal state, the occlusal contact surfaces of the upper and lower teeth have extremely low grayscale contrast due to physical contact, resulting in highly blurred crown boundaries, which can easily lead to segmentation and adhesion problems. Although the problem of tooth contact surface adhesion can be avoided in the non-occlusal state (such as open mouth), other types of segmentation errors still occur due to similar tooth shapes and close proximity. Due to the imaging principle and the characteristics of human tissue, the tooth root and the surrounding alveolar bone exhibit similar attenuation characteristics, resulting in low contrast between the two, which poses a challenge to accurately delineating the tooth root contour.

[0004] The inherent, localized boundary blurring defects in these CBCT scans mean that clinical analysis of CBCT scans still heavily relies on physician experience, resulting in inefficiency and high subjectivity.

[0005] In recent years, to overcome this technological bottleneck, artificial intelligence technology, especially deep learning methods based on CNN (Convolutional Neural Network), has been widely used to achieve automatic segmentation of teeth and bone structures. From the classic U-Net architecture and its extensions to three-dimensional data (such as 3D U-Net), to the highly adaptive nnU-Net framework, and AttentionUNet3D, which improves upon the shortcomings of 3D U-Net, all have demonstrated excellent performance in oral CBCT segmentation tasks.

[0006] Furthermore, architectures such as ToothNet, specifically designed for oral CBCT images, further improve the accuracy of tooth instance segmentation and recognition by introducing strategies such as edge feature maps and spatial topological relationship encoding. Specifically, ToothNet employs a carefully designed two-stage processing flow: In the first stage, edge feature maps of teeth are extracted from the volume data of the input raw CBCT images using specific image processing algorithms (such as edge detection operators). This step aims to enhance the saliency of tooth morphological boundaries. In the second stage, the extracted edge feature maps are concatenated with the original CBCT images and input into a 3D-based RPN (Region Proposal Network) instance segmentation framework, ultimately generating an accurate voxel-level mask for each tooth. This 3D RPN framework introduces a learnable similarity matrix mechanism, effectively filtering and removing redundant candidate boxes by calculating the feature similarity between candidate regions. This design significantly accelerates model training and reduces memory consumption. Furthermore, to address the ambiguity in identifying different teeth with similar morphological features, ToothNet encodes the spatial topological relationships between teeth as structured features and inputs them as additional prior information into the recognition module, thereby effectively improving the accuracy of tooth number recognition. Experimental results show that ToothNet surpasses other methods of its time in segmentation and recognition accuracy on publicly available CBCT tooth datasets, making it the first research work to successfully apply deep neural networks to CBCT tooth instance segmentation tasks. This work has been published at CVPR 2019, a top conference in the field of computer vision, and no related publicly available patent documents have been found to date.

[0007] While these methods have made significant progress in overall segmentation, they are mostly end-to-end global optimization architectures or lack specificity, and still lack effective targeted mechanisms when facing the inherent, local boundary ambiguity in CBCT. Below, we will explain traditional methods in two categories from the perspectives of common problems and specific problems.

[0008] Category 1: Common problems of general architectures.

[0009] Examples of models: classic U-Net, 3D U-Net, nnU-Net, and later Transformer architectures (such as UNETR, SwinUNETR) and hybrid architectures (such as 3DUX-Net).

[0010] Key drawback: These are end-to-end, general-purpose segmentation models. They are designed for overall segmentation accuracy, rather than specifically addressing localized pathological issues in tooth segmentation (such as intermaxillary occlusal adhesions).

[0011] Specific manifestations: such as Figure 1 As shown, these models are prone to undersegmentation, boundary confusion, or topological errors when dealing with low-contrast areas such as tooth interproximal surfaces and root apex. They lack a built-in "diagnosis-repair" mechanism to identify and specifically handle these difficult areas.

[0012] Type Two: Specific problems addressed by targeted methods.

[0013] Examples of models: ToothNet, BA-Net, EBA-Net, PMFSNet.

[0014] Core flaw: These models have recognized the boundary problem and attempted to solve it (e.g., ToothNet uses edge graphs, BA-Net uses boundary branches), but the solutions have limitations in terms of mechanism.

[0015] In summary, existing methods have the following main shortcomings: 1. Insufficient targeting of inherent defects in CBCT: U-Net and its variants, Transformer and other models are end-to-end holistic optimization architectures designed for overall segmentation accuracy, rather than specifically addressing specific, localized pathological issues in tooth segmentation; while ToothNet adopts a global, homogeneous enhancement strategy. Therefore, these existing segmentation models lack specific mechanisms to address the inherent defects of CBCT, making it difficult to simultaneously maintain global morphological consistency and finely depict local boundaries in complex scenarios. This leads to segmentation results that are prone to adhesion or topological errors in key areas, directly affecting the accuracy of subsequent analysis.

[0016] 2. Lack of Adaptability and Targeting in Boundary Refinement Strategies: While models like BA-Net improve overall edge quality, their boundary refinement strategies are typically uniform, applying equal attention and computational resources to all boundary regions. ToothNet fails to provide targeted and adaptive modeling and optimization for the inherent limitations of CBCT; its network architecture lacks targeted optimization modules and strong geometric constraints. Therefore, these existing boundary-sensitive segmentation models fail to identify and focus on the most uncertain and challenging fuzzy boundaries within the model itself, resulting in unreasonable allocation of computational resources and limited optimization effects on the most difficult-to-segment regions.

[0017] 3. Lack of explicit modeling and utilization of segmentation uncertainty: End-to-end models such as U-Net, V-Net, and AttentionUNet3D usually directly output segmentation results without effectively quantifying and utilizing the uncertainty in the model prediction process. As a result, they cannot self-diagnose which regions have segmentation difficulties, and therefore cannot use this uncertainty information as a feedback signal to guide subsequent segmentation refinement or risk warning steps, which limits the robustness and interpretability of the model in challenging cases.

