A three-dimensional implant crown intelligent selection and generation method based on U-Net

Abstract

The application provides a three-dimensional implant crown intelligent selection and generation method based on U-Net, comprising the following steps: converting a three-dimensional tooth model in STL format into a high-resolution two-dimensional projection image; taking a U-Net network structure with VGG16 as a backbone network, combining a bar pooling module, and designing a neural network model for accurately capturing tooth sequential arrangement features; training the neural network model by using an adaptive learning rate adjustment strategy and a cosine annealing learning rate scheduling strategy, and performing semantic segmentation on the two-dimensional projection image to obtain a 2D tooth classification result image; identifying missing tooth labels based on the 2D tooth classification result image, calculating the center points of left and right adjacent teeth of the missing tooth, calculating the center position coordinates of the missing tooth in the 2D image through the center points of the left and right adjacent teeth, and mapping the center position coordinates to a three-dimensional space to obtain three-dimensional center coordinates of the missing tooth; and moving the template tooth center point to the three-dimensional center coordinates of the missing tooth to generate a three-dimensional crown model that matches the missing tooth position.

Description

Technical Field

[0001] This invention relates to the field of oral medical image processing and computer-aided design technology, and more particularly to a method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net. Background Technology

[0002] The intelligent design and fabrication of dental implant crowns is a crucial aspect of dental implant treatment. With the development of digital dental technology, automated crown design methods based on 3D intraoral scan data and deep learning have gradually become a research hotspot. However, existing technologies still face multiple technical bottlenecks in practical clinical applications, specifically in the following aspects:

[0003] Cone-beam computed tomography (CBCT), a key imaging technology for dental implantology, faces significant challenges in its automated tooth segmentation methods. Current techniques lack generalization ability when dealing with imaging artifacts and individual anatomical differences, often requiring manual correction and hindering end-to-end fully automated processing. Studies have shown that artifacts and scattering phenomena caused by metal implants in CBCT images severely degrade image quality, manifesting as streaks and shadows that directly affect segmentation accuracy.

[0004] Teeth in corresponding positions on the upper and lower jaws (such as the left mandibular first molar and the right mandibular first molar) and adjacent teeth are often confused, leading to incorrect localization of the missing tooth area. Existing general segmentation networks (such as U-Net) do not fully consider the U-shaped and discrete distribution characteristics of teeth on the dental arch, making it difficult to model global spatial relationships and limiting recognition accuracy.

[0005] U-Net is one of the most widely used deep learning architectures in medical image segmentation, and it also performs well in tooth segmentation tasks. Studies have shown that models based on U-Net and its variants can achieve high pixel accuracy in dental X-ray images and intraoral scan models. However, existing U-Net architectures still have limitations in handling tooth overlap, morphological variations, and artifact interference, especially in the high confusion rate of recognizing corresponding teeth in the upper and lower jaws. Tooth segmentation tasks can be divided into semantic segmentation (dividing the image into tooth and non-tooth regions) and instance segmentation (assigning independent labels to each tooth). Instance segmentation is more critical for implant localization, but it is also more difficult to implement. Existing methods have low accuracy when dealing with crowded teeth, varied third molar morphology, and incompletely erupted teeth, mainly because of insufficient training data for these complex cases and the susceptibility of dental X-ray images to artifact interference and insufficient clarity.

[0006] When converting a 3D STL model into a 2D projection image, traditional methods do not consider the occlusion relationship in the height direction (Y-axis), which causes points that should be at the top of the projection image to be covered by points at the bottom, resulting in poor projection image quality and affecting the accuracy of subsequent recognition.

[0007] Projecting a 3D mesh model onto a 2D plane is a crucial preprocessing step in 3D tooth analysis. Traditional projection methods draw the triangular faces directly in order, neglecting the occlusion relationships between faces of different heights. This results in lower faces being covered by upper faces, leading to information loss. For 3D models lacking surface information, depth information needs to be introduced when projecting onto a 2D plane to compensate for the loss of depth information caused by dimensionality reduction. In dental intraoral scan models, the distribution of triangular faces along the Y-axis (vertical direction) varies significantly. If they are not drawn in Y-axis order, the normal vector color points of lower faces will be covered by points above, resulting in the loss of key geometric features in the projected image.

[0008] Existing methods are mostly based on two-dimensional images or manual marking of missing tooth locations using CBCT, lacking precise algorithms for automatically mapping two-dimensional segmentation results to three-dimensional space, making it difficult to achieve automated positioning and movement of the template tooth.

[0009] Traditional dental implant crown design primarily relies on computer-aided design (CAD) technology. Dentists must select a suitable template tooth from a morphological library and then manually adjust the template's position, orientation, and details. While the introduction of CAD technology has significantly improved design efficiency, early tooth templates still had issues with morphological style and deviations from the target tooth shape, making the design process still time-consuming. In recent years, researchers have begun exploring automated methods based on deep learning, such as generating 3D point cloud reconstructions of missing teeth by predicting the tooth centroid and combining it with the geometric constraints of adjacent teeth. Studies have shown that deep learning-based template generation methods can achieve higher restoration accuracy compared to traditional CAD methods, providing personalized templates that more closely approximate the original morphology of the damaged tooth. However, existing methods still require improvement in areas such as automated estimation of the center point of the edentulous area and accurate mapping from 2D segmentation results to 3D space.

[0010] First proposed in 2020, strip pooling aims to capture long-range dependencies in images, making it particularly suitable for scene parsing tasks with anisotropic structures. Compared to traditional square pooling kernels, strip pooling uses long and narrow kernels, enabling it to better model the contextual information of elongated or discretely distributed objects.

[0011] In the field of medical image segmentation, strip pooling technology has been successfully applied to tasks such as retinal vessel segmentation, effectively improving segmentation accuracy. Similarly, in the glomerular basement membrane segmentation task, researchers also found that the target shape is strip-shaped. By introducing a strip pooling module to extract strip and square features, high-precision segmentation results were obtained. However, the application of strip pooling technology to the field of tooth recognition, especially the specialized network design for the sequential arrangement features of teeth on the dental arch curve, has not yet been systematically studied.

[0012] In summary, existing intelligent design and fabrication technologies for dental implant crowns have the following main drawbacks: (1) CBCT image segmentation is susceptible to metal artifacts, has insufficient generalization ability, and is difficult to achieve fully automated processing; (2) The general U-Net network does not fully consider the U-shaped and discrete distribution characteristics of teeth, resulting in serious confusion in the identification of corresponding teeth and adjacent teeth in the upper and lower jaws; (3) The occlusion relationship of the Y-axis was not considered when the 3D STL model was projected to the 2D model, resulting in the loss of geometric features; (4) The estimation of the center point of the missing tooth area relies on manual intervention and lacks an accurate automatic mapping algorithm from the two-dimensional segmentation results to the three-dimensional space; (5) Strip pooling technology has not been systematically applied to the field of tooth recognition and has failed to effectively model the global spatial relationship of the sequential arrangement of teeth on the dental arch curve.

