An Automatic Caries Identification Method Based on Transformer-U-shaped Network in CBCT

The use of Transformer-U-shaped networks for automatic CBCT caries identification solves the problem of caries segmentation and localization separation in existing technologies, achieving high-precision caries identification and localization, and improving the stability and clinical usability of automatic caries identification.

CN122597401BActive Publication Date: 2026-11-10XUZHOU MEDICAL UNIVERSITY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611062776.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-17
Publication Date
2026-11-10
Estimated Expiration
2046-07-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively utilize the continuous contextual information between CBCT slices, leading to unstable identification of small lesions. Furthermore, the separation of caries segmentation and localization tasks affects the accuracy and efficiency of automatic caries identification.

Method used

A Transformer-U network is used for 3D context modeling. Through a 3D block embedding module, encoder, bottleneck layer and decoder, combined with a multi-task loss function, end-to-end caries segmentation and localization is achieved. The long-range dependency between slices is explicitly modeled by a cross-slice window self-attention mechanism.

Benefits of technology

It significantly improves the stability and consistency of early small-volume caries and proximal caries identification, provides accurate caries segmentation masks and subvoxel localization heat maps, and assists in the development of clinical treatment plans.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122597401B_ABST
    Figure CN122597401B_ABST
Patent Text Reader

Abstract

This invention belongs to the field of intelligent CT image analysis technology, and relates to an automatic CBCT caries identification method based on a Transformer-U network. This method performs grayscale cropping, effective slice cropping, size standardization, and pixel normalization preprocessing on CBCT volume data to construct a sliding window slice sequence input unit. It constructs a Transformer-U network based on Swin-Unet, employing 3D window self-attention in the encoding and decoding paths to achieve cross-slice context modeling, and injecting 3D relative position bias to enhance spatial perception. Through multi-task learning, it simultaneously outputs the three-category caries segmentation results and 3D localization heatmap. This invention significantly improves the stability and 3D consistency of CBCT small caries lesion identification, achieving integrated output of segmentation, localization, and quantization, and can be directly used for auxiliary diagnosis of caries in oral clinical practice.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image analysis technology, specifically relating to an automatic CBCT caries identification method based on Transformer-U-shaped network. Background Technology

[0002] Dental caries is one of the most common chronic infectious oral diseases in clinical practice, mainly characterized by progressive demineralization and destruction of the hard tissues of the teeth. In clinical diagnosis and treatment, accurately identifying the spatial location, morphological extent, and depth of caries lesions is crucial for developing treatment strategies (such as minimally invasive caries removal, preservation of healthy tooth tissue, and selection of appropriate restorations). Cone-beam computed tomography (CBCT) provides richer three-dimensional spatial information compared to traditional two-dimensional X-rays.

[0003] However, automatic caries identification based on CBCT images still faces many challenges. First, caries often appear in CBCT images as having low contrast and blurred boundaries, especially in early or occult caries where the grayscale difference between the caries and adjacent normal tooth tissue is limited, easily leading to missed detections or underestimation of the lesion area. Second, CBCT images are often accompanied by noise, scattering artifacts, and streak artifacts produced by metal restorations, further increasing the risk of false positives and false negatives.

[0004] Existing deep learning-based caries identification solutions often process CBCT volumetric data by dividing it into independent two-dimensional slices or using traditional three-dimensional convolutional networks. These solutions struggle to explicitly and adequately model the continuous spatiotemporal context information between slices, resulting in insufficient stability in identifying complex spatial structures such as proximal caries and pit and fissure caries, as well as small-volume early caries, and are prone to inconsistencies across slice boundaries. Furthermore, existing solutions often separate caries segmentation and localization tasks, lacking an end-to-end integrated solution, which impacts the efficiency and clinical usability of assisted diagnosis.

[0005] Therefore, there is an urgent need for an automatic caries identification method that can effectively perform three-dimensional context modeling of CBCT data and simultaneously achieve accurate segmentation and localization. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an automatic CBCT caries identification method based on Transformer-U network, so as to solve the problems that the prior art is difficult to effectively utilize the continuous context information between CBCT slices, resulting in unstable identification of small lesions and separation of segmentation and localization tasks.

[0007] To achieve the objectives of this invention, the following technical solutions are adopted.