[0018] In summary, when faced with the inherent, localized boundary ambiguity inherent in CBCT, existing technologies struggle to improve overall recognition rates while ensuring precise segmentation and topological accuracy of local boundaries, particularly at the occlusal contact surfaces and root contours. They cannot guarantee that each segmented tooth is independent and separate in three-dimensional space, nor can they ensure the correct relative positional and connectivity relationships between teeth. Therefore, developing new intelligent segmentation methods capable of explicit modeling and resolving boundary ambiguity while incorporating anatomical priors has become an inevitable trend driving digital diagnostic systems towards higher precision. Summary of the Invention

[0019] This invention provides a CBCT tooth segmentation method and system based on an anatomical perception cascade network, which solves the problem that existing technologies lack effective targeted mechanisms when facing the inherent, local boundary ambiguity defects in CBCT, and can achieve accurate and robust segmentation of 42 types of instances.

[0020] One embodiment of the present invention provides a CBCT tooth segmentation method based on an anatomical perception cascade network. The anatomical perception cascade network adopts a two-stage segmentation framework, and the method includes: In the first stage, based on the original CBCT images, a simplified binary classification segmentation is performed using a coarse segmentation network with 3D U-Net as the backbone, outputting probability maps of maxillary and mandibular teeth, which are used as prior information to guide the generation of SDM. In the second stage, the original CBCT images are stitched together with the calibrated probabilistic maps of the maxilla and mandible teeth. The resulting multi-channel input tensor is then fed into a fine segmentation network. This network uses Residual U-Net as its backbone and integrates AGBR and SDMAA modules in its encoder-decoder architecture to perform the following operations: At the bottleneck layer, an ambiguity map is first calculated based on the probability maps of maxillary and mandibular teeth. The AGBR module guided by the ambiguity map is then used to perform gated refinement of the deepest high semantic feature map of the encoder, thereby targeting and eliminating the segmentation adhesion problem. On the jump connection paths at each scale, the SDM-guided SDMAA module is used to generate a spatial attention map based on the multi-scale feature map, followed by anatomical weighted pooling, and then feature excitation and recalibration to enhance shape consistency. Finally, the decoder merges all the refined and recalibrated feature maps and upsamples to reconstruct the topologically correct and clearly defined 42-class instance segmentation results.

[0021] Optionally, the first stage of the 3D U-Net uses the SGD optimizer with an initial learning rate of 0.01, momentum of 0.99, and a poly learning rate decay strategy, and is trained for 500 epochs.

[0022] Optionally, the calculation formula for the ambiguity map is as follows: , , In the above formula, Foreground probability, This is a probability diagram of maxillary teeth. This is a probability diagram of mandibular teeth. This is an ambiguous image.

[0023] Optionally, the gating fusion formula for the AGBR module is as follows: , In the above formula, This is the final refined feature map output. The input is a high semantic feature map. It is a hyperparameter that controls the refinement intensity. It is a binary gated mask. This represents element-wise multiplication. This is the refined feature map after gating.

[0024] Optionally, the hyperparameter controlling the refinement intensity The specific value is 0.5, and the binary gated mask The high uncertainty region is specifically the range where the ambiguity value is between 0.94 and 0.99.

[0025] Optionally, the direct calculation formula for the spatial attention map is as follows: , In the above formula, This is a spatial attention map. For spatial voxel indexing, It is an exponential transformation function. For the normalized SDM, It is a learnable temperature coefficient.

[0026] Optionally, the adapter calculation method for the spatial attention map specifically includes: projecting the SDM onto multi-channel features through a 1×1 convolutional layer, and then calculating the average absolute value of the channel dimensions to generate the adapter. .

[0027] Optionally, the anatomical weighted pooling step specifically involves: using a spatial attention map as spatial weights to perform pooling on the feature map at the current scale within the multi-scale feature map. A weighted average is performed to generate an anatomically perceived channel descriptor vector. In order to prioritize features located near anatomical boundaries when assessing the importance of channels.

[0028] Optionally, during the training process in the second phase, a composite loss function is also used to penalize activation of anatomically inappropriate regions. The formula for the composite loss function is as follows: , In the above formula, For composite loss function, and To balance the hyperparameters of the two loss functions, The loss function for LCE, This is the loss function for LDC.

[0029] On the other hand, embodiments of the present invention also provide a CBCT tooth segmentation system based on an anatomical sensing cascade network, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned CBCT tooth segmentation method based on an anatomical sensing cascade network.

[0030] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: To address the topological problems of tooth erroneous "adhesion" in the segmentation results of existing CBCT tooth segmentation methods when processing images of natural occlusion, such as low gray-scale contrast and blurred boundaries of the contact surfaces of the upper and lower teeth, and the lack of explicit modeling of boundary uncertainties and effective utilization of anatomical shape priors in existing deep learning models, this invention proposes an anatomically-aware cascaded network. This network innovatively integrates a two-stage segmentation framework from coarse to fine, an AGBR module, and a SDMAA module to construct an end-to-end cascaded network, forming a complete workflow from CBCT image input to simulation output. It explicitly identifies and quantifies blurred boundary regions during the segmentation process and utilizes a shape-prior-guided feature learning mechanism to adaptively focus on uncertain regions for refined segmentation. This significantly reduces tooth adhesion, improves the accuracy of segmentation boundaries and the correctness of topological structures, and enables high-precision 3D tooth instance segmentation and reconstruction of key maxillofacial structures. This targeted optimization strategy not only enables intelligent allocation of computing resources, but also provides a reliable model foundation for downstream digital dental applications such as quantitative calculation, analysis and planning, risk assessment, orthodontic planning, implant navigation and surgical simulation, significantly improving the accuracy and safety of clinical decision-making.

[0031] To overcome the shortcomings of traditional models, the technical means employed in this invention will be described in a list below.