[0013] Therefore, there is an urgent need for a fully automated, high-precision method that does not rely on traditional CBCT images and can achieve end-to-end generation of three-dimensional crown models adapted to the missing tooth location from raw STL data. Summary of the Invention

[0014] To address the technical problems in existing technologies, such as CBCT image segmentation being susceptible to metal artifacts, severe confusion in tooth recognition using the general U-Net network, loss of 2D projection geometric features in 3D STL models, reliance on manual intervention for center point estimation of edentulous areas, and the lack of systematic application of strip pooling technology in the field of tooth recognition, this invention provides a U-Net-based intelligent selection and generation method for 3D implant crowns. This invention mainly utilizes 2D projection image generation technology from 3D intraoral scan data, an improved U-Net network based on a VGG16 encoder and integrating a strip pooling module, a center point estimation method for edentulous areas based on weighted average of adjacent tooth center points, and automatic translation and positioning technology for template teeth. This achieves the technical effect of end-to-end fully automatic generation of 3D crown models adapted to the edentulous location from raw STL data.

[0015] The technical means employed in this invention are as follows: A U-Net-based intelligent selection and generation method for three-dimensional dental implant crowns, comprising: S1. Based on the STL format, a 3D tooth model is generated by calculating and normalizing the normal vector of each triangular facet so that each triangular facet corresponds to a pixel. The 3D normal vector information is mapped to the color space to generate a high-resolution scatter plot, thereby generating a 2D projection map that reflects the normal vector characteristics of the 3D model. S2. Using VGG16 as the backbone network, a U-Net network structure is combined with a strip pooling module to design a neural network model that accurately captures the sequential arrangement features of teeth. S3. An adaptive learning rate adjustment strategy and a cosine annealing learning rate scheduling strategy are used to train the neural network model, and the trained neural network model is used to perform semantic segmentation on the two-dimensional projection image to identify each tooth category and obtain a 2D tooth classification result image. S4. Based on the 2D tooth classification result image, identify the missing tooth label, calculate the center point of the left and right adjacent teeth of the missing tooth, calculate the center position coordinate of the missing tooth in the 2D image through the center point of the left and right adjacent teeth, and map it to the three-dimensional space to obtain the three-dimensional center coordinate of the missing tooth. S5. Move the center point of the template tooth to the three-dimensional center coordinates of the missing tooth to generate a three-dimensional crown model that matches the position of the missing tooth.

[0016] Further, step S1 includes: S11. Considering that the 3D STL format tooth model to be detected is a collection of triangular facets, each with three vertices and a normal vector, therefore, it will have... The 3D model data of each triangular facet is represented as a set. :

[0017] in, Indicates the first A triangular facet, , Indicates the first The three vertices of a triangle. Indicates the first The normal vectors of the triangular faces. Depend on Three components Normal vector 3D coordinate vector; Indicates the first The first triangular face One vertex, , Represents vertices The three-dimensional coordinate components; S12, Calculate each triangular face The center of gravity is determined by the following formula:

[0018] in, Indicates the center of gravity The three-dimensional coordinate components, , , ; S13, Regarding the first Normal vectors of each face The three components are used to generate the normal vectors of all triangle faces. coordinate, coordinates and The set of minimum and maximum coordinate values ​​is given by the following formula:

[0019]

[0020] in, Represents the normal vector of all triangle faces. The set of minimum values ​​of each component. Represents the normal vector of all triangle faces. The set of maximum values ​​of each component. ,correspond , Indicates the first The first normal vector of the triangular face One component; S14. Calculate the normal vector components of all triangle faces. Normalize to the [0,1] interval to generate the normal vector, as follows:

[0021] in, This represents the normalized normal vector.

[0022] S15. Based on the center of gravity coordinates Sort all triangle faces in ascending order to ensure Points with smaller values ​​are projected first. Project the points with larger values ​​back; S16. Define a two-dimensional projected image as a collection The set of points is given by the following formula:

[0023] Each point It is a two-dimensional data. ; S17, Regarding according to Each triangle face after value sorting The calculated center points of the triangular faces are then arranged in sequence. Mapped to A plane is used to generate a two-dimensional projection point. To obtain the projection point coordinates and Coordinates are as follows:

[0024]

[0025] Finally, 3D model data is generated. Two-dimensional projection of the normal vector of the plane.

[0026] Further, step S2 includes: S21. Construct a U-Net network with VGG16 as the encoder. The encoder adopts a 5-layer downsampling structure, including 13 convolutional layers and 5 max pooling layers; the decoder adopts a 4-layer upsampling structure. S22. Add a strip pooling module to each downsampling structure of the encoder to collect long-distance contextual information from different spatial dimensions using horizontal and vertical strip pooling operations.

[0027] Further, step S22 includes: S221, Let For a two-dimensional input tensor, where and These represent the spatial height and width of the two-dimensional input tensor, respectively, and the bar pooling module uses... The pooling window performs a horizontal bar pooling operation and outputs... The formula is as follows:

[0028] in, This indicates the first horizontal bar pooling step. The height feature value of the row; S222, Let For a two-dimensional input tensor, where and These represent the spatial height and width of the two-dimensional input tensor, respectively, and the bar pooling module uses... The pooling window performs a vertical bar pooling operation and outputs... The formula is as follows:

[0029] in, This indicates the first vertical bar pooling step. The width characteristic value of the column; S223, Let For the input tensor, where Indicates the number of channels. and The height and width of the input tensor space are respectively... The input is fed into two parallel paths, each containing a horizontal or vertical strip pooling layer followed by a one-dimensional convolutional layer with a kernel size of 3 to modulate the current position and its neighboring features. The output is... and ; S224, Output and By performing fusion, fusion characteristics are obtained. The fusion formula is as follows:

[0030] in, This indicates that the fused feature tensor is in the channel. ,Location eigenvalues ​​at that location This indicates that the horizontal pooling output is in the channel. , No. The eigenvalues ​​of the row Indicates the vertical pooling output in the channel , No. The characteristic values ​​of the column; S225, Based on fusion features The calculation yields an output that contains more global prior information. The formula is as follows:

[0031] in, This represents element-wise multiplication. This represents the Sigmoid function. Represents a 1×1 convolution; S226. Use the Sigmoid function to map the convolution output to... The interval, expressed as follows:

[0032] in, Indicates the input value. The output range is in Between these, the generated spatial attention weights are used to reweight the input features.