[0008] An automatic caries identification method based on Transformer-U-shaped network in CBCT includes the following steps: S1. Obtain raw oral caries CBCT image data, perform three-dimensional annotation of caries lesions, and generate annotated CBCT volume data; S2. Perform grayscale cropping, effective slice interval cropping, size standardization and pixel normalization preprocessing on the CBCT volume data, perform sliding window segmentation and resampling on the effective slices, construct slice sequence input units containing the context of adjacent slices, and obtain training set, validation set and test set by random division at the patient level. S3. Construct a Transformer-U network, which includes a 3D block embedding module, an encoder, a bottleneck layer, a decoder, and dual output heads for segmentation and localization. The 3D block embedding module maps the input to a feature sequence and injects 3D position information. The encoder uses 3D window self-attention combined with 3D relative position bias, and introduces a mask under the shift window to achieve cross-slice context modeling. The decoder fuses multi-scale features through upsampling and skip connections. The network is trained using a multi-task loss function consisting of segmentation loss and heatmap regression loss. S4. Repeat step S2 for preprocessing and input construction of the CBCT body data to be tested. Input the trained network to obtain the segmentation output and three-dimensional localization heatmap. Fill the segmentation mask back into the original body data space to obtain the global caries segmentation result. Perform peak search and weighted center calculation on the heatmap, and combine the segmentation mask to carry out connected component analysis to output the physical coordinates and volume of the caries lesion center point. Simultaneously obtain the caries segmentation result and localization result.

[0009] Furthermore, the original oral caries CBCT image data were acquired under consistent scanning and reconstruction conditions.

[0010] Furthermore, during the 3D annotation process, a threshold segmentation method is used to adjust the window width and window level to set the threshold within a preset range, thereby enhancing the visualization effect of the tooth structure. Median filtering and closure operation-cavity filling morphological operation are combined to optimize the segmentation results.

[0011] Furthermore, the preprocessing includes: Grayscale cropping cuts off the voxel intensity of CBCT volumetric data to a preset range to eliminate interference from irrelevant tissues and enhance the contrast of the target area. Effective slice interval trimming: Determine the start and end slice positions of the tooth region along the Z-axis, trim and retain only continuous slice segments containing tooth structure, and remove background slices without tooth information to reduce computational redundancy; Size standardization was achieved by uniformly adjusting all annotated original oral caries CBCT image data to the same resolution. Pixel normalization maps the pixel values ​​of all annotated original oral caries CBCT image data to a preset range.

[0012] Furthermore, the construction of the slice sequence input unit involves cutting the continuous segments after the effective slice interval is truncated, using a sliding window to divide them along the Z-axis to obtain a series of three-dimensional data blocks, and then resampling them to a uniform size as the input unit of the network.

[0013] Furthermore, the Transformer-U-shaped network includes: The input terminal is equipped with a three-dimensional block embedding module, which is used to divide the input slice sequence input unit into three-dimensional blocks without overlap, and to map each block into a high-dimensional token through linear projection to obtain a token sequence; at the same time, a learnable three-dimensional absolute position code is injected into the token sequence to preserve the global spatial position information of the token. The encoding path consists of four stages. Each stage includes a block embedding layer that performs spatial downsampling through block merging and two stacked Transformer blocks for feature extraction. It alternates between window multi-head self-attention and shift window multi-head self-attention. The decoding path consists of four stages. Each stage includes a block expansion layer that completes spatial upsampling and two Transformer blocks for cross-scale feature fusion and detail reconstruction. The decoding path and the encoding path are connected by a skip connection to pass multi-scale features. The Transformer block embeds a cross-slice window self-attention module, which calculates self-attention within a preset three-dimensional window, wherein: it queries tokens that come only from the central slice, the key and value come from tokens of the central slice and its adjacent slices, and injects a relative position offset based on the three-dimensional relative coordinate index of the tokens within the window; The bottleneck layer, located between the encoder and decoder, is used for global semantic modeling and feature integration. Its input is the high semantic low-resolution features at the end of the encoding path. It only enhances spatial features and does not perform downsampling to avoid the loss of fine-grained information. It completes local modeling and cross-window information interaction through stacked sliding window multi-head self-attention modules, and outputs high semantic features as the starting input of the decoding path, providing semantic constraints and structural priors for upsampling and detail reconstruction. Segmentation output branch: The feature map output from the decoding path is mapped into a three-class segmentation probability map through a convolutional projection layer, and the segmentation results of background, tooth body and caries are output; Localization output branch: The feature map output from the decoding path is used to generate a single-channel 3D localization heatmap through 3D convolution, normalization and non-linear activation layers.