[0032]

[0033] 1. Two-stage segmentation framework: The first stage performs a simplified binary classification task, generating probability maps of maxillary and mandibular teeth. These maps are used as prior information to guide the generation of the SDM (Segmented Data Map), quantitatively identifying and explicitly recognizing regions with blurred boundaries. The second stage uses the original CBCT images and the probability maps of maxillary and mandibular teeth as multi-channel inputs to perform refined segmentation of 42 classes of instances. This two-stage segmentation framework effectively solves the adhesion problem of maxillary and mandibular teeth at the occlusal contact surface, improving the topological accuracy of the segmentation.

[0034] Both this invention and ToothNet employ a two-stage network framework, rather than single-stage end-to-end models such as U-Net, nnUNet, UNETR, and SwinUNETR. This "two-stage task decomposition" approach is common. Despite the similar architectural paradigms, ToothNet's solution to the core problem of inherent CBCT imaging defects (especially blurred occlusal contact surfaces) differs fundamentally from this invention. This invention, while inheriting the "two-stage task decomposition" idea, undergoes a fundamental paradigm shift—from "feature enhancement" to "uncertainty-guided targeted optimization and deep fusion of shape constraints," thereby precisely tackling the core challenge of "intermaxillary adhesions in natural occlusion," a problem that ToothNet has failed to address effectively.

[0035] 2. AGBR Module: The AGBR module is integrated into the bottleneck layer of the second-stage fine segmentation network. It uses the generated ambiguity map as a spatial gating signal to adaptively concentrate feature refinement computation resources on the most uncertain boundary regions of the model (such as the occlusal contact surface). Through the gating mask mechanism, targeted refinement is performed only on highly ambiguous regions, significantly reducing the segmentation adhesion problem and improving boundary segmentation accuracy.

[0036] 3. SDMAA Module: This invention introduces SDM (Signed Distance Map) as an explicit shape prior and designs the SDMAA module, embedding it at the jump connections of each scale. The SDMAA module replaces traditional global average pooling with anatomically weighted pooling to generate a spatial attention map. During feature recalibration, it prioritizes features near the anatomical boundaries, thereby enhancing the structural consistency and boundary smoothness of the segmentation results.

[0037] Specifically, anatomically, teeth are rigid structures with smooth, continuous surfaces. SDM assigns a signed distance value to each voxel in 3D space, precisely quantifying the geometric relationship from that point to the tooth surface (positive for the exterior, zero for the surface, and negative for the interior). This mathematical description explicitly encodes the key shape prior: "a tooth is an internally continuous, smoothly bounded entity." When Residual U-Net learns to fuse this key shape prior information, it is guided to generate segmentation results with similar topological properties, thus naturally resisting topological errors that lead to segmentation adhesion problems (such as unreasonable holes, breaks, or merging). SDM's powerful geometric representation capabilities can directly compensate for the inherent limitations of CBCT imaging. Furthermore, SDM's differentiable and dense data format can be innovatively and deeply integrated with deep learning architectures into SDMAA modules, thereby achieving a leap from passively relying on image data to actively utilizing anatomical knowledge.

[0038] Traditional global average pooling assigns equal weights to all spatial locations, assuming all regions contribute equally to the evaluation of channel importance. This ignores a crucial fact in medical image segmentation tasks: features near anatomical boundaries are essential for accurate segmentation, while regions inside objects or far from boundaries contribute less to distinguishing the importance of different channels. To address this, this invention uses anatomically weighted pooling. Through an attention map generated by SDM, geometric priors are transformed into spatial weights. The evaluation criterion shifts from global intensity uniformity to focusing on boundary regions, directly addressing and alleviating segmentation adhesion and boundary blurring problems.

[0039] 4. Composite Loss Function: This invention not only introduces the AGBR and SDMAA modules into the Residual U-Net network structure, but also incorporates a composite loss function during training to penalize activation in anatomically unreasonable regions (such as predicting mandibular teeth within the maxilla), thereby forcibly imposing topological constraints at the output level and further improving the clinical rationality of the segmentation results.

[0040] Specifically, the composite loss function is a weighted sum of multi-class cross-entropy loss and soft Dice loss. Multi-class cross-entropy loss is used to measure the accuracy of voxel-level classification, while soft Dice loss is used to directly optimize the spatial overlap between the predicted segmentation results and the true labels. It is particularly effective in handling the class imbalance problem commonly found in tooth segmentation. Attached Figure Description

[0041] Figure 1 The images show segmentation results for traditional models (AttentionUNet3D, 3DUX-Net, UNETR, SwinUNETR, nnFormer, nnUNet, PMFSNet), this model, and ground truth labeled models. Figure 2 This is a schematic diagram of the framework of the anatomical perception cascade network of the present invention, which includes a first-stage and a second-stage segmentation network structure, an AGBR module, and an SDMSAA module; Figure 3 Three-view diagram of CBCT tooth segmentation results under the action of the AGBR module; Figure 4 This is a schematic diagram of the SDMAA module mechanism; Figure 5 A comparison chart of performance parameters and evaluation metrics of traditional models (AttentionUNet3D, 3DUX-Net, UNETR, SwinUNETR, nnFormer, nnUNet, PMFSNet) and our model (AACN); Figure 6 This is a schematic diagram showing the segmentation results for 42 types of instances. Detailed Implementation

[0042] This invention provides a CBCT tooth segmentation method and system based on an anatomical perception cascade network, which solves the problem that existing technologies lack effective targeted mechanisms when facing the inherent, local boundary ambiguity defects in CBCT, and can achieve accurate and robust segmentation of 42 types of instances.

[0043] First, the terms appearing in the instruction manual will be explained.

[0044] The study includes 42 categories of examples, comprising 32 teeth and 10 categories of key oral structures. The 10 categories of key oral structures include 2 categories of jawbones, 4 categories of tubular / cavitary structures, 2 categories of soft tissues, and 2 categories of restorations. For details, please refer to [reference needed]. Figure 6 , Figure 6 The segmentation results for 42 types of instances are shown.

