A deep learning-based method for detecting lingual canal in CBCT data

Abstract

This invention belongs to the field of medical image processing technology and discloses a deep learning-based method for detecting the lingual canal in CBCT data. The specific steps are as follows: S1: Extract mandibular slice images from CBCT data; automatically segment the tooth region in the CBCT data using a three-dimensional segmentation model, determine the target region containing the lingual canal based on the spatial distribution of the teeth, and generate equidistant continuous two-dimensional slices within this range along the x-axis. Automatically extracting the tooth region using a three-dimensional segmentation model and determining the target region containing the lingual canal accordingly avoids the manual selection of slices or regions in traditional methods, significantly improving the automation and consistency of the detection process; using a deep learning model to detect the lingual canal in the mandibular slice images, enhancing local contrast through preprocessing, and combining confidence screening and non-maximum suppression strategies effectively improves the accuracy of the lingual canal candidate boxes.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically a method for detecting the lateral lingual tube in CBCT data based on deep learning. Background Technology

[0002] Cone-beam computed tomography (CBCT) is widely used in orthodontics, implantology, and jawbone structure analysis. It can provide high-resolution three-dimensional bone tissue information. The lingual canal is an important microscopic anatomical structure within the mandible. Its location is related to the safety of implant placement and accelerated orthodontic surgery. Inaccurate identification of the lingual canal may lead to damage to blood vessels or nerves during surgery, causing risks such as bleeding or numbness. Therefore, improving the detection efficiency and accuracy of the lingual canal is of great clinical significance.

[0003] Currently, clinical practice mainly relies on doctors manually observing the lingual canal layer by layer in CBCT slices. The lingual canal has a small structure, low grayscale contrast, and is easily affected by noise. Manual identification is time-consuming and prone to missed detections. Traditional image processing algorithms, such as threshold segmentation and edge detection, lack robustness under different equipment and patient conditions, making it difficult to reliably identify the lingual canal. Some studies have attempted to introduce deep learning models, but most models are designed for larger anatomical structures and lack specialized mechanisms for small targets like the lingual canal, resulting in unsatisfactory localization accuracy. Furthermore, existing methods generally cannot achieve end-to-end automatic detection, requiring manual preprocessing or region cropping, leading to high application barriers and insufficient generalization ability. Therefore, it is necessary to propose a more accurate and stable deep learning-based automatic detection method for the lingual canal to address the problems existing in current technologies. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for detecting the lateral spiracle in CBCT data, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting the lateral spiracle in CBCT data based on deep learning, the specific steps of which are as follows: S1: Extracting mandibular slice images from CBCT data The tooth region in CBCT data is automatically segmented using a three-dimensional segmentation model. The target region containing the lingual canal is determined based on the spatial distribution of the teeth, and equidistant continuous two-dimensional slices are generated within this region along the x-axis. S2: Use the detection model to obtain slices of the detection frame with tongue-shaped tubes. The mandibular slice image is preprocessed, and the preprocessed slice is input into a pre-trained lingual canal detection model to obtain candidate detection boxes and corresponding confidence scores for the lingual canal. The candidate detection boxes are then subjected to confidence thresholding and non-maximum suppression to obtain the final detection box of the lingual canal in the slice. S3: Restore the detection results to the three-dimensional coordinate system. Determine whether the detection frames between slices belong to the same tongue-shaped tube. Once it is determined that they belong to the same tongue-shaped tube, select the slice with the most central position in the detection frame sequence and calculate the coordinates and radius of its detection frame center point. Then, map the center point back to the three-dimensional coordinate system to realize the three-dimensional positioning and reconstruction of the tongue-shaped tube.

[0006] Preferably, when the CBCT data in S1 is preprocessed, it is downsampled to 0.4 spacing. The three-dimensional segmentation model adopts nnUNet, and a three-dimensional mask of the tooth is obtained through this model. The distribution boundary of the tooth in three-dimensional space is calculated based on the tooth segmentation results. The target area of ​​the lingual canal is estimated according to the relative positional relationship between the tooth and the lingual canal. The generated two-dimensional slice has the same spatial resolution and pixel spacing as the original CBCT data.