[0033] Further, step S3 includes: S31. An adaptive learning rate adjustment strategy is adopted to normalize the actual batch size with the baseline batch size, dynamically adjusting the initial learning rate to ensure that the neural network model obtains consistent convergence characteristics under different training configurations. The specific process is as follows: The adjusted initial learning rate is calculated using the following formula:

[0034] in, This indicates the actual batch size during the current training phase. Indicates the baseline batch size. This represents the user-defined initial learning rate. This represents the maximum learning rate limit associated with the optimizer. This represents the minimum learning rate lower bound related to the optimizer. This represents the adjusted actual initial learning rate. This represents the adjusted actual minimum learning rate. The function will input values Limited to the range Inside, ; Calculate the adjusted minimum learning rate The calculation formula is as follows:

[0035] S32. A cosine annealing learning rate scheduling strategy is adopted to simulate the changing law of the cosine function, so that the learning rate smoothly decays from the initial value to the minimum value, avoiding training oscillations caused by step-like descent, and promoting the neural network model to converge to the optimal solution more stably. The specific process is as follows: Calculate the number of preheating stages The formula is as follows:

[0036] in, Indicates the total number of training rounds; Calculate the number of rounds in the smooth termination phase The formula is as follows:

[0037] Based on the adjusted actual initial learning rate Calculate the warm-up initial learning rate The formula is as follows:

[0038] Based on the number of preheating stages The number of rounds in the smooth conclusion phase and warm-up initial learning rate The learning rate change curve is described using a piecewise function, as follows:

[0039] in, This indicates the current training round number.

[0040] Further, step S4 includes: S41. Based on the classification results of a single tooth in the two-dimensional projection image, 8-connected regions are extracted from the classification regions of different tooth labels. The extracted connected regions that satisfy the minimum patch area threshold minPatchSize are statistically analyzed. If a tooth category is missing, that tooth category label is the missing tooth label. Specifically, this includes: Define tooth category space The numbering of all teeth is denoted as... ; For tooth category space any of the following ,calculate The maximum number of pixels in the 8-connected region corresponding to the class ,like Then Add to collection middle; Search Sets There is a lack of space for tooth categories in China. Which category label in, i.e. ,but That is, a set consisting of categories of missing teeth; S42. Calculate the center points of the left and right adjacent teeth of the missing tooth, specifically including: Assume the missing tooth is numbered The tooth numbers of its left and right adjacent teeth are and ,like ,but , Otherwise, if the tooth is an incisor or the last molar, a non-molar area will be indicated, and no action will be taken. In the two-dimensional image of tooth classification results, the teeth labeled as follows are extracted respectively. and For the tooth region sub-image, the 8-connected region extraction method was used to count all connected regions in the left and right adjacent tooth region sub-images. Considering the possibility of misclassification, the connected region of the left and right adjacent teeth with the largest patch area was regarded as the actual range of the left and right adjacent teeth, while other smaller connected regions were regarded as noise removal. set up It is the set of pixels within all connected regions of a given neighboring tooth, containing... If there are 100 pixels, then define The average value of each pixel is the center of the adjacent tooth. The formula is as follows:

[0041] S43. The coordinates of the center position of the missing tooth in the 2D image are calculated using the center points of the left and right adjacent teeth, specifically including: Let the area of ​​a neighboring tooth in the missing region be... It is the sum of the number of pixels in the connected region corresponding to the adjacent tooth. That is ;area The calculation formula is as follows:

[0042] Since molars are generally circular in shape, we assume that the shape of this region is approximately circular, and then based on the area... The approximate diameter of the tooth was calculated. The formula is as follows:

[0043] Assuming the diameter of the tooth adjacent to the missing tooth is... The central position is The diameter of the right adjacent tooth is The central position is Calculate the center coordinates of the missing tooth in the 2D image. The formula is as follows:

[0044]

[0045] S44. Calculate the location of missing teeth in the 3D model, specifically including: Suppose a 3D STL model is in The maximum and minimum boundaries on the plane are If the two-dimensional projected image Width in the axial direction is , The height in the axial direction is Then calculate and Directional scaling factor and The formula is as follows:

[0046]

[0047] Then the coordinates of the two-dimensional image To three-dimensional coordinates The linear mapping relationship is as follows: ; Define with A 3D tooth model with vertices ,as follows:

[0048] in, ; Using two-dimensional image coordinates To three-dimensional coordinates linear mapping relationship Calculate a three-dimensional tooth model Coordinates of the center of the missing middle tooth The formula is as follows:

[0049]

[0050]

[0051] in, express Scale factor in the axial direction express Scale factor in the axial direction.

[0052] Further, step S5 includes: S51. Enter the missing tooth number. Coordinates of the center of the missing tooth location in the 3D tooth model dental implant template data set , ; S52, Regarding the dental implant template data set Calculate the centroid coordinates of the template tooth. The formula is as follows:

[0053] in, This indicates the number of triangular faces in the template tooth model; S53, the center point of the obtained dental implant template. Move to the center coordinates of the missing tooth location in the 3D tooth model Calculate the displacement of the center point of the dental implant template tooth. The calculation formula is as follows:

[0054] S54, Center point of the dental implant template With displacement Add them together to get ; S55, The points in the original dental implant template... move ,get The new coordinates of all points in the implant template tooth are determined using the following formula:

[0055] S56. After the above movements, a three-dimensional crown STL model data that matches the position of the missing tooth is obtained. The formula is as follows:

[0056] in, , indicating the translated first A triangular facet, Indicates the translation of the first... The normal vectors of the triangular faces. , representing the coordinates of the new vertex after translation. This represents the three-dimensional coordinates of the vertex after translation.

[0057] Compared with the prior art, the present invention has the following advantages: 1. This invention significantly improves the accuracy of tooth recognition, with the average intersection-union ratio (mIoU) increasing from 87.69% to 90.73%, an improvement of 3.04 percentage points; the average pixel accuracy increasing from 94.53% to 96.35%, an improvement of 1.82 percentage points; and the confusion rate of symmetrical teeth (such as 34 and 44, 45 and 34) is reduced to almost zero.

[0058] 2. The strip pooling module of the present invention can capture long-distance contextual information in the horizontal and vertical directions. Each position in the output tensor can establish a relationship with multiple positions in the input tensor, effectively modeling the global spatial relationship between teeth and reducing confusion between adjacent teeth and teeth of the same number on the left and right.

[0059] 3. The STL file is converted into a 2D projection image, and then a trained neural network model performs tooth recognition on the projection image. After recognition, the center position of the missing tooth is located, and the conversion of the missing tooth position from the 2D image to a 3D model and the movement of the template tooth are completed. This complete process takes about 10 seconds, but traditional manual methods often take more than 30 seconds to determine a relatively accurate center position. Moreover, the missing tooth center position obtained by this invention through precise mathematical calculation is more accurate and has less error compared to the position estimated manually.

[0060] 4. This invention does not rely on traditional CBCT images, but uses a method that combines three-dimensional oral scan data with two-dimensional image processing, effectively avoiding the impact of artifacts and scattering phenomena caused by metal implants on segmentation accuracy.