[0045] SDM (Signed Distance Map) is a method used in medical imaging to describe the boundaries of anatomical structures. It is obtained by discretizing the data.

[0046] The core innovation of this invention lies in decomposing the complex task of segmenting 42 types of instances into two sequential and clearly defined sub-tasks. Specifically, it provides a two-stage segmentation framework from coarse to fine. In the first stage, a coarse segmentation network is used to coarsely locate the uncertainty of the boundary. In the second stage, a fine segmentation network is used, and targeted optimization is performed through ambiguity recognition and prior guidance to achieve fine segmentation.

[0047] To better understand, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described in this invention are only a part of the embodiments of this invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0048] like Figure 2 As shown, in the first stage (i.e. Figure 2 In Phase I of the process, a coarse segmentation network is used to coarsely locate the uncertainty of the boundary.

[0049] Specifically, the first stage uses the standard 3D U-Net architecture as the coarse segmentation network. The input is the volume data of the raw CBCT images from a single channel. A simplified binary classification segmentation task is performed, and the output is a probability map of maxillary teeth. Probability diagram of mandibular teeth The binary classification segmentation task treats all tooth categories, including maxillary and mandibular teeth, as the foreground and all non-tooth categories, such as jawbone and dental nerve canals, as the background. This significantly reduces task complexity, allowing 3D U-Net to focus on learning the macroscopic anatomy of the dental arch. The dental arch is the arch-shaped structure formed by the arrangement of maxillary and mandibular teeth on the alveolar bone, composed of teeth, alveolar bone, and gingiva, and is an important component of the oral masticatory system.

[0050] In the latter part of the first stage of training, an auxiliary supervision task is added to predict the SDM generated from the coarse segmentation labels. The SDM, which serves as the anatomical prior map, is used as the final output, along with the previously generated maxillary tooth probability map. Probability diagram of mandibular teeth This prior information is used to guide the generation of a more accurate SDM. This auxiliary supervision task aims to enable the coarse segmentation network to learn distance information of anatomical boundaries early on, thereby improving the probability map of maxillary teeth it generates. Probability diagram of mandibular teeth The accuracy of the boundary and the geometric rationality enable earlier and deeper integration of shape priors in the anatomical perception cascade network, forming a complete closed loop from coarse-grained shape perception to fine-grained shape correction, which can effectively enhance the anatomical consistency and robustness of the segmentation results.

[0051] Please continue to refer to this. Figure 2 In the second stage (i.e. Figure 2 In Stage II of the process, a fine-grained segmentation network is used, and targeted optimization is performed through ambiguity identification and prior guidance to achieve fine segmentation. The following sections will provide a detailed explanation from four aspects: input construction, network structure and task, integration of core innovative modules, and composite loss function.

[0052] 1. Input Construction: The probability map of maxillary teeth generated in the first stage. Probability diagram of mandibular teeth The features are transformed separately through a 1×1×1 convolutional layer with shared weights to obtain calibrated maxillary and mandibular prior features. Then, the volume data of the original CBCT images, the calibrated maxillary and mandibular prior features are concatenated along the channel dimension to form a multi-channel input tensor. This multi-channel input tensor is then input into the encoder of the second-stage network.

[0053] 2. Network Structure and Task: The fine segmentation network uses Residual U-Net as its backbone to perform the final 42-class instance segmentation task (32 teeth, upper and lower jaws, dental nerve canals, etc.) on multi-channel input tensors.

[0054] 3. Integration of core innovative modules: The fine segmentation network integrates two key modules: AGBR (Ambiguity-Gated Boundary Refiner) and SDMAA (SDM-Guided Anatomical Attention). These modules are used to target and solve the boundary ambiguity problem and enhance shape consistency. This is the key to the invention's superiority over existing end-to-end models (such as nnUNet and SwinUNETR) and single-policy models.

[0055] 3.1 AGBR Module Improvement Motivation: To address the problem that existing boundary refinement methods treat all boundaries equally and lack focus on highly challenging regions, the AGBR module utilizes ambiguity maps to dynamically and selectively refine features of highly uncertain boundary regions.

[0056] Input: High semantic feature map of the deepest layer of the encoder (Right now Figure 2 Feature 1), and the maxillary tooth probability map generated after the first stage. Probability diagram of mandibular teeth .

[0057] Specifically, high semantic feature maps This originates from the encoder end of the second-stage Residual U-Net. After the multi-channel input tensor undergoes successive convolutions and downsampling operations in the encoder, its spatial size is gradually compressed to a minimum, while the channel dimension increases to a maximum. The resulting high-semantic feature map... Located between the encoder and decoder, the bottleneck layer carries the most global and high-level semantic information aggregated by the network from the entire input region, in order to make decisions and solve the most fundamental problem of tooth adhesion at this highest semantic level.

[0058] The AGBR module, positioned at the bottleneck layer of the fine segmentation network, is located at the deepest feature level between the encoder and decoder, making the final decision on the fundamental category ambiguity caused by the occlusal contact surface. It dynamically activates a lightweight fine segmentation network using the ambiguity map, correcting features only in high-uncertainty boundary regions. This eliminates tooth adhesion problems globally at extremely low computational cost, ensuring topological correctness and guaranteeing that each segmented tooth is independent and separated in 3D space, while maintaining correct relative positional and connectivity relationships between teeth.

[0059] Workflow: Includes four steps: ambiguity map generation, gating mask generation, feature refinement, and gating fusion.

[0060] 3.1.1 Ambiguity Map Generation: Utilizing the maxillary tooth probability map from the first stage. Probability diagram of mandibular teeth Calculate the ambiguity graph The formula is as follows: , , In the above formula, Foreground probability.

[0061] In the above formula, and The value can be interpreted as the probability that the current voxel belongs to the maxillary tooth or the mandibular tooth, respectively. Both are close to 0.5, which is a strong signal from the algorithm level, indicating that the current position is the anatomical position where the image is blurred due to physical contact and segmentation adhesion is most likely to occur.