[0014] Furthermore, during the training phase, a multi-task loss function is used to optimize the Transformer-U-shaped network model. The multi-task loss function consists of a weighted sum of segmentation loss and heatmap regression loss.

[0015] Furthermore, the caries segmentation result is a three-category segmentation mask output by the segmentation output head, consisting of a background class, a tooth class, and a caries class. The three-category segmentation masks are then backfilled and merged into the original volume data space to obtain a global segmentation result.

[0016] Furthermore, the caries localization result is a single-channel three-dimensional localization heatmap output by the localization output head. The voxel coordinates of the caries center point are extracted from the three-dimensional localization heatmap through peak search and local neighborhood weighted center calculation, and converted into physical space coordinates by combining voxel spacing and origin information. At the same time, connected component analysis is performed based on the three-category segmentation mask to calculate the volume of each caries connected component, and the localization information containing the physical coordinates of the center point and the volume is output.

[0017] Compared with the prior art, the present invention has the following advantages: 1. High recognition accuracy and more robust to small lesions: By constructing slice sequence input units and introducing a cross-slice window self-attention mechanism, the network can explicitly and efficiently model the long-range dependencies and three-dimensional contextual information between slices in CBCT data, which significantly improves the recognition stability and three-dimensional consistency of lesions with low contrast and blurred boundaries, such as early small-volume caries and proximal caries, and reduces missed detections and false detections.

[0018] 2. Integrated task and strong clinical usability: It adopts an end-to-end Transformer-U-shaped network, which simultaneously outputs pixel-level caries segmentation masks and sub-voxel precision localization heatmaps, providing clinicians with comprehensive quantitative information including the precise contour of the lesion, the three-dimensional physical coordinates of the center, and the volume, which can directly assist in the formulation of surgical or treatment plans.

[0019] 3. Advanced technical solution: The Transformer architecture is creatively applied to the specific task of CBCT caries analysis. Through customized cross-slice attention calculation and three-dimensional relative position encoding, it effectively overcomes the limitations of traditional methods in processing three-dimensional temporal medical image data. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the CBCT automatic caries identification method based on Transformer-U-shaped network of the present invention; Figure 2 This is a schematic diagram of the Transformer-U-shaped network structure of the present invention, showing the encoding path, decoding path, segmentation branch, and positioning branch; Figure 3This is a schematic diagram of the three-dimensional window self-attention and cross-slice interaction mechanism of the present invention; Figure 4 This is a graph showing the change in the caries segmentation performance of the present invention. Detailed Implementation

[0021] The present invention will be further described in conjunction with the accompanying drawings and embodiments.

[0022] As an embodiment of the present invention, such as Figure 1 As shown, a CBCT automatic caries identification method based on Transformer-U-shaped network includes the following steps: Obtain raw CBCT images of dental caries and annotate the raw dental caries images; The CBCT volume data is preprocessed and a slice sequence input unit is constructed; Construct a Transformer-U-shaped network model and train it based on the slice sequence input unit; After training, the trained Transformer-U network is used to perform inference processing on the new CBCT volume data, and output the caries segmentation results and caries localization results.