[0061] 5. This invention is achieved through... Axis sorting ensures that higher points in three-dimensional space ( Larger values ​​cover lower points to avoid projection occlusion and ensure the integrity of key geometric features in the projection map. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of the overall process of the method of the present invention.

[0064] Figure 2 This is a flowchart of the tooth projection process of the present invention.

[0065] Figure 3 This is a diagram showing the effect of generating a two-dimensional projection image of the three-dimensional data of this invention.

[0066] Figure 4 This is a diagram of the traditional U-Net network structure.

[0067] Figure 5 This is a diagram of the VGG16 network structure.

[0068] Figure 6 This is a diagram of the U-Net network structure that incorporates VGG16 in this invention.

[0069] Figure 7 This is a diagram of the neural network architecture of the present invention.

[0070] Figure 8 This is a structural diagram of the strip pooling module of the present invention.

[0071] Figure 9 A tooth numbering diagram provided for an embodiment of the present invention.

[0072] Figure 10 A single tooth recognition effect diagram provided in an embodiment of the present invention. Figure 11 This is a comparison chart of the extraction effects of the strip pooling module of the present invention.

[0073] Figure 12 This is a comparison diagram of tooth confusion in the strip pooling module of the present invention.

[0074] Figure 13 This is a flowchart illustrating the process of estimating the center point location of a three-dimensional model of a missing tooth in this invention.

[0075] Figure 14 This is a schematic diagram illustrating the calculation of the center point of the tooth-deficient region in this invention.

[0076] Figure 15 This is a comparison diagram of the coordinate mapping between the three-dimensional model and the two-dimensional projected image of this invention.

[0077] Figure 16 An example diagram illustrating the location of missing teeth provided in an embodiment of the present invention.

[0078] Figure 17 This is a three-dimensional dental implant crown template generation diagram provided in an embodiment of the present invention. Detailed Implementation

[0079] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0080] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product or device.

[0081] like Figure 1 As shown, this invention provides a method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net, including: S1. Based on the STL format, a 3D tooth model is generated by calculating and normalizing the normal vector of each triangular facet so that each triangular facet corresponds to a pixel. The 3D normal vector information is mapped to the color space to generate a high-resolution scatter plot, thereby generating a 2D projection map that reflects the normal vector characteristics of the 3D model. S2. Using VGG16 as the backbone network, a U-Net network structure is combined with a strip pooling module to design a neural network model that accurately captures the sequential arrangement features of teeth. S3. An adaptive learning rate adjustment strategy and a cosine annealing learning rate scheduling strategy are used to train the neural network model, and the trained neural network model is used to perform semantic segmentation on the two-dimensional projection image to identify each tooth category and obtain a 2D tooth classification result image. S4. Based on the 2D tooth classification result image, identify the missing tooth label, calculate the center point of the left and right adjacent teeth of the missing tooth, calculate the center position coordinate of the missing tooth in the 2D image through the center point of the left and right adjacent teeth, and map it to the three-dimensional space to obtain the three-dimensional center coordinate of the missing tooth. S5. Move the center point of the template tooth to the three-dimensional center coordinates of the missing tooth to generate a three-dimensional crown model that matches the position of the missing tooth.

[0082] In specific implementation, as a preferred embodiment of the present invention, such as Figure 2 As shown, step S1 includes: S11. Considering that the 3D STL format tooth model to be detected is a collection of triangular facets, each with three vertices and a normal vector, therefore, it will have... The 3D model data of each triangular facet is represented as a set. :

[0083] in, Indicates the first A triangular facet, , Indicates the first The three vertices of a triangle. Indicates the first The normal vectors of the triangular faces. Depend on Three components Normal vector 3D coordinate vector; Indicates the first The first triangular face One vertex, , Represents vertices The three-dimensional coordinate components; S12, Calculate each triangular face The center of gravity is determined by the following formula:

[0084] in, Indicates the center of gravity The three-dimensional coordinate components, , , ; S13, Regarding the first Normal vectors of each face The three components are used to generate the normal vectors of all triangle faces. coordinate, coordinates and The set of minimum and maximum coordinate values ​​is given by the following formula:

[0085]

[0086] in, Represents the normal vector of all triangle faces. The set of minimum values ​​of each component. Represents the normal vector of all triangle faces. The set of maximum values ​​of each component. ,correspond , Indicates the first The first normal vector of the triangular face One component; S14. Calculate the normal vector components of all triangle faces. Normalize to the [0,1] interval to generate the normal vector, as follows:

[0087] in, This represents the normalized normal vector.

[0088] S15. Based on the center of gravity coordinates Sort all triangle faces in ascending order to ensure Points with smaller values ​​are projected first. Project the points with larger values ​​back; In this embodiment, considering The axis represents the height data of the tooth, and there may be multiple triangular facets whose centroids project onto the same 2D plane point. For example, there may be... and All will be projected onto On a two-dimensional point and To avoid points with low elevation High coverage points For the projection, we need to press The coordinate values ​​(vertical direction in the original 3D space) are sorted, and points with lower coordinates are projected first, followed by points with higher coordinates. After this processing, Points with larger values ​​(higher points in 3D space) will be back-projected and covered. Smaller projection pixel values ​​avoid occlusion during projection.

[0089] S16. Define a two-dimensional projected image as a collection The set of points is given by the following formula:

[0090] Each point It is a two-dimensional data. ; S17, Regarding according to Each triangle face after value sorting The calculated center points of the triangular faces are then arranged in sequence. Mapped to A plane is used to generate a two-dimensional projection point. To obtain the projection point coordinates and Coordinates are as follows:

[0091]

[0092] Finally, three-dimensional model data is generated. A two-dimensional projection of the normal vector of a plane. For example... Figure 3 As shown, by Figure 3 (b) It can be seen that, according to In the 2D projection image after axis coordinate sorting, there is no occlusion issue in the anterior incisors, ensuring that all points projected into the image are aligned with the axes coordinate sorting. Image generation based on axis size order avoids the situation where the color point of the normal vector corresponding to the lower triangle is projected onto a point of the upper triangle, such as... Figure 3 As shown in (c).

[0093] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. Construct a U-Net network with VGG16 as the encoder. The encoder adopts a 5-layer downsampling structure, including 13 convolutional layers and 5 max pooling layers; the decoder adopts a 4-layer upsampling structure. In this embodiment, based on the traditional U-Net network architecture, VGG16 convolutional and pooling layers are combined to serve as the backbone network of the neural network structure for extracting input data features. Figure 4 , Figure 5 These represent the traditional U-Net network architecture and the VGG16 network architecture, respectively. Figure 4 , Figure 5 It is known that the encoder part of the traditional U-Net network structure consists of 5 downsampling modules, each consisting of two convolutional layers and one max-pooling layer, for a total of 10 convolutional layers; while the downsampling part of VGG16 has 13 convolutional layers and 5 pooling layers. Combining the two, using a U-Net network structure with VGG16 as the backbone network, as follows... Figure 6As shown, neural networks have a stronger ability to extract information from input data. In the traditional U-Net network structure, the number of image channels is increased from 512 to 1024 through convolution in the last downsampling module, which leads to a significant increase in the number of subsequent parameters and computational cost. However, the U-Net network structure using VGG16 as the encoder goes through 13 convolutional layers, performing sufficient and progressive feature extraction and compression through layer-by-layer convolution. At this point, the 512-channel feature map already contains very high-level and dense semantic information, eliminating the need to blindly increase the number of channels and reducing the additional computational cost caused by increasing the number of channels.