[0062] The value can be understood as the total probability that the voxel belongs to any tooth (i.e., the foreground).

[0063] It is a three-dimensional matrix with the same size as the input CBCT image, where each value The model's spatial location was quantified. The uncertainty of the segmentation is expressed in the range [0,1]. The higher the value, the more uncertain the model is about this segmentation, and the more difficult the segmentation is. Regions with high values ​​strongly point to the most intractable topological error risk area—the "adult-maxillary tooth adhesion problem"—that is, a boundary region with high uncertainty. Therefore, through... The value can help pinpoint the location where tooth adhesion problems are most likely to occur.

[0064] The above formula ensures that in areas where the probability of belonging to either the maxillary or mandibular tooth is close to 0.5, such as the occlusal surface of teeth where it is impossible to determine whether they belong to the maxillary or mandibular tooth, let... The value is close to 1, thus accurately representing the highly ambiguous boundary region between the upper and lower jaw teeth.

[0065] 3.1.2 Gated Mask Generation: A binary gated mask is generated by setting a high uncertainty interval. The mask is set to 1 only at highly blurred voxel locations, achieving precise positioning.

[0066] 3.1.3 Feature Refinement: Input High Semantic Feature Map (Right now Figure 2 Feature 1) is fed into a lightweight residual refinement network, which outputs a gated refined feature map. .like Figure 2 As shown, Conv3D and ReLU in the residual refinement network are standard and universal building blocks for deep neural networks. Conv3D (3D convolutional layer) is the core and fundamental operation for processing 3D volumetric data and is used to extract spatial features. ReLU (rectified linear unit) is one of the most commonly used nonlinear activation functions and is used to introduce nonlinear transformation capabilities into the network.

[0067] 3.1.4 Gated fusion: The final refined feature map output From the original high semantic feature map Refined feature map after gating The result is obtained by adding the ingredients together, and the calculation formula is as follows: , In the above formula, This represents element-wise multiplication. It is a hyperparameter that controls the refinement intensity.

[0068] The AGBR module ensures that computational resources are focused on solving the most challenging dental adhesion problems. Figure 3 The image shows three views of the CBCT tooth segmentation results generated by the AGBR module. The areas marked by the blue rectangles highlight the high-uncertainty boundary regions near the occlusal contact surfaces of the maxillary and mandibular teeth. This visually illustrates how the AGBR module can utilize the ambiguous information generated in the first stage (i.e., the probability map of the maxillary teeth). Probability diagram of occlusal teeth It can accurately locate and segment fuzzy regions, and adaptively enhance the local features of the region through a gating refinement mechanism, thereby directing computing resources to the most challenging segmentation problems.

[0069] The final refined feature map output by the AGBR module This feature map, directly used as the starting input to the decoder path in the second-stage fine-segmentation network, carries global semantic information optimized by the gating mechanism. It is then gradually restored to spatial details through upsampling operations in the decoder. During this process, each layer of the decoder simultaneously receives features from two sources: one from… The network employs two main methods: first, features derived from layer-by-layer upsampling that have resolved the macroscopic adhesion problem; and second, features from the corresponding layer of the encoder that have been calibrated based on shape priors by the SDMAA module at the same level. These two methods are fused in the decoder, enabling the network to ensure overall segmentation accuracy based on the output of the AGBR module, while simultaneously ensuring the precise anatomy of local boundaries with the guidance of the SDMAA module. This collaborative process generates the final high-quality segmentation result.

[0070] 3.2 SDMAA Module Improvement Motivation: Addressing the limitation of existing shape prior methods, which often impose constraints on the loss function or output layer and fail to directly guide feature learning, SDMAA innovatively integrates explicit tooth shape priors deep into the network's forward propagation process, enhancing the perception of anatomical boundaries at the feature level. The explicit tooth shape prior specifically refers to the SDM map, which is not implicit knowledge learned by the network during training but rather an explicit prior pre-calculated from high-quality real-world annotations. For each voxel in 3D space, the SDM value represents the distance from that point to the nearest tooth (or other target anatomical structure) surface. A positive SDM value indicates the point is outside the tooth; a zero SDM value indicates the point is exactly on the tooth surface, i.e., at the boundary; and a negative SDM value indicates the point is exactly inside the tooth.

[0071] Location: The SDMAA module is embedded at various hop connections in the decoder path and operates on multi-scale feature maps.

[0072] Input: The feature map at the current scale from the multi-scale feature map. ( Figure 2 Feature 2) originates from the output of each downsampling layer in the encoder path. During feature extraction, the network saves a copy of the intermediate output of each layer after passing through a block consisting of convolutions and downsampling. These feature maps at the current scale... With different spatial dimensions and semantic granularities, they are extracted separately through skip connections and input to SDMAA modules at each scale to achieve cross-scale, shape-prior-guided feature recalibration.

[0073] Spatial attention map generation: Given a feature map at the current scale And the corresponding SDM (calculated from real annotations in the first stage), first generate a spatial attention map. ( Figure 2 In the M_spatial plot, this plot has high weights near the object's surface (where the SDM value is close to 0), and can be implemented in two ways: 1) Direct calculation: Intermediate graph after normalization of SDM (Its zero-value surface corresponds to the anatomical surface) undergoes an exponential transformation to directly calculate the spatial attention map. The formula is as follows: , In the above formula, For spatial voxel indexing, It is a learnable temperature coefficient used to control the sharpness of focus of attention near the anatomical surface.

[0074] This method directly utilizes the geometric distance information of SDM, making the calculation efficient and possessing clear physical meaning.

[0075] 2) Adapter method: Project the SDM onto the multi-channel features through a 1×1 convolutional layer, and then calculate the mean absolute value of the channel dimensions to generate... This provides the model with the flexibility to learn the most relevant shape information.