[0007] Preferably, the preprocessing of the mandibular slice image in S2 is as follows: Gaussian filtering is used to reduce speckle noise in the CBCT image while preserving the detailed structure of the trabecular bone; CLAHE is used to enhance the mandibular region to strengthen the contrast between the bone wall around the lingual canal and the surrounding tissue.

[0008] Preferably, the lingual tube detection model in S2 adopts a YOLOv11 network structure with a cross-scale memory module; the training process of the lingual tube detection model includes: collecting 3000 CBCT slices containing the median lingual tube and the lateral lingual tube, and using Labelme software to annotate the lingual tube detection boxes; performing data augmentation on the original samples, wherein the data augmentation adopts the YOLOv11 data augmentation strategy, including random translation, random scaling, illumination perturbation, color enhancement, Mosaic stitching enhancement and MixUp image mixing enhancement.

[0009] Preferably, the operation of the cross-scale memory module is as follows: Let the input image be... ,in and These represent the height and width of the image, respectively. Extract feature representations at S scales from the Backbone network. ,in Indicates the scale quantity. Let s be the height and width of the scale s. For the number of channels, Backbone will input... Mapping to multi-scale features ; pass A linear transformation is performed on the number of channels at different scales. Implement as Convolution or learnable linear projections make all scales map to the same memory space dimension D, i.e. ; Construct a unified memory structure Each memory unit Consistent with the feature resolution at the corresponding scale; In the forward inference phase, the stability of the current features is enhanced through attention-based weighted fusion, and the fusion formula is as follows: ,in The matching weight is obtained by dot product attention. This indicates element-wise multiplication.

[0010] Preferably, the process of obtaining the candidate detection box and corresponding confidence score for the lateral spiracular tube in S2 is as follows: The fused multi-scale features are... The input is a tongue-shaped tube detection head. The head outputs a position regression branch and a class confidence branch at each scale. The position regression branch is used to predict the center coordinates, width, and height parameters of the tongue-shaped tube detection box, and the class confidence branch is used to output the probability value of the detection box being a tongue-shaped tube. The detection head adopts a prediction method that combines multi-scale anchored boxes and anchorless boxes.

[0011] Preferably, the confidence threshold screening and non-maximum suppression process in S2 is as follows: based on the confidence of the tongue tube output by the detection model, detection boxes with confidence below a preset threshold are removed; the remaining candidate detection boxes are sorted from high to low confidence, the cross-union ratio between any two detection boxes is calculated, and when the cross-union ratio is greater than the set overlap threshold, only the detection boxes with higher confidence are retained, and the remaining overlapping boxes are suppressed.

[0012] Preferably, the process of determining whether the detection frames between slices belong to the same tongue-shaped tube in S3 is as follows: Slices with tongue-shaped tube detection frames are numbered according to their spatial order along the x-axis, let the number be... The x-axis coordinate of the slice in the three-dimensional coordinate system is: ; When two adjacent slices satisfy: ,in, If the preset slice interval threshold is used, the corresponding slice will be used as a candidate associated slice pair. Let the detection boxes of two adjacent slices in a candidate associated slice pair be respectively and The ratio of their overlapping areas is: ; When the overlapping area ratio satisfies: ,in, If the preset overlap area threshold is used, then adjacent detection boxes are determined to have high continuity in spatial location; For the sequence of center points of the detection box in n consecutive slices Calculate the displacement vector of adjacent center points ; When the displacement vector satisfies ,in, The preset threshold for smooth change; When adjacent detection boxes in a candidate associated slice pair simultaneously satisfy: If the adjacent detection frames belong to the same tongue-side tube, then it is determined that they are adjacent to the same tongue-side tube.