[0094] S22. Add strip pooling modules to each downsampling structure of the encoder, and use horizontal and vertical strip pooling operations to collect long-distance contextual information from different spatial dimensions. In this embodiment, as shown... Figure 7 As shown, in Figure 6 Building upon this foundation, a strip pooling module was added during the downsampling process, which better models the global spatial relationships between teeth and reduces tooth obfuscation. The core idea of ​​strip pooling is to use a long and narrow pooling kernel to capture long-distance dependencies in the image. Compared with traditional square pooling kernels, strip pooling can better model anisotropic contextual information, making it particularly suitable for scene parsing tasks with long strip-shaped or discretely distributed objects.

[0095] In a specific implementation, as a preferred embodiment of the present invention, step S22 includes: S221, Let For a two-dimensional input tensor, where and These represent the spatial height and width of the two-dimensional input tensor, respectively, and the bar pooling module uses... The pooling window performs a horizontal bar pooling operation and outputs... The formula is as follows:

[0096] in, This indicates the first horizontal bar pooling step. The height feature value of the row; S222, Let For a two-dimensional input tensor, where and These represent the spatial height and width of the two-dimensional input tensor, respectively, and the bar pooling module uses... The pooling window performs a vertical bar pooling operation and outputs... The formula is as follows:

[0097] in, This indicates the first vertical bar pooling step. The width characteristic value of the column; S223, Let For the input tensor, where Indicates the number of channels. and The height and width of the input tensor space are respectively... The input is fed into two parallel paths, each containing a horizontal or vertical strip pooling layer followed by a one-dimensional convolutional layer with a kernel size of 3 to modulate the current position and its neighboring features. The output is... and ; S224, Output and By performing fusion, fusion characteristics are obtained. The fusion formula is as follows:

[0098] in, This indicates that the fused feature tensor is in the channel. ,Location eigenvalues ​​at that location This indicates that the horizontal pooling output is in the channel. , No. The eigenvalues ​​of the row Indicates the vertical pooling output in the channel , No. The characteristic values ​​of the column; S225, Based on fusion features The calculation yields an output that contains more global prior information. The formula is as follows:

[0099] in, This represents element-wise multiplication. This represents the Sigmoid function. Represents a 1×1 convolution; S226. Use the Sigmoid function to map the convolution output to... The interval, expressed as follows:

[0100] in, Indicates the input value. The output range is in Between these, the generated spatial attention weights are used to reweight the input features.

[0101] In the above process, each position in the output tensor is allowed to establish relationships with multiple positions in the input tensor. For example, in Figure 8In the output tensor, the squares surrounded by yellow boxes are connected to all positions with the same horizontal or vertical coordinates (surrounded by red and green boxes). Therefore, by repeating the above aggregation process several times, it is possible to establish long-range dependencies across the entire scene, thereby improving the ability to capture long-range spatial dependencies and utilize inter-channel dependencies.

[0102] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. An adaptive learning rate adjustment strategy is adopted to normalize the actual batch size with the baseline batch size, dynamically adjusting the initial learning rate to ensure that the neural network model obtains consistent convergence characteristics under different training configurations. The specific process is as follows: The adjusted initial learning rate is calculated using the following formula:

[0103] in, This indicates the actual batch size during the current training phase. Indicates the baseline batch size. This represents the user-defined initial learning rate. This represents the maximum learning rate limit associated with the optimizer. This represents the minimum learning rate lower bound related to the optimizer. This represents the adjusted actual initial learning rate. This represents the adjusted actual minimum learning rate. The function will input values Limited to the range Inside, ; Calculate the adjusted minimum learning rate The calculation formula is as follows:

[0104] S32. A cosine annealing learning rate scheduling strategy is adopted to simulate the changing law of the cosine function, so that the learning rate smoothly decays from the initial value to the minimum value, avoiding training oscillations caused by step-like descent, and promoting the neural network model to converge to the optimal solution more stably. The specific process is as follows: Calculate the number of preheating stages The formula is as follows:

[0105] in, Indicates the total number of training rounds; Calculate the number of rounds in the smooth termination phase The formula is as follows:

[0106] Based on the adjusted actual initial learning rate Calculate the warm-up initial learning rate The formula is as follows:

[0107] Based on the number of preheating stages The number of rounds in the smooth conclusion phase and warm-up initial learning rate The learning rate change curve is described using a piecewise function, as follows:

[0108] in, This indicates the current training round number.

[0109] In this embodiment, 168 lower tooth data points were collected. LabelMe was used to create the labeled data, and the images were divided into 11 semantic categories. Each medical label tooth was classified into one type. Since our goal is to generate implant models for teeth 34, 35, 36, 44, 45, and 46, we only considered these 6 teeth and their adjacent teeth, ignoring other teeth. Therefore, there are a total of 10 tooth labels (33, 34, 35, 36, 37, 43, 44, 45, 46, 47). The background category is one category, such as... Figure 9 As shown, Figure 9 (a) is a diagram of tooth identification and labeling. Figure 9 (b) A total of 112 training samples, 33 validation samples, and 23 test samples were created for the tooth color chart. Figure 10 The test experiments shown demonstrate that the network architecture designed in this invention can identify all individual teeth with high accuracy. Furthermore, to verify the effectiveness of the added strip pooling module, a comparative experiment was conducted with and without the module, as shown below. Figure 11 As shown. From Figure 11 (a) It can be seen that before the addition, adjacent upper and lower teeth and corresponding left and right teeth were very easy to identify and confuse. Figure 11 (b) It can be seen that after adding the strip pooling module, the accuracy of recognition is significantly improved, and the confusion between left and right adjacent teeth has basically disappeared. Figure 12 The diagram illustrates the overlap and confusion of teeth before and after adding the strip pooling module. Figure 12 (a) represents the confusion matrix between the predicted and true labels for tooth recognition without the addition of the bar pooling module. Figure 12(b) shows the confusion matrix for tooth recognition after adding the bar pooling module. As can be seen from the figure, before adding this module, tooth 46 was easily confused as tooth 47 and tooth 45, and tooth 45 was easily confused as tooth 34. After adding this module, the above confusion rates decreased significantly. For example, the occurrence of tooth 47 being confused as tooth 45 was almost zero, and the confusion of bilaterally symmetrical teeth such as tooth 44 and tooth 34 also decreased. Furthermore, the overlap and accuracy between the predicted labels and the correct real labels of the teeth also improved to varying degrees.