[0076] Anatomical weighted pooling: This step is key to innovation; it uses... As spatial weights, for the feature map at the current scale A weighted average is performed to generate an anatomically perceived channel descriptor vector. This replaces the spatially insensitive Global Average Pooling (GAP) in the traditional SE module, forcing the network to prioritize features located near anatomical boundaries when evaluating channel importance.

[0077] Figure 4 This is a schematic diagram of the SDMAA module mechanism, showing the spatial attention map generated by the SDMAA module using SDM. Instead of global average pooling, this method prioritizes anatomical boundary features, with the ultimate goal of eliminating tooth adhesion issues during segmentation. Specifically, Figure 4 Using three classic medical image views (i.e., the three views of CBCT images displayed in the software: top view, front view, and side view), this visually demonstrates how the SDMAA module transforms the prior knowledge of tooth shape into spatial weights that can be used to guide the network: Axial slice (Z=200): A horizontal slice of the top view.

[0078] Coronal slice (Y=325): A vertical slice of the front view.

[0079] Sagittal slice (X=325): A vertical slice of a side view.

[0080] Background: Raw CBCT image.

[0081] Overlay heatmap (color): A color map converted from SDM values.

[0082] Dark areas (black): These represent regions where the SDM value is close to 0, i.e., the anatomical surface or boundary of the tooth. These areas have the highest spatial weight in the spatial attention map.

[0083] The SDMAA technology upon which this invention relies can generate a spatially accurate, anatomically sound heatmap that clearly distinguishes the boundaries of adjacent structures. This intuitively demonstrates that using this heatmap to guide network attention is an effective and direct technical approach to solving the problems of blurred tooth boundaries and segmentation adhesion in CBCT images.

[0084] Excitation and recalibration: Then through a bottleneck structure containing two fully connected layers (i.e.) Figure 2 In the MLP (Multi-Level Processing), this structure receives channel descriptor vectors generated by anatomical weighted pooling. Learn the non-linear dependencies between channels and output the channel attention weight vector. Ultimately, Compared with the original feature map (i.e., the feature map at the current scale) Perform channel-by-channel multiplication (i.e.) Figure 2 The output is a recalibrated feature map guided by shape priors (1×1×1×C). The recalibrated feature map is a feature map output by the SDMAA module that has been reweighted by channel attention guided by shape priors. Its core feature is the selective enhancement of channel dimensions, which strengthens feature channels related to anatomical boundaries.

[0085] Subsequent processing: The recalibrated feature map is passed to the corresponding layer of the decoder through skip connections. After being fused with the decoder features, it continues to participate in upsampling and semantic information reconstruction, ultimately helping to generate segmentation results with consistent anatomy and accurate boundaries.

[0086] In summary, the SDMAA module is integrated into the scale-hopping connection paths between the encoder and decoder. As multi-scale features are passed from the encoder to the decoder, the anatomical attention generated by SDM is used to calibrate the features step by step. This forces the network to prioritize features near anatomical boundaries at all scales, thereby continuously applying shape constraints during detail reconstruction and ultimately outputting smooth, continuous, and anatomically sound segmentation boundaries.

[0087] 4. Composite Loss Function This invention not only introduces AGBR and SDMAA modules into the fine segmentation network, but also incorporates an anatomical exclusive loss function during training to penalize activation in anatomically unreasonable regions (such as predicting mandibular teeth within the maxilla), thereby forcibly imposing topological constraints at the output level and further improving the clinical rationality of the segmentation results.

[0088] Specifically, the compound loss function used in this invention is a weighted sum of two classic loss functions: LCE (Multi-class Cross-Entropy Loss) and LDC (Soft Dice Loss). LCE measures the accuracy of voxel-level classification, while LDC directly optimizes the spatial overlap between the predicted segmentation results and the ground truth labels, proving particularly effective in addressing the common class imbalance problem in tooth segmentation. The formula for the compound loss function is as follows: , In the above formula, For composite loss function, and To balance the hyperparameters of the two loss functions, The loss function for LCE, This is the loss function for LDC.

[0089] Example 1: The patient CBCT image data acquired in this example has a voxel spacing of 0.2 mm. The original DICOM format data was preprocessed, including window width and level adjustment and Gaussian noise filtering. The preprocessed volume data was randomly cropped into 128×224×224 patch blocks and normalized to zero mean and unit variance. The processed data was input into the first stage's coarse segmentation network based on 3D U-Net. This network uses the SGD optimizer with an initial learning rate of 0.01, momentum of 0.99, and a poly learning rate decay strategy, trained for 500 epochs. The first-stage network outputs probability maps of the maxillary and mandibular teeth. In the second stage, the probability maps are obtained using the formula... The ambiguity map is calculated, and the original CBCT data is concatenated with the probability maps of the maxillary and mandibular teeth to form a three-channel input. This input is then fed into a fine-grained segmentation network based on Residual U-Net, with an AGBR module integrated at the bottleneck layer and an SDMAA module integrated at the skip connections. The final output is a 42-class segmentation result, such as... Figure 6 As shown.

[0090] In this process, the ambiguity map generated in the first stage is used to select high-uncertainty voxel regions with ambiguity values ​​in the range of 0.94-0.99. A residual thinning network based on the AGBR module is applied to these regions. This network consists of four 3D convolutional layers with a kernel size of 3×3×3 and uses the ReLU activation function. A gating mechanism is used to control the hyperparameters of the thinning intensity. =0.5, enabling targeted boundary optimization. Simultaneously, SDMAA modules are integrated at the scale-skip connections of the decoder, and an adapter method is used to project the SDM map onto the feature space to generate a spatial attention map. Feature recalibration is then achieved through anatomical weighted pooling.