[0013] Preferably, the calculation process of the center point coordinates and radius of the detection box in S3 is as follows: sort the slices in the detection box sequence according to their spatial position in the x-axis direction; select the slice located in the middle of the detection box sequence as the slice with the most central position according to the distribution of the detection box sequence in the x-axis direction; obtain the corresponding detection box in the slice; calculate the center point coordinates according to the geometric position of the detection box in the slice plane; calculate the equivalent radius according to the size information of the detection box in the slice plane; the equivalent radius is the arithmetic mean of the major axis radius and the minor axis radius.

[0014] Preferably, the process of three-dimensional positioning and reconstruction of the tongue-side tube in S3 is as follows: based on the spatial coordinates of the center point in the three-dimensional coordinate system and the corresponding equivalent radius information, the position and spatial scale of the tongue-side tube in three-dimensional space are determined. The spatial scale includes the equivalent tube diameter and axial extension length of the tongue-side tube. The center point sequence of the same tongue-side tube in adjacent slices is fitted with a B-spline curve, and the spatial continuity information of the tongue-side tube in adjacent slices is fused to realize the positioning and reconstruction of the tongue-side tube in three-dimensional space. The fitting error of the B-spline curve is ≤0.3mm.

[0015] The beneficial effects of this invention are as follows: The method automatically extracts tooth regions using a 3D segmentation model and determines the target area containing the lingual canal, avoiding the manual selection of slices or regions required in traditional methods, thus significantly improving the automation and consistency of the detection process. A deep learning model is used to detect the lingual canal in mandibular slice images. Preprocessing enhances local contrast, and confidence filtering and non-maximum suppression strategies effectively improve the accuracy of lingual canal candidate boxes, maintaining high detection stability even in low-contrast, noisy, or complex root overlap conditions. Multiple features, such as slice continuity, overlap area, and centroid position, are fused to determine whether slices belong to the same lingual canal. The optimal slice position is selected for center point and radius calculation, effectively reducing errors caused by single-slice anomalies and improving the reliability of 3D localization. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for detecting the lateral spiracle in CBCT data based on deep learning. The specific steps are as follows: S1: Extracting mandibular slice images from CBCT data The tooth region in CBCT data is automatically segmented using a three-dimensional segmentation model. The target region containing the lingual canal is determined based on the spatial distribution of the teeth, and equidistant continuous two-dimensional slices are generated within this region along the x-axis. S2: Use the detection model to obtain slices of the detection frame with tongue-shaped tubes. The mandibular slice image is preprocessed, and the preprocessed slice is input into a pre-trained lingual canal detection model to obtain candidate detection boxes and corresponding confidence scores for the lingual canal. The candidate detection boxes are then subjected to confidence thresholding and non-maximum suppression to obtain the final detection box of the lingual canal in the slice. S3: Restore the detection results to the three-dimensional coordinate system. Determine whether the detection frames between slices belong to the same tongue-shaped tube. Once it is determined that they belong to the same tongue-shaped tube, select the slice with the most central position in the detection frame sequence and calculate the coordinates and radius of its detection frame center point. Then, map the center point back to the three-dimensional coordinate system to realize the three-dimensional positioning and reconstruction of the tongue-shaped tube.

[0019] The target area of ​​the lingual canal is automatically delineated using a 3D segmentation model, eliminating the need for doctors to manually examine CBCT slices layer by layer. This completely solves the problems of time-consuming manual identification and the risk of missed detection. Furthermore, confidence thresholding and non-maximum suppression effectively eliminate noise interference, improving the accuracy of lingual canal detection and reducing the risk of intraoperative damage to blood vessels and nerves. Compared to traditional algorithms such as threshold segmentation and edge detection, this solution relies on a deep learning model to complete target area segmentation and lingual canal detection, overcoming the sensitivity of traditional algorithms to equipment and individual patient differences. It can stably output detection results under different CBCT data input conditions, significantly improving the robustness of the detection method. Furthermore, by fusing multiple features such as slice continuity, overlap area, and centroid position, it determines whether slices belong to the same lingual canal and selects the slice with the optimal position for center point and radius calculation, effectively reducing errors caused by single-slice anomalies and improving the reliability of 3D localization.