[0110] To investigate the effect of the strip pooling module on tooth recognition performance and the optimal placement of the module, Figure 6 The VGG16 shown serves as the backbone of the U-Net network, which has five downsampling modules. Bar pooling modules were added to each downsampling module before comparative experiments were conducted. In Table 1, Base U-Net represents the baseline network, and SPM represents the bar pooling module. Stage 2 indicates that a bar pooling module was added after the second downsampling module; the same applies to Stage 3, Stage 4, and Stage 5 in the table.

[0111] Table 1 Ablation analysis of the strip pooling module

[0112] As shown in Table 1, the U-Net baseline network achieved an average crossover union (CUB) of 87.69% on the validation set, while the network with added strip pooling showed improved CUB on the validation set. The architecture that added strip pooling modules after the 2nd, 3rd, 4th, and 5th downsampling modules achieved the most significant improvement, exceeding 3%. Therefore, unlike traditional square pooling kernels, strip pooling uses long and narrow kernels, which can better model the global spatial relationships of sequential tooth arrangement and reduce the confusion rate of left and right symmetrical teeth.

[0113] In a specific implementation, as a preferred embodiment of the present invention, step S4 is as follows: Figure 13 As shown, it includes: S41. Based on the classification results of a single tooth in the two-dimensional projection image, 8-connected regions are extracted from the classification regions of different tooth labels. The extracted connected regions that satisfy the minimum patch area threshold minPatchSize are statistically analyzed. If a tooth category is missing, that tooth category label is the missing tooth label. Specifically, this includes: Define tooth category space The numbering of all teeth is denoted as... ; For tooth category space any of the following ,calculate The maximum number of pixels in the 8-connected region corresponding to the class ,like Then Add to collection middle; Search Sets There is a lack of space for tooth categories in China. Which category label in, i.e. ,but That is, a set consisting of categories of missing teeth; S42. Calculate the center points of the left and right adjacent teeth of the missing tooth, specifically including: Assume the missing tooth is numbered The tooth numbers of its left and right adjacent teeth are and ,like ,but , Otherwise, if the tooth is an incisor or the last molar, a non-molar area will be indicated, and no action will be taken. In the two-dimensional image of tooth classification results, the teeth labeled as follows are extracted respectively. and For the tooth region sub-image, the 8-connected region extraction method was used to count all connected regions (patches) in the left and right adjacent tooth region sub-images. Considering the possibility of misclassification, the left and right adjacent tooth connected region with the largest patch area was regarded as the actual range of the left and right adjacent teeth, and other smaller connected regions were regarded as noise removal. set up It is the set of pixels within all connected regions of a given neighboring tooth, containing... If there are 100 pixels, then define The average value of each pixel is the center of the adjacent tooth. The formula is as follows:

[0114] S43. The coordinates of the center position of the missing tooth in the 2D image are calculated using the center points of the left and right adjacent teeth, specifically including: Let the area of ​​a neighboring tooth in the missing region be... It is the sum of the number of pixels in the connected region corresponding to the adjacent tooth. That is ;area The calculation formula is as follows:

[0115] Since molars are generally circular in shape, we assume that this area is approximately circular in shape, such as... Figure 14 As shown, it is based on area The approximate diameter of the tooth was calculated. The formula is as follows:

[0116] Assuming the diameter of the tooth adjacent to the missing tooth is... The central position is The diameter of the right adjacent tooth is The central position is Calculate the center coordinates of the missing tooth in the 2D image. The formula is as follows:

[0117]

[0118] It can be seen that the calculation method of the present invention is in When the adjacent tooth on the left is larger, that is... The center point is closer to the right adjacent tooth, while... When this happens, the formula degenerates into a simple average. Using this calculation method, the position of the missing tooth on the dental arch curve is approximately the midpoint between the positions of the left and right adjacent teeth, which matches the actual situation.

[0119] S44. Calculate the location of missing teeth in the 3D model, specifically including: Suppose a 3D STL model is in The maximum and minimum boundaries on the plane are If the two-dimensional projected image Width in the axial direction is , The height in the axial direction is Then calculate and Directional scaling factor and The formula is as follows:

[0120]

[0121] Then the coordinates of the two-dimensional image To three-dimensional coordinates The linear mapping relationship is as follows: This is because in two-dimensional images Mapped to a 3D image When both are in the same direction, increasing to the right, they can be directly scaled using the scaling factor. However, in a two-dimensional image... Mapped to a 3D image When the coordinate axes are expressed slightly differently, in two dimensions... Points with smaller values ​​are located at the top, while in three dimensions... Points with larger values ​​are located at the top, so a negative sign is required during mapping.

[0122] Define with A 3D tooth model with vertices ,as follows:

[0123] in, ; Using two-dimensional image coordinates To three-dimensional coordinates linear mapping relationship Calculate a three-dimensional tooth model Coordinates of the center of the missing middle tooth The formula is as follows:

[0124]

[0125]

[0126] in, express Scale factor in the axial direction express Scale factor in the axial direction.

[0127] Because the generated image does not have blank borders in the top, bottom, left, and right directions when generating a 3D to 2D projection map, all pixels correspond to the 3D model. Data points on a plane, therefore, can be used to calculate the 3D STL model. coordinates and The maximum and minimum values ​​of the coordinates are used to obtain the 3D STL model. Width and height of the plane Then the width and height of the two-dimensional image are calculated. With 3D model Length and width on a plane scaling factor ,like Figure 15 As shown. Thus, for any coordinate point in a two-dimensional image, it can be determined according to the scaling factor. The coordinates of the missing tooth's center point on the 3D STL model are mapped to coordinates on the model. In this embodiment, the method of the present invention makes the center point closer to the right neighboring tooth when the left neighboring tooth is larger, and degenerates into a simple average when the left and right neighboring teeth are equal, so that the position of the missing tooth on the dental arch curve is approximately the midpoint of the positions of the left and right neighboring teeth, which conforms to the actual anatomical structure.