[0091] During model training, the Spatial Attention Model (SDM) is calculated based on real labeled data. For each training sample, the SDM and the feature map at the corresponding scale are input into the SDMAA module. Two spatial attention generation methods are employed: one is a direct calculation method, with a temperature coefficient set. =2.0, after normalizing the SDM, a spatial attention map is generated through an exponential function; the second is the adapter method, which uses 1×1×1 convolution to project the single-channel SDM map onto the multi-channel feature space. Anatomical weighted pooling is used instead of traditional global average pooling, so that the network prioritizes features near the anatomical boundary when evaluating channel importance, thereby enhancing the topological correctness of the segmentation results.

[0092] To more intuitively compare the segmentation performance of the anatomical perception cascade network of this invention (hereinafter referred to as the "model") with that of traditional models, Figure 1 The visualizations of segmentation results for the same test sample are presented sequentially using AttentionUNet3D, 3DUX-Net, UNETR, SwinUNETR, nnFormer, nnUNet, PMFSNEt, the current model, and ground truth annotations. Because ToothNet's code is not publicly available, the results cannot be reproduced. Figure 1 ToothNet is not provided. Each model is presented with four result images arranged from top to bottom, showcasing its comprehensive segmentation capabilities for complex oral anatomy, including teeth and jawbones, from axial, sagittal, coronal, and 3D reconstruction perspectives. Green blocks represent the mandible, yellow blocks represent the maxilla, and red blocks represent the maxillary sinus. Using ground truth annotations as an ideal reference group, the specific segmentation performance of traditional models is explained below.

[0093] 1) AttentionUNet3D Segmentation characteristics: It is a deep learning model that integrates attention gating mechanism with 3D U-Net to adapt to tooth segmentation scenarios. While retaining the spatial structure modeling capability of 3D U-Net, it achieves "feature focusing" through attention mechanism to accurately capture the 3D anatomical features of teeth.

[0094] Existing defects: Figure 1 The AttentionUNet3D segmentation results show severe undersegmentation in the crown and root apex regions, resulting in the loss of significant tooth details. Furthermore, the jawbone contour segmentation is incomplete, severely disrupting the anatomical relationship between the teeth and alveolar bone.

[0095] 2) 3DUX-Net Segmentation characteristics: Hybrid architecture, which combines the local convolutional feature extraction of U-Net with the global dependency capture of Transformer, and achieves deep fusion of semantic and spatial features through a parallel dual-branch structure.

[0096] Existing defects: Figure 1 The 3DUX-Net model can roughly segment the tooth region, but it leads to incorrect fusion of multiple teeth in adjacent areas, resulting in significant boundary confusion. This indicates that even with the fusion of local and global features, without an adaptive refinement mechanism for boundary uncertainties and strong anatomical shape constraints, the model still struggles to make accurate instance distinctions in complex, densely packed tooth regions.

[0097] 3) UNETR Segmentation characteristics: It adopts a pure Transformer encoder, treats 3D voxels as sequences for global attention modeling, and is good at capturing long-range dependencies and global contextual information.

[0098] Existing defects: Figure 1 The results show that UNETR's global modeling capabilities failed to effectively translate into local boundary accuracy in tooth segmentation tasks. Similar to AttentionUNet3D, its segmentation results exhibited severe undersegmentation in the crown and root apex regions, resulting in the loss of significant tooth details. Simultaneously, the jawbone contour segmentation was incomplete, severely disrupting the anatomical relationship between the teeth and alveolar bone. This indicates that the pure Transformer structure, lacking a targeted local optimization mechanism, is insufficient in handling low-contrast, high-detail tooth boundaries, and is prone to topological errors.

[0099] 4) SwinUNETR Segmentation features: Based on UNETR, a hierarchical sliding window self-attention (Swin Transformer) is introduced, which reduces computational complexity and improves feature extraction efficiency while maintaining global modeling capabilities.

[0100] Existing defects: Figure 1 The segmentation capabilities of SwinUNETR are similar to those of 3DUX-Net. While it can roughly segment the tooth region, it performs poorly in handling the boundaries of adjacent teeth, leading to the incorrect merging of multiple teeth and segmentation adhesion problems. This boundary confusion also extends to the junction of teeth and jawbone, causing distortion of alveolar bone morphology. This indicates that even with improved computational methods, this type of end-to-end general architecture still lacks explicit recognition and targeted optimization mechanisms for specific ambiguous boundaries between teeth, and cannot effectively solve the adhesion problem at intermaxillary contact surfaces.

[0101] 5) nnFormer Segmentation Features: This is a Transformer-based volumetric data segmentation network. Its core lies in the design of an interleaved self-attention mechanism and convolution operations. Local convolution and global self-attention are used alternately across different resolution levels, effectively modeling long-range dependencies and global contextual information while maintaining computational efficiency. This design gives it strong feature fusion and semantic modeling capabilities in complex 3D medical image segmentation tasks.

[0102] Existing defects: Figure 1 The results show that nnFormer still exhibits issues such as incorrect tooth fusion and jawbone morphology distortion when processing low-contrast intermaxillary regions. This indicates that even with the fusion of local and global features through an interleaving mechanism, nnFormer lacks an explicit recognition and adaptive optimization mechanism for specific ambiguous areas such as the occlusal contact surface, and cannot effectively resolve the problem of low-contrast boundary adhesion caused by natural occlusion.

[0103] 6) nnUNet Segmentation characteristics: A highly adaptive framework, not a single network, but a powerful workflow that automatically configures network structure, preprocessing, and post-processing based on dataset characteristics. It performs robustly in numerous medical image segmentation tasks and is often considered a strong baseline model.

[0104] Existing defects: Figure 1 While nnUNet outperforms other traditional models in overall segmentation, its segmented jawbone surfaces lack the smoothness of real-world annotations, and its segmented maxillary sinus surfaces also lack the features of real-world annotations, failing to achieve ideal anatomical accuracy. This indicates that its automated framework has limitations when dealing with tasks requiring strong shape priors to ensure anatomical plausibility (such as precise delineation of teeth and jawbones).