[0023] In this embodiment of the invention, the CBCT automatic caries identification method based on Transformer-U-shaped network includes the following specific steps: S1. Obtain raw CBCT images of dental caries and annotate the raw caries images; the specific process is as follows: S101, CBCT imaging data of dental caries were obtained as follows: When using the Italian NewTom VGi cone-beam CT to acquire images of the patient's teeth, it is necessary to ensure that all images are acquired under consistent scanning and reconstruction conditions during data acquisition. S102, Image Data Labeling: All volumetric data were labeled using 3D Slicer software. Threshold segmentation was employed, with window width and level adjusted (WW2500 / WL 1500), and the threshold range set to 1000–3000 HU to enhance the visualization of tooth structure. Morphological operations such as median filtering (1.00 mm) and closure operation-cavity filling (1.50 mm) were combined to optimize the segmentation results. For caries annotation, axial images were used as a basis to delineate the lesion area layer by layer, followed by meticulous manual calibration in coronal and sagittal images to ensure accurate reconstruction of the carious area in three-dimensional space. All CBCT images were independently completed by two researchers with oral medicine backgrounds and then verified by a senior endodontist to ensure the accuracy of label boundaries and consistency of annotation level. Samples with deviations were discussed and corrected to ensure high-quality and clinically interpretable labeled data. Finally, the labels were exported to generate NifiTI format files for subsequent model training. The size of the CBCT raw data obtained in the above steps is , can be represented as: , In the formula, Represents the original body data; Represents the real number space; D is the depth in the z-axis direction; H is the slice height in the y-axis direction; W is the slice width in the x-axis direction; The corresponding segmentation tag size is 400. , can be represented as: ; In the formula, This indicates the voxel-level segmentation label corresponding to the original CBCT volume data; The depth is along the z-axis. The slice height is along the y-axis. The slice width along the x-axis; S2. Preprocess the CBCT volume data and construct a slice sequence input unit; The specific process of step S2 is as follows: S201, Grayscale cropping: Voxel intensity is truncated to the interval In the formula, and These correspond to the lower and upper boundaries of the gray-level concentration intervals in the gray-level histogram of the training data, with corresponding values ​​of -1000 and 2500. S202, Effective slice interval trimming: Determine the start and end slice positions of the tooth region, and cut along the Z-axis to retain only continuous slice segments containing tooth structure, while removing a large number of background slices that are pure air or have no tooth information, thus creating the effective slice sequence; S203, Size Standardization: All original CBCT images of dental caries were uniformly adjusted to the same resolution. This example uses ; S204, Pixel Normalization: A normalization strategy is adopted to map all pixel values ​​to the range of 0,1; S205. Constructing the slice sequence input unit: The effective slice sequence obtained from the cropping is divided into slices using a sliding window with a window size of 16 and a sliding step size of 8, thus obtaining the slice sequence input unit. In the formula, This represents the input unit for the slice sequence generated by the nth sliding window; This indicates the number of slices within the window, in this embodiment. =16; H and W represent the slice height and width, respectively; and it is resampled to a uniform size. The tensor of the above input network can be represented as: , In the formula, This represents the five-dimensional tensor input to the network; B is the batch size; 1 is the number of channels. , , The dimensions after resampling are: depth 16, height 224, width 224. S206. Dataset partitioning: A patient-level random partitioning strategy is adopted, and all cases are divided into training set, validation set and test set in an 8:1:1 ratio; S3. Construct a Transformer-U network model and train it based on the slice sequence input units, such as... Figures 2-3 As shown; the specific process is as follows: S301, Block Embedding Module: Utilizing the 3D block embedding module to... The non-overlapping partitions are projected into a sequence, and the output is: , In the formula, C is the tensor output by the block embedding module; C is the embedding dimension. , , , This refers to the size of the 3D block. These represent the number of blocks in the depth, height, and width directions after the block embedding module is divided. Simultaneously, learnable absolute position encoding is injected, and the grid coordinates of each image token are cached. For 3D position modeling, the process can be described as follows: , In the formula, This indicates a learnable absolute position encoding function; Indicates an update operation; S302, Cross-slice sliding window attention module and relative position bias: Divide image tokens by window, and apply the window token sequence... Multi-head attention is computed, and tokens within the window are split according to the slice dimension. In attention calculation, the query comes only from the central slice, while the key and value come from the central slice and neighboring slices. The linear mapping process is defined as follows: , , , , In the formula, The number of tokens within each window; , , These represent the number of tokens contained in the window in the depth, height, and width directions, respectively. Indicates the sequence of tokens within the window; This represents the set of sub-tokens corresponding to the central slice; and These represent the token sets corresponding to the adjacent slices before and after the central slice; Represents the neighborhood slice radius; Q represents the query matrix; Represents the key matrix; Represents the key mapping matrix; Represents a value matrix; Represents a value mapping matrix; Indicates the weight of the linear mapping in the query; Subsequently, scaled dot product attention is calculated, and a 3D relative position offset is introduced to display the spatial positional relationship of tokens within the modeling window. Simultaneously, a mask matrix is ​​introduced under the shift window mechanism. To limit illegal attentional connections, the calculation formula is as follows: , In the formula, This is the attention weight matrix; This is the transpose of the key matrix; A score is given for attention similarity. Dimensions for each attention head; This is the scaling factor; The relative position offset is obtained by indexing the three-dimensional relative displacement within the window to the parameter table; For the attention mask matrix; Softmax ( ) is the normalization function; Finally, the normalized attention weights are used to perform weighted aggregation of the cross-slice Value features to obtain the context-enhanced representation of the central slice: , In the formula, A is the normalized attention weight matrix; V is the value matrix; and Y is the attention output result. The output of multi-head attention, after concatenation and linear mapping, is combined with the input features through residual connection box normalization, thereby forming a central slice feature representation that integrates 3D spatial structure information and cross-slice context dependencies. This process can be described as follows: , , , In the formula, MSA is the multi-head self-attention operator; To stitch together multiple attention head outputs along the channel dimension; For the first The output features of each attention head; h is the number of attention heads; The output projection matrix is ​​used to fuse the multi-head attention results; These are intermediate features after multi-head self-attention and residual connections; For regularization operators; It is a feedforward network; For layer normalization operators; The final output features are centered on the slice; S303, Caries Segmentation Result Output Module: This module includes the encoding path, bottleneck layer, decoding path, and skip connections. The encoding path consists of four stages, each of which first performs downsampling through block merging (this process maintains...). The dimension remains unchanged, only for the spatial dimension. (Downsampling and progressively increasing the number of channels), followed by stacking two sliding window attention blocks for feature extraction, where window multi-head self-attention and shift window multi-head self-attention are used alternately.