[0128] In a specific implementation, as a preferred embodiment of the present invention, step S5 includes: S51. Enter the missing tooth number. Coordinates of the center of the missing tooth location in the 3D tooth model dental implant template data set , ; S52, Regarding the dental implant template data set Calculate the centroid coordinates of the template tooth. The formula is as follows:

[0129] in, This indicates the number of triangular faces in the template tooth model; S53, the center point of the obtained dental implant template. Move to the center coordinates of the missing tooth location in the 3D tooth model Calculate the displacement of the center point of the dental implant template tooth. The calculation formula is as follows:

[0130] S54, Center point of the dental implant template With displacement Add them together to get ; S55, The points in the original dental implant template... move ,get The new coordinates of all points in the implant template tooth are determined using the following formula:

[0131] S56. After the above movements, a three-dimensional crown STL model data that matches the position of the missing tooth is obtained. The formula is as follows:

[0132] in, , indicating the translated first A triangular facet, Indicates the translation of the first... The normal vectors of the triangular faces. , representing the coordinates of the new vertex after translation. This represents the three-dimensional coordinates of the vertex after translation. For example... Figure 17 As shown, after determining the specific location, the template tooth can be moved to the center of the missing tooth, thus completing the initial repositioning of the template tooth. The final output is a 3D STL model of the crown that matches the missing tooth location. .

[0133] Example 1: Estimating the center point of the edentulous region Step 1: Identify missing teeth Input: Semantic segmentation results of a two-dimensional projected image.

[0134] Parameter settings: Minimum patch area threshold Pixel.

[0135] Processing procedure: Extract the largest 8-connected region for each tooth category and identify the missing category.

[0136] Results: Set of missing tooth labels This indicates that tooth number 35 is missing.

[0137] Step 2: Calculate the center point of the tooth Missing tooth label Left and right adjacent tooth numbers , .

[0138] Extract the maximum connected region of tooth #36 and the number of pixels. Center point coordinates ; Extract the maximum connected region of tooth #34 and the number of pixels. Center point coordinates .

[0139] Step 3: Calculate the center point of the missing tooth Calculate the approximate diameter of the left and right adjacent teeth (assuming the teeth are approximately circular). and The formulas are as follows:

[0140]

[0141] Weighted calculation of the center point coordinates of the missing tooth region ( As shown in the following formula:

[0142] Step 4: Mapping 3D coordinates 3D STL model in The boundary of the plane is as follows:

[0143] in , These represent three-dimensional tooth models. Minimum and maximum values ​​of the axis coordinates and These represent three-dimensional tooth models. Minimum and maximum values ​​of axis coordinates.

[0144] Set the size of the 2D projected image to the width. ,high ;calculate On the axis and Scale factor on axis and The result is as follows:

[0145] Mapping yields the coordinates of the three-dimensional missing center point ( The calculation result is as follows:

[0146] In the above formula, regarding The calculation, in , Take within range Maximum value, for example .

[0147] Example 2: Positioning template teeth Step 1: Determine the template tooth data Dental implant template using tooth size 35 Its definition is as follows, and the specific meaning of its internal variables is the same as above. The template tooth contains the number of triangular faces. .

[0148]

[0149] Calculate the centroid of the template tooth (the template has been centered) to obtain ( = (0.0, 0.0, 0.0).

[0150] Step 2, Translation Transformation The translation vector (from the template center to the missing center) is calculated as follows:

[0151] Apply translation to all vertices of the template tooth, and calculate as follows:

[0152] Step 3, Output Results A three-dimensional STL model of the crown that matches the location of the missing tooth is generated. This model can be directly used for subsequent fine morphological adjustments and implant surgery planning.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net, characterized in that, include: S1. Based on the STL format, a 3D tooth model is generated by calculating and normalizing the normal vector of each triangular facet so that each triangular facet corresponds to a pixel. The 3D normal vector information is mapped to the color space to generate a high-resolution scatter plot, thereby generating a 2D projection map that reflects the normal vector characteristics of the 3D model. S2. Using VGG16 as the backbone network, a U-Net network structure is combined with a strip pooling module to design a neural network model that accurately captures the sequential arrangement features of teeth. S3. An adaptive learning rate adjustment strategy and a cosine annealing learning rate scheduling strategy are used to train the neural network model, and the trained neural network model is used to perform semantic segmentation on the two-dimensional projection image to identify each tooth category and obtain a 2D tooth classification result image. S4. Based on the 2D tooth classification result image, identify the missing tooth label, calculate the center point of the left and right adjacent teeth of the missing tooth, calculate the center position coordinate of the missing tooth in the 2D image through the center point of the left and right adjacent teeth, and map it to the three-dimensional space to obtain the three-dimensional center coordinate of the missing tooth. S5. Move the center point of the template tooth to the three-dimensional center coordinates of the missing tooth to generate a three-dimensional crown model that matches the position of the missing tooth.

2. The method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net according to claim 1, characterized in that, Step S1 includes: S11. Considering that the 3D STL format tooth model to be detected is a collection of triangular facets, each with three vertices and a normal vector, therefore, it will have... The 3D model data of each triangular facet is represented as a set. : in, Indicates the first A triangular facet, , Indicates the first The three vertices of a triangle. Indicates the first The normal vectors of the triangular faces. Depend on Three components Normal vector 3D coordinate vector; Indicates the first The first triangular face One vertex, , Represents vertices The three-dimensional coordinate components; S12, Calculate each triangular face The center of gravity is determined by the following formula: in, Indicates the center of gravity The three-dimensional coordinate components, , , ; S13, Regarding the first Normal vectors of each face The three components are used to generate the normal vectors of all triangle faces. coordinate, coordinates and The set of minimum and maximum coordinate values ​​is given by the following formula: in, Represents the normal vector of all triangle faces. The set of minimum values ​​of each component. Represents the normal vector of all triangle faces. The set of maximum values ​​of each component. ,correspond , Indicates the first The first normal vector of the triangular face One component; S14. Calculate the normal vector components of all triangle faces. Normalize to the interval [0, 1] to generate the normal vector, as follows: in, This represents the normalized normal vector. S15. Based on the center of gravity coordinates Sort all triangle faces in ascending order to ensure Points with smaller values ​​are projected first. Project the points with larger values ​​back; S16. Define a two-dimensional projected image as a collection The set of points is given by the following formula: Each point It is a two-dimensional data. ; S17, Regarding according to Each triangle face after value sorting The calculated center points of the triangular faces are then arranged in sequence. Mapped to A plane is used to generate a two-dimensional projection point. To obtain the projection point coordinates and Coordinates are as follows: Finally, 3D model data is generated. Two-dimensional projection of the normal vector of the plane.

3. The method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net according to claim 1, characterized in that, Step S2 includes: S21. Construct a U-Net network with VGG16 as the encoder. The encoder adopts a 5-layer downsampling structure, including 13 convolutional layers and 5 max pooling layers; the decoder adopts a 4-layer upsampling structure. S22. Add a strip pooling module to each downsampling structure of the encoder to collect long-distance contextual information from different spatial dimensions using horizontal and vertical strip pooling operations.