[0020] In S1, the CBCT data is downsampled to 0.4 spacing during preprocessing. The 3D segmentation model is nnUNet, which is used to obtain the 3D mask of the teeth. Based on the tooth segmentation results, the distribution boundary of the teeth in 3D space is calculated. The target area of ​​the lingual canal is estimated according to the relative positional relationship between the teeth and the lingual canal. The generated 2D slices have the same spatial resolution and pixel spacing as the original CBCT data.

[0021] 0.4spacing downsampling reduces computational load and improves processing efficiency while ensuring no loss of lingual canal details. nnUNet is used for automatic tooth segmentation without manual intervention, offering high segmentation accuracy and robustness. The target area is calculated based on the relative position of the teeth and lingual canals, significantly reducing the detection range and minimizing noise interference. The 2D slices maintain the same spatial resolution and pixel spacing as the original CBCT, ensuring the accuracy of subsequent 3D localization and reconstruction.

[0022] The preprocessing of the mandibular slice image in S2 is as follows: Gaussian filtering is used to reduce speckle noise in the CBCT image while preserving the detailed structure of the trabecular bone; CLAHE is used to enhance the mandibular region to improve the contrast between the bone wall around the lingual canal and the surrounding tissue.

[0023] Gaussian filtering can accurately remove speckle noise from CBCT images while avoiding the loss of fine structures such as trabeculae, thus preserving key anatomical information for lingual canal detection. CLAHE enhancement processing for the mandibular region can directionally enhance the grayscale contrast between the lingual canal and surrounding tissues, solving the problem of indistinct grayscale features of the lingual canal. This significantly improves the accuracy and stability of subsequent detection models in recognizing small targets like the lingual canal, reducing the risk of missed detections.

[0024] In S2, the lingual tube detection model adopts a YOLOv11 network structure with a cross-scale memory module. The training process of the lingual tube detection model includes: collecting 3,000 CBCT slices containing the median and lateral lingual tubes, and using Labelme software to annotate the lingual tube detection boxes; data augmentation is performed on the original samples, and the data augmentation adopts the YOLOv11 data augmentation strategy, including random translation, random scaling, illumination perturbation, color enhancement, Mosaic stitching enhancement, and MixUp image mixing enhancement.

[0025] The cross-scale memory module empowers YOLOv11, enhancing the capture and memorization of cross-scale features of small targets such as lateral spirilla tubes and improving detection accuracy; 3,000 labeled samples containing two types of lateral spirilla tubes, coupled with precise Labelme annotation, provide sufficient and high-quality data support for model training.

[0026] The working process of the cross-scale memory module is as follows: Let the input image be... ,in and These represent the height and width of the image, respectively. Extract feature representations at S scales from the Backbone network. ,in Indicates the scale quantity. Let s be the height and width of the scale s. For the number of channels, Backbone will input... Mapping to multi-scale features ; pass A linear transformation is performed on the number of channels at different scales. Implement as Convolution or learnable linear projections make all scales map to the same memory space dimension D, i.e. ; Construct a unified memory structure Each memory unit Consistent with the feature resolution at the corresponding scale; In the forward inference phase, the stability of the current features is enhanced through attention-based weighted fusion, and the fusion formula is as follows: ,in The matching weight is obtained by dot product attention. This indicates element-wise multiplication.

[0027] The cross-scale memory module maps features of different scales to a unified-dimensional memory space through multi-scale feature extraction and channel linear transformation, eliminating feature fragmentation caused by scale differences. The attention-weighted fusion strategy during forward inference can enhance the key feature representation of small targets in the lingual tube, suppress background noise interference, and significantly improve the detection accuracy and stability of YOLOv11 for small anatomical structures. By enhancing the memory information of high-attention areas, the discrimination ability of weak texture areas in the lingual tube can be improved.