[0024] The bottleneck layer, located between the encoding and decoding paths, is responsible for global semantic modeling and feature integration. Its input consists of high-semantic, low-resolution features from the end of the encoding path, while maintaining... The dimension remains unchanged, only for the spatial dimension. The smallest-scale features are enhanced and no further downsampling is performed to avoid loss of fine-grained information. This layer stacks sliding window multi-head self-attention modules to achieve local modeling and cross-window information interaction. The high semantic features output by the bottleneck layer serve as the starting input to the decoding path, participating in the stepwise recovery process together with the multi-scale features from the encoding end, providing strong semantic constraints and structural prior information for the subsequent upsampling stage.

[0025] The multi-scale features output from each stage of the encoding path are passed to the corresponding stage of the decoding path via skip connections. Each stage of the decoding path first performs upsampling through block expansion (this process preserves...). The dimension remains unchanged, only for the spatial dimension. The number of channels is downsampled and augmented layer by layer, and then concatenated with features of the same scale as the encoding path in the channel dimension. After linear lamination back to the target number of channels, cross-scale fusion and detail reconstruction are performed through two sliding window attention blocks. Finally, the embedded block mesh is strictly mapped back to the original voxel mesh using a pure geometric upsampling module, and multi-class segmentation results are output through convolutional projection. , In the formula, This represents the three-class segmentation result output by the network; 3 indicates the number of classes. S304, Caries Localization Result Output Module: This module outputs features based on the final decoding path. Heatmaps are output after 3D convolution stacking. The threshold is Threshold screening is performed on the heatmap to determine the presence of credible lesion responses. If the global maximum value of the heatmap is below the threshold, it is determined as "no caries detected". When a valid response is found, the coarse center position of the heatmap is first searched by peak value, i.e., the coordinates of the largest corresponding voxel. The process can be described as follows: , , , In the formula, t is the depth coordinate; y is the height coordinate; and x is the width coordinate. coordinates The intensity of the caries response at the site; This indicates that the maximum value is taken at all three-dimensional locations. This represents the maximum response value of the heatmap. Detect represents the threshold; Detect represents the detection result; 0 indicates that no tooth decay was detected, i.e., the maximum response value is below the threshold; 1 indicates that tooth decay was detected, i.e., the maximum response value reaches or exceeds the threshold. This indicates the location of the maximum value. To predict the coordinates of the center of the carious tooth; Define the peak neighborhood radius as Then calculate the local area of ​​the peak. By weighting the center point within the region, we obtain the fine coordinates of the center point. The process can be described as follows: , , , In the formula, Indicates coarse positioning center A local peak neighborhood constructed around the center; These represent the radii of the neighborhood in the depth, height, and width directions, respectively. Represents the coordinates of any voxel within the neighborhood; This represents the weighting coefficient corresponding to the voxel; in this paper, it is taken as the thermographic response value. ; This represents the fine-grained central voxel coordinates obtained through local weighted center calculation; Final binding voxel spacing Origin Information Global starting slice index Further, the central voxel grid coordinates Convert to physical coordinates The process can be described as follows: , , In the formula, Indicates the starting slice index of the valid slice in the original CBCT; Represents the depth coordinates in the original volume data; , These represent the height and width coordinates in the original volume data, respectively; Indicates the physical coordinates of the center of the tooth decay; , , These represent the origin of the physical coordinates of the volume data; , , These represent the voxel spacing in the three directions, respectively. S305. Multi-task loss function: A multi-task loss function is used during the training phase. ,in: This is the total loss function; This refers to the multi-task loss weighting coefficient; The segmentation loss includes multi-class cross-entropy loss. and loss ; For heatmap regression loss; The formulas for the above calculation steps are as follows: , , , For the heatmap regression loss, the mean squared error loss is used, and its calculation formula is as follows: , In the formula, Indicates the number of voxels or samples involved in the loss calculation; Indicates a category index; This represents the true label of the i-th voxel belonging to category c; This represents the probability that the model predicts the i-th voxel belongs to category c; and Let represent the predicted segmentation result and the true segmentation label of the i-th voxel, respectively; This is a smoothing term used to avoid the denominator being zero; and Let represent the predicted heatmap response value and the actual heatmap response value at the i-th position, respectively; S306, Training Configuration: The AdamW optimizer was used for training the model in this example, with the base learning rate and weight decay set to 0.0008 and 0.0001, respectively. The batch size was 2, and the model was trained for a total of 100 iterations. The overall model performance was evaluated using accuracy, AUC, F1 score, recall, and IoU. S4. After training is completed, the trained Transformer-U network is used to perform inference processing on the new CBCT volume data, and output the caries segmentation results and caries localization results. The new CBCT volume data undergoes preprocessing consistent with the training phase, and a slice sequence input unit is constructed. This slice sequence input unit is then input into the Transformer-U network to obtain segmentation and localization outputs. The segmentation and localization outputs are then processed to generate the final output. The specific process is as follows: S401, Inference output: Dental caries segmentation result: The segmentation output head outputs three types of segmentation score tensors and applies them to the three types of caries segmentation masks, namely background class, tooth body class and caries class. The caries segmentation masks are then backfilled and fused into the original volume data space to obtain the global segmentation result.

[0026] S402, Inference output: Caries localization results: The positioning output head outputs a single-channel three-dimensional positioning heat map and extracts the voxel coordinates of the caries center point from the three-dimensional positioning heat map. Based on the caries segmentation mask, connected component analysis is performed to calculate the voxel coordinates and volume of the caries connected component, and the positioning information containing the physical coordinates of the center point and the caries is output.

[0027] As an experimental verification example of the present invention, such as Figure 4As shown, the Caries IoU for dental caries reached 0.91, indicating that the model stably identified lesion areas in the mid-to-late stages. The Caries Dice coefficient was 0.95, and the Dice coefficient gradually increased with training, indicating that the model can accurately reproduce the details of lesion edges.

[0028] The test results are shown in Table 1, which provides the test results of this example model, including the comparison results with the classic model. Table 1: Comparison of Test Results Model IoU Dice F1 score Recall rate 3D Unet 0.66 0.75 0.74 0.72 UNTER 0.68 0.64 0.64 0.61 TransUNet 0.80 0.77 0.77 0.75 This invention 0.97 0.95 0.98 0.98 As shown in Table 1, Swin-Unet significantly outperformed other comparative models in both Caries IoU and Caries Dice evaluation metrics, demonstrating stronger lesion area localization capabilities and comprehensive detection capabilities.