[0105] 7) PMFSNet Segmentation Features: A lightweight model for tooth segmentation in 3D CBCT, employing a U-Net-based structure and integrating a polarization multi-scale feature self-attention mechanism to reduce computational complexity.

[0106] Existing defects: Figure 1 The results show that its segmentation performance is only average, and there are obvious missegmentation problems in the sagittal plane. This confirms that its lightweight model architecture is insufficient to cope with such complex, multi-objective, and high-precision tooth segmentation tasks, and it has shortcomings in feature representation ability and handling of complex boundaries.

[0107] In summary, traditional models exhibit significant shortcomings in segmentation results, which precisely explains their deficiencies in quantitative metrics. AttentionUNet3D and UNETR not only show severe undersegmentation in the crown and root apex regions, losing substantial tooth details, but their segmentation of the jawbone contour is also incomplete, severely disrupting the anatomical relationship between teeth and alveolar bone. While 3DUX-Net and SwinUNETR can roughly segment the tooth region, they cause multiple teeth to be incorrectly merged when processing adjacent regions. This boundary confusion extends to the junction of teeth and jawbone, distorting the alveolar bone morphology. Even the high-performance nnUNet lacks the features of realistically labeled maxillary sinus surfaces, failing to achieve ideal anatomical accuracy. nnFormer employs a hybrid design of staggered Transformers and convolutions to fuse local and global features, but it still suffers from incorrect tooth fusion and jawbone morphology distortion when processing low-contrast intermaxillary regions, demonstrating a lack of specific optimization mechanisms for specific blurred boundaries. PMFSNet's segmentation performance was also relatively poor, with missegmentation occurring in the sagittal plane, confirming that its model architecture is not suitable for such complex, multi-object segmentation tasks.

[0108] In comparison, this model demonstrates significant advantages in visualization, with segmentation results highly consistent with real annotations. It not only completely and clearly delineates the independent contours of each tooth, effectively avoiding erroneous adhesions between the upper and lower jaws, but also accurately reconstructs fine structures such as the root apex. Simultaneously, the model's segmentation of the upper and lower jaws is equally precise, with smooth bone surfaces and natural morphology, fully preserving the correct three-dimensional spatial anatomical relationship between teeth and alveolar bone. This superior segmentation performance is attributed to the invention's unique two-stage framework, ambiguity-gated targeted boundary optimization, and shape-prior-guided feature learning mechanism, enabling it to specifically overcome inherent challenges in CBCT and ultimately providing a high-quality, high-fidelity three-dimensional anatomical model foundation for downstream clinical applications.

[0109] To quantitatively analyze the segmentation performance of the traditional model and this model, the DEC (Dice Similarity Coefficient), HD95 (95th Percentile Hausdorff Distance), ASSD (Average Symmetric Surface Distance), and SEN (Sensitivity) of the segmentation results output by each model were evaluated. Detailed evaluation results can be found in [link to evaluation]. Figure 5 . Figure 5The results show that this model improves DSC to 90.17% on the test set, while reducing HD95 to 3.63mm, ASSD to 1.54mm, and SEN to 90.00%. It not only surpasses the existing best method in voxel-level segmentation accuracy, but also achieves significant optimization in boundary quality, resulting in a significant overall performance improvement in tooth segmentation tasks.

[0110] In a second aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a CBCT tooth segmentation method based on an anatomical sensing cascade network.

[0111] The memory can be volatile or non-volatile, or a combination of both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memory of this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0112] The processor can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0113] The method steps of this invention can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0114] Software implementation can be achieved by executing functional modules (such as procedures, functions, etc.). Software code can be stored in memory and executed by the processor. Memory can be implemented in the processor or outside the processor.

[0115] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include such modifications and variations.

Claims

1. A CBCT tooth segmentation method based on an anatomically-aware cascading network, characterized in that, The anatomy-aware cascaded network adopts a two-stage segmentation framework, and the method comprises the following steps: In the first stage, based on the original CBCT image, a simplified binary segmentation is performed by a coarse segmentation network with 3D U-Net as the backbone, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational complexity, and a max-pooling layer is added at the end of the network to reduce the computational ​ ​ ​ ​ 2. The method of claim 1, wherein, ​ 3. The method of claim 1, wherein, ​ ， ， In the above formulae, is a foreground probability, is a maxillary teeth probability map, is a mandibular teeth probability map, is an ambiguity map.

4. The method of claim 1, wherein, ​ ， In the above formula, is the refined feature map for final output, is the input high semantic feature map, is a hyper-parameter to control the refinement intensity, is the binary gating mask, represents element-wise multiplication, is the gated refined feature map.

5. The method of claim 4, wherein, The control refinement intensity super parameter The specific value of the binary gating mask The high uncertainty region of the binary gating mask is specifically an interval of 0.94-0.99 of the ambiguity value.

6. The method of claim 1, wherein, ​ ， In the above formula, for a spatial attention map, for a spatial voxel index, for an exponential transform function, for a normalized SDM, for a learnable temperature coefficient.

7. The method of claim 1, wherein, The adapter computing method of the spatial attention map specifically comprises: projecting the SDM to multi-channel features through a 1x1 convolution layer, and then calculating the average absolute value of the channel dimension to generate .

8. The method of claim 1, wherein, The step of the anatomy-weighted pooling specifically comprises: using a spatial attention map as a spatial weight to perform weighted average on a feature map at a current scale in multi-scale feature maps to generate an anatomy-aware channel descriptor vector , so as to give priority to features located near an anatomical boundary when evaluating channel importance.

9. The method of claim 1, wherein, ​ ， In the above formula, is a composite loss function, and is a hyperparameter for balancing the two loss functions, is a loss function of LCE, is a loss function of LDC.

10. A CBCT tooth segmentation system based on an anatomically-aware cascading network, comprising a memory, a processor and a computer program stored on the memory, characterized in that, ​

Images