[0028] The process of generating candidate detection boxes and corresponding confidence scores for the lingual tube in S2 is as follows: The fused multi-scale features are... The input is a tongue-shaped tube detection head. The head outputs a position regression branch and a class confidence branch at each scale. The position regression branch is used to predict the center coordinates, width, and height parameters of the tongue-shaped tube detection box, and the class confidence branch is used to output the probability value of the detection box being a tongue-shaped tube. The detection head uses a prediction method that combines multi-scale anchored boxes and anchorless boxes.

[0029] The detection head is divided into two branches: location regression and category confidence, which enables precise decoupling of the detection box parameters and target probability of the lateral lateral tube, improving the accuracy of detection, localization and classification. The prediction method that combines multi-scale anchor boxes and no anchor boxes not only adapts to the detection needs of lateral lateral tubes of different sizes, but also avoids the limitation of a single anchor box for small targets, greatly enhancing the robustness of the model for detecting small anatomical structures such as lateral lateral tubes.

[0030] The process of confidence threshold screening and non-maximum suppression in S2 is as follows: based on the confidence of the tongue tube output by the detection model, detection boxes with confidence below the preset threshold are removed; the remaining candidate detection boxes are sorted from high to low confidence, and the cross-union ratio between any two detection boxes is calculated. When the cross-union ratio is greater than the set overlap threshold, only the detection boxes with higher confidence are retained, and the remaining overlapping boxes are suppressed.

[0031] Confidence threshold screening can directly eliminate low-confidence spurious detection boxes, reduce invalid interference, and improve the purity of detection results; non-maximum suppression accurately eliminates duplicate detection boxes of the same tongue tube through confidence ranking and cross-union ratio calculation, avoids redundant annotation, and ensures the uniqueness and accuracy of detection boxes.

[0032] In S3, the process of determining whether the detection frames between slices belong to the same tongue-shaped tube is as follows: Slices with tongue-shaped tube detection frames are numbered according to their spatial order along the x-axis. Let the number of slices be... The x-axis coordinate of the slice in the three-dimensional coordinate system is: ; When two adjacent slices satisfy: ,in, If the preset slice interval threshold is used, the corresponding slice will be used as a candidate associated slice pair. Let the detection boxes of two adjacent slices in a candidate associated slice pair be respectively and The ratio of their overlapping areas is: ; When the overlapping area ratio satisfies: ,in, If the preset overlap area threshold is used, then adjacent detection boxes are determined to have high continuity in spatial location; For the sequence of center points of the detection box in n consecutive slices Calculate the displacement vector of adjacent center points ; When the displacement vector satisfies ,in, The preset threshold for smooth change; When adjacent detection boxes in a candidate associated slice pair simultaneously satisfy: If the adjacent detection frames belong to the same tongue-side tube, then it is determined that they are adjacent to the same tongue-side tube.

[0033] By using a triple threshold determination based on slice spatial interval, detection box overlap area ratio, and center point displacement vector smoothness, a rigorous association rule for tongue-shaped tube detection boxes was constructed. This rule accurately identifies continuous slice detection boxes of the same tongue-shaped tube and effectively eliminates interference from discrete pseudo-detection boxes. The progressive triple determination conditions ensure the spatial continuity and consistency of the detection box sequence, providing high-precision and high-reliability two-dimensional detection box sequence support for subsequent three-dimensional localization and reconstruction of the tongue-shaped tube.

[0034] The calculation process for the center point coordinates and radius of the detection box in S3 is as follows: sort the slices in the detection box sequence according to their spatial position in the x-axis direction; select the slice located in the middle of the detection box sequence as the slice with the most central position based on the distribution of the detection box sequence in the x-axis direction; obtain the corresponding detection box in the slice; calculate the center point coordinates based on the geometric position of the detection box in the slice plane; calculate the equivalent radius based on the size information of the detection box in the slice plane; the equivalent radius is the arithmetic mean of the major axis radius and the minor axis radius.

[0035] By sorting the slices in the middle of the sequence according to their spatial position along the x-axis, the defects of edge slices being susceptible to noise interference are avoided, ensuring the representativeness and reliability of the selected detection frame. The coordinates of the center point are calculated based on the geometric position of the detection frame, and the equivalent radius is calculated by combining the arithmetic mean of the major and minor axis radii. The quantitative indicators are clear and the calculation logic is rigorous, accurately representing the spatial position and morphological dimensions of the tongue tube, providing high-precision and standardized parameter support for subsequent three-dimensional coordinate mapping and structural reconstruction.