Claims

1. A CBCT automatic caries identification method based on Transformer-U-shaped network, characterized in that, Includes the following steps: S1. Obtain raw oral caries CBCT image data, perform three-dimensional annotation of caries lesions, and generate annotated CBCT volume data; S2. Perform grayscale cropping, effective slice interval cropping, size standardization and pixel normalization preprocessing on the CBCT volume data, perform sliding window segmentation and resampling on the effective slices, construct slice sequence input units containing the context of adjacent slices, and obtain training set, validation set and test set by random division at the patient level. S3. Construct a Transformer-U network, which includes: a 3D block embedding module at the input end, used to divide the input slice sequence into 3D blocks without overlap, and map each block to a high-dimensional token through linear projection to obtain a token sequence; simultaneously, inject learnable 3D absolute position encoding into the token sequence; the encoding path consists of four stages, each stage including a block embedding layer that performs spatial downsampling through block merging and two stacked Transformer blocks for feature extraction, and alternately using window multi-head self-attention and shift window multi-head self-attention; the decoding path consists of four stages, each stage including a block expansion layer that performs spatial upsampling through block expansion and two Transformer blocks for cross-scale feature fusion and detail reconstruction, and the decoding path and encoding path are connected by skip connections to pass multi-scale features; the Transformer block embeds a cross-slice window self-attention module, which in Self-attention is computed within a predefined 3D window, where: the query only retrieves tokens from the central slice, the key and value come from tokens in the central slice and its neighboring slices, and a relative position bias is injected based on the 3D relative coordinate index of the tokens within the window; the bottleneck layer is located between the encoder and decoder, used for global semantic modeling and feature integration, its input is high semantic low-resolution features at the end of the encoding path, it only enhances spatial features and does not perform downsampling, and completes local modeling and cross-window information interaction through stacked sliding window multi-head self-attention modules, outputting high semantic features as the starting input of the decoding path; the segmentation output branch maps the feature map output by the decoding path to a three-class segmentation probability map through a convolutional projection layer, outputting the segmentation results of background, tooth body and caries; the localization output branch generates a single-channel 3D localization heatmap from the feature map output by the decoding path through 3D convolution, normalization and nonlinear activation layers; the training phase uses a multi-task loss function consisting of a weighted sum of segmentation loss and heatmap regression loss to complete network training; S4. Repeat step S2 for preprocessing and input construction of the CBCT volume data to be tested. Input the trained network to obtain three-category segmentation masks and three-dimensional localization heatmaps for background, tooth, and caries classes. Backfill and fuse the three-category segmentation masks into the original volume data space to obtain caries segmentation results. Extract the voxel coordinates of the caries center point from the three-dimensional localization heatmap through peak search and local neighborhood weighted center calculation, and convert them into physical space coordinates by combining voxel spacing and origin information. At the same time, perform connected component analysis based on the three-category segmentation masks, calculate the volume of each caries connected component, and output the caries localization results containing the physical coordinates of the center point and the volume.

2. The CBCT automatic caries identification method based on Transformer-U-shaped network according to claim 1, characterized in that, The original CBCT image data of dental caries were acquired under consistent scanning and reconstruction conditions.

3. The CBCT automatic caries identification method based on Transformer-U-shaped network according to claim 1, characterized in that, During the 3D annotation process, the threshold segmentation method is used to adjust the window width and window level to set the threshold within a preset range, thereby enhancing the visualization effect of the tooth structure. The segmentation results are optimized by combining median filtering and closure operation-cavity filling morphological operation.

4. The CBCT automatic caries identification method based on Transformer-U-shaped network according to claim 1, characterized in that, The preprocessing includes: Grayscale cropping cuts off the voxel intensity of CBCT volumetric data to a preset range to eliminate interference from irrelevant tissues and enhance the contrast of the target area. Effective slice interval trimming: Determine the start and end slice positions of the tooth region along the Z-axis, trim and retain only continuous slice segments containing tooth structure, and remove background slices without tooth information to reduce computational redundancy; Size standardization was achieved by uniformly adjusting all annotated original oral caries CBCT image data to the same resolution. Pixel normalization maps the pixel values ​​of all annotated original oral caries CBCT image data to a preset range.

5. The CBCT automatic caries identification method based on Transformer-U-shaped network according to claim 1, characterized in that, The construction of the slice sequence input unit involves cutting continuous segments after effective slice intervals, using a sliding window to divide them along the Z-axis to obtain a series of three-dimensional data blocks, and then resampling them to a uniform size as the input unit of the network.

Citation Information

Patent Citations

  • Decayed tooth identification method based on oral panoramic film and double attention module

    CN113221945A

  • Oral panoramic film decayed tooth segmentation method based on semi-supervised multi-level uncertainty perception

    CN116228639A