4. The method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net according to claim 3, characterized in that, Step S22 includes: S221, Let For a two-dimensional input tensor, where and These represent the spatial height and width of the two-dimensional input tensor, respectively, and the bar pooling module uses... The pooling window performs a horizontal bar pooling operation and outputs... The formula is as follows: in, This indicates the first horizontal bar pooling step. The height feature value of the row; S222, Let For a two-dimensional input tensor, where and These represent the spatial height and width of the two-dimensional input tensor, respectively, and the bar pooling module uses... The pooling window performs a vertical bar pooling operation and outputs... The formula is as follows: in, This indicates the first vertical bar pooling step. The width characteristic value of the column; S223, Let For the input tensor, where Indicates the number of channels. and The height and width of the input tensor space are respectively... The input is fed into two parallel paths, each containing a horizontal or vertical strip pooling layer followed by a one-dimensional convolutional layer with a kernel size of 3 to modulate the current position and its neighboring features. The output is... and ; S224, Output and By performing fusion, fusion characteristics are obtained. The fusion formula is as follows: in, This indicates that the fused feature tensor is in the channel. ,Location eigenvalues ​​at that location This indicates that the horizontal pooling output is in the channel. , No. The eigenvalues ​​of the row Indicates the vertical pooling output in the channel , No. The characteristic values ​​of the column; S225, Based on fusion features The calculation yields an output that contains more global prior information. The formula is as follows: in, This represents element-wise multiplication. This represents the Sigmoid function. Represents a 1×1 convolution; S226. Use the Sigmoid function to map the convolution output to... The interval, expressed as follows: in, Indicates the input value. The output range is in Between these, the generated spatial attention weights are used to reweight the input features.

5. The method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net according to claim 1, characterized in that, Step S3 includes: S31. An adaptive learning rate adjustment strategy is adopted to normalize the actual batch size with the baseline batch size, dynamically adjusting the initial learning rate to ensure that the neural network model obtains consistent convergence characteristics under different training configurations. The specific process is as follows: The adjusted initial learning rate is calculated using the following formula: in, This indicates the actual batch size during the current training phase. Indicates the baseline batch size. This represents the user-defined initial learning rate. This represents the maximum learning rate limit associated with the optimizer. This represents the minimum learning rate lower bound related to the optimizer. This represents the adjusted actual initial learning rate. This represents the adjusted actual minimum learning rate. The function will input values Limited to the range Inside, ; Calculate the adjusted minimum learning rate The calculation formula is as follows: S32. A cosine annealing learning rate scheduling strategy is adopted to simulate the changing law of the cosine function, so that the learning rate smoothly decays from the initial value to the minimum value, avoiding training oscillations caused by step-like descent, and promoting the neural network model to converge to the optimal solution more stably. The specific process is as follows: Calculate the number of preheating stages The formula is as follows: in, Indicates the total number of training rounds; Calculate the number of rounds in the smooth termination phase The formula is as follows: Based on the adjusted actual initial learning rate Calculate the warm-up initial learning rate The formula is as follows: Based on the number of preheating stages The number of rounds in the smooth conclusion phase and warm-up initial learning rate The learning rate change curve is described using a piecewise function, as follows: in, This indicates the current training round number.

6. The method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net according to claim 1, characterized in that, Step S4 includes: S41. Based on the classification results of a single tooth in the two-dimensional projection image, 8-connected regions are extracted from the classification regions of different tooth labels. The extracted connected regions that satisfy the minimum patch area threshold minPatchSize are statistically analyzed. If a tooth category is missing, that tooth category label is the missing tooth label. Specifically, this includes: Define tooth category space The numbering of all teeth is denoted as... ; For tooth category space any of the following ,calculate The maximum number of pixels in the 8-connected region corresponding to the class ,like Then Add to collection middle; Search Sets There is a lack of space for tooth categories in China. Which category label in, i.e. ,but That is, a set consisting of categories of missing teeth; S42. Calculate the center points of the left and right adjacent teeth of the missing tooth, specifically including: Assume the missing tooth is numbered The tooth numbers of its left and right adjacent teeth are and ,like ,but , Otherwise, if the tooth is an incisor or the last molar, a non-molar area will be indicated, and no action will be taken. In the two-dimensional image of tooth classification results, the teeth labeled as follows are extracted respectively. and For the tooth region sub-image, the 8-connected region extraction method was used to count all connected regions in the left and right adjacent tooth region sub-images. Considering the possibility of misclassification, the connected region of the left and right adjacent teeth with the largest patch area was regarded as the actual range of the left and right adjacent teeth, while other smaller connected regions were regarded as noise removal. set up It is the set of pixels within all connected regions of a given neighboring tooth, containing... If there are 100 pixels, then define The average value of each pixel is the center of the adjacent tooth. The formula is as follows: S43. The coordinates of the center position of the missing tooth in the 2D image are calculated using the center points of the left and right adjacent teeth, specifically including: Let the area of ​​a neighboring tooth in the missing region be... It is the sum of the number of pixels in the connected region corresponding to the adjacent tooth. That is ;area The calculation formula is as follows: Since molars are generally circular in shape, we assume that the shape of this region is approximately circular, and then based on the area... The approximate diameter of the tooth was calculated. The formula is as follows: Assuming the diameter of the tooth adjacent to the missing tooth is... The central position is The diameter of the right adjacent tooth is The central position is Calculate the center coordinates of the missing tooth in the 2D image. The formula is as follows: S44. Calculate the location of missing teeth in the 3D model, specifically including: Suppose a 3D STL model is in The maximum and minimum boundaries on the plane are If the two-dimensional projected image Width in the axial direction is , The height in the axial direction is Then calculate and Directional scaling factor and The formula is as follows: Then the coordinates of the two-dimensional image To three-dimensional coordinates The linear mapping relationship is as follows: ; Define with A 3D tooth model with vertices ,as follows: in, ; Using two-dimensional image coordinates To three-dimensional coordinates linear mapping relationship Calculate a three-dimensional tooth model Coordinates of the center of the missing middle tooth The formula is as follows: in, express Scale factor in the axial direction express Scale factor in the axial direction.

7. The method for intelligent selection and generation of three-dimensional dental implant crowns based on U-Net according to claim 1, characterized in that, Step S5 includes: S51. Enter the missing tooth number. Coordinates of the center of the missing tooth location in the 3D tooth model dental implant template data set , ; S52, Regarding the dental implant template data set Calculate the centroid coordinates of the template tooth. The formula is as follows: in, This indicates the number of triangular faces in the template tooth model; S53, the center point of the obtained dental implant template. Move to the center coordinates of the missing tooth location in the 3D tooth model Calculate the displacement of the center point of the dental implant template tooth. The calculation formula is as follows: S54, Center point of the dental implant template With displacement Add them together to get ; S55, The points in the original dental implant template... move ,get The new coordinates of all points in the implant template tooth are determined using the following formula: S56. After the above movements, a three-dimensional crown STL model data that matches the position of the missing tooth is obtained. The formula is as follows: in, , indicating the translated first A triangular facet, Indicates the translation of the first... The normal vectors of the triangular faces. , representing the coordinates of the new vertex after translation. This represents the three-dimensional coordinates of the vertex after translation.

Images