[0036] The process of three-dimensional positioning and reconstruction of the tongue-side tube in S3 is as follows: Based on the spatial coordinates of the center point in the three-dimensional coordinate system and the corresponding equivalent radius information, the position and spatial scale of the tongue-side tube in three-dimensional space are determined. The spatial scale includes the equivalent tube diameter and axial extension length of the tongue-side tube. The B-spline curve is used to fit the sequence of center points of the same tongue-side tube in adjacent slices, and the spatial continuity information of the tongue-side tube in adjacent slices is fused to realize the positioning and reconstruction of the tongue-side tube in three-dimensional space. The fitting error of the B-spline curve is ≤0.3mm.

[0037] Based on the three-dimensional coordinates of the center point and the equivalent radius, the spatial position and morphological scale of the tongue-side tube can be accurately quantified, and its key parameters such as tube diameter and axial extension length can be presented intuitively. The center point sequence is fitted with a B-spline curve with a fitting error ≤0.3mm, which effectively integrates the spatial continuity information of adjacent slices, ensuring the accuracy and smoothness of the three-dimensional reconstruction of the tongue-side tube.

[0038] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the lateral spiracle in CBCT data based on deep learning, characterized in that, The specific steps are as follows: S1: Extracting mandibular slice images from CBCT data The tooth region in CBCT data is automatically segmented using a three-dimensional segmentation model. The target region containing the lingual canal is determined based on the spatial distribution of the teeth, and equidistant continuous two-dimensional slices are generated within this region along the x-axis. S2: Use the detection model to obtain slices of the detection frame with tongue-shaped tubes. The mandibular slice image is preprocessed, and the preprocessed slice is input into a pre-trained lingual canal detection model to obtain candidate detection boxes and corresponding confidence scores for the lingual canal. The candidate detection boxes are then subjected to confidence thresholding and non-maximum suppression to obtain the final detection box of the lingual canal in the slice. S3: Restore the detection results to the three-dimensional coordinate system. Determine whether the detection frames between slices belong to the same tongue-shaped tube. Once it is determined that they belong to the same tongue-shaped tube, select the slice with the most central position in the detection frame sequence and calculate the coordinates and radius of its detection frame center point. Then, map the center point back to the three-dimensional coordinate system to realize the three-dimensional positioning and reconstruction of the tongue-shaped tube.

2. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: When the CBCT data mentioned in S1 is preprocessed, it is downsampled to 0.4 spacing. The three-dimensional segmentation model adopts nnUNet, and the three-dimensional mask of the teeth is obtained through this model. Based on the tooth segmentation results, the distribution boundary of the teeth in three-dimensional space is calculated. The target area of ​​the lingual canal is estimated according to the relative positional relationship between the teeth and the lingual canal. The generated two-dimensional slices have the same spatial resolution and pixel spacing as the original CBCT data.

3. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The preprocessing of the mandibular slice image described in S2 is as follows: Gaussian filtering is used to reduce speckle noise in the CBCT image while preserving the detailed structure of the trabecular bone; CLAHE is used to enhance the mandibular region to improve the contrast between the bone wall around the lingual canal and the surrounding tissue.

4. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The lingual tube detection model described in S2 adopts a YOLOv11 network structure with a cross-scale memory module. The training process of the lingual tube detection model includes: collecting 3000 CBCT slices containing the median and lateral lingual tubes, and using Labelme software to annotate the lingual tube detection boxes; performing data augmentation on the original samples, which adopts the YOLOv11 data augmentation strategy, including random translation, random scaling, illumination perturbation, color enhancement, Mosaic stitching enhancement, and MixUp image mixing enhancement.

5. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 4, characterized in that: The operation of the cross-scale memory module is as follows: Let the input image be... ,in and These represent the height and width of the image, respectively. Extract feature representations at S scales from the Backbone network. ,in Indicates the scale quantity. Let s be the height and width of the scale s. For the number of channels, Backbone will input... Mapping to multi-scale features ; pass A linear transformation is performed on the number of channels at different scales. Implement as Convolution or learnable linear projections make all scales map to the same memory space dimension D, i.e. ; Construct a unified memory structure Each memory unit Consistent with the feature resolution at the corresponding scale; In the forward inference phase, the stability of the current features is enhanced through attention-based weighted fusion, and the fusion formula is as follows: ,in The matching weight is obtained by dot product attention. This indicates element-wise multiplication.

6. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The process of obtaining the candidate detection box and corresponding confidence score for the lateral spiracular tube in S2 is as follows: The fused multi-scale features... The input is a tongue-shaped tube detection head. The head outputs a position regression branch and a class confidence branch at each scale. The position regression branch is used to predict the center coordinates, width, and height parameters of the tongue-shaped tube detection box, and the class confidence branch is used to output the probability value of the detection box being a tongue-shaped tube. The detection head adopts a prediction method that combines multi-scale anchored boxes and anchorless boxes.

7. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The confidence threshold screening and non-maximum suppression process described in S2 is as follows: based on the confidence of the tongue tube output by the detection model, detection boxes with confidence levels lower than the preset threshold are removed; the remaining candidate detection boxes are sorted from high to low confidence levels, and the cross-union ratio between any two detection boxes is calculated. When the cross-union ratio is greater than the set overlap threshold, only the detection boxes with higher confidence levels are retained, and the remaining overlapping boxes are suppressed.

8. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The process described in S3 for determining whether the detection frames between slices belong to the same tongue-shaped tube is as follows: Slices with tongue-shaped tube detection frames are numbered according to their spatial order along the x-axis. Let the number of slices be... The x-axis coordinate of the slice in the three-dimensional coordinate system is: ; When two adjacent slices satisfy: ,in, If the preset slice interval threshold is used, the corresponding slice will be used as a candidate associated slice pair. Let the detection boxes of two adjacent slices in a candidate associated slice pair be respectively and The ratio of their overlapping areas is: ; When the overlapping area ratio satisfies: ,in, If the preset overlap area threshold is used, then adjacent detection boxes are determined to have high continuity in spatial location; For the sequence of center points of the detection box in n consecutive slices Calculate the displacement vector of adjacent center points ; When the displacement vector satisfies ,in, The preset threshold for smooth change; When adjacent detection boxes in a candidate associated slice pair simultaneously satisfy: If the adjacent detection frames belong to the same tongue-side tube, then it is determined that they are adjacent to the same tongue-side tube.

9. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The calculation process of the center point coordinates and radius of the detection box described in S3 is as follows: sort the slices in the detection box sequence according to their spatial position in the x-axis direction; select the slice located in the middle of the detection box sequence as the slice with the most central position based on the distribution of the detection box sequence in the x-axis direction; obtain the corresponding detection box in the slice; calculate the center point coordinates based on the geometric position of the detection box in the slice plane; calculate the equivalent radius based on the size information of the detection box in the slice plane; the equivalent radius is the arithmetic mean of the major axis radius and the minor axis radius.

10. The method for detecting the lateral spiracle in CBCT data based on deep learning according to claim 1, characterized in that: The process of three-dimensional positioning and reconstruction of the tongue-side tube described in S3 is as follows: Based on the spatial coordinates of the center point in the three-dimensional coordinate system and the corresponding equivalent radius information, the position and spatial scale of the tongue-side tube in three-dimensional space are determined. The spatial scale includes the equivalent tube diameter and axial extension length of the tongue-side tube. The center point sequence of the same tongue-side tube in adjacent slices is fitted using a B-spline curve, and the spatial continuity information of the tongue-side tube in adjacent slices is fused to realize the positioning and reconstruction of the tongue-side tube in three-dimensional space. The fitting error of the B-spline curve is ≤0.3mm.

Images