An oral implant planning method, device, equipment, medium and program product

Abstract

The present disclosure relates to an oral implant planning method, device, equipment, medium and program product, the method comprising: acquiring an oral three-dimensional image of a to-be-tested sample, an implant prediction model composed of a backbone prediction network and a linear modulation network which has been trained; segmenting and positioning the oral three-dimensional image to obtain a tooth loss feature map and a tooth loss tooth position identifier; inputting the tooth loss tooth position identifier into the linear modulation network to obtain output modulation parameter data; inputting the tooth loss feature map into the backbone prediction network, and performing affine transformation on the intermediate feature map of the backbone prediction network according to the modulation parameter data to obtain target implant parameters output by the backbone prediction network; and determining a target implant model corresponding to the to-be-tested sample according to the target implant parameters, thereby solving the problem of insufficient cross-tooth position generalization performance caused by oral full-tooth position unified modeling, reducing the prediction deviation caused by anatomical differences of different tooth positions, and improving the reliability of implant selection planning.

Description

Technical Field

[0001] This disclosure relates to the field of oral medical image processing technology, and in particular to an oral implant planning method, device, equipment, medium and program product. Background Technology

[0002] In dental implantology, precise implant selection is crucial for determining long-term survival and functional stability. Currently, clinical practice primarily relies on cone-beam computed tomography (CBCT) to reconstruct the three-dimensional anatomical structure of the jawbone. This, combined with anatomical landmarks such as the root morphology of adjacent teeth at the edentulous site and the integrity of the labial, buccal, and lingual bone plates, allows for the selection of the appropriate implant model in terms of diameter and length.

[0003] Significant heterogeneity exists in the anatomical structure and mechanical characteristics of different edentulous sites within the oral cavity. Taking the molar and premolar regions as examples, the molar region has sufficient bone volume, but its vertical bone height is limited due to its proximity to the inferior alveolar nerve, and it is subjected to multi-directional shear stress over a long period of time; the premolar region has moderate bone density and relatively abundant horizontal bone volume, but it has high requirements for the aesthetic effect and vertical occlusal precision of implant restorations.

[0004] In recent years, deep learning-based prediction models have provided a new solution for implant selection planning. However, most existing models use a global unified framework to model all tooth positions in the oral cavity, which makes it difficult to adapt to the differentiated needs of implant selection for different tooth positions. The generalization performance across tooth positions is insufficient, which has caused implant selection planning to fall into an accuracy bottleneck. Summary of the Invention

[0005] This disclosure provides a method, apparatus, device, medium, and program product for dental implant planning, which addresses the problem of insufficient generalization performance across tooth positions caused by unified modeling of all tooth positions, improves the accuracy of the prediction model in selecting implants for different tooth positions, and ensures the reliability of implant selection planning.

[0006] One aspect of this disclosure provides a method for dental implant planning, including: Acquire three-dimensional images of the oral cavity of the sample to be tested and a trained implant prediction model, wherein the implant prediction model consists of a backbone prediction network and a linear modulation network; The oral cavity three-dimensional image is segmented and localized to obtain a tooth loss feature atlas and tooth loss position markers; The missing tooth position identifier is input into the linear modulation network to obtain the output modulation parameter data, which includes feature modulation parameters corresponding to at least one target network layer in the backbone prediction network. The missing tooth feature map is input into the trunk prediction network, and the intermediate feature map output by the corresponding target network layer is subjected to affine transformation according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network. Based on the target implant parameters, determine the target implant model corresponding to the sample to be tested.

[0007] Another aspect of this disclosure provides a dental implant planning device, comprising: The implant prediction model acquisition module is used to acquire the oral three-dimensional image of the sample to be tested and the trained implant prediction model, which consists of a backbone prediction network and a linear modulation network. The missing tooth position identification module is used to segment and locate the oral cavity three-dimensional image to obtain a missing tooth feature map and missing tooth position identification; The modulation parameter data determination module is used to input the missing tooth position identifier into the linear modulation network to obtain the output modulation parameter data, wherein the modulation parameter data includes feature modulation parameters corresponding to at least one target network layer in the backbone prediction network. The target implant parameter output module is used to input the missing tooth feature map into the trunk prediction network, and perform affine transformation on the intermediate feature map output by the corresponding target network layer according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network. The target implant model determination module is used to determine the target implant model corresponding to the sample to be tested based on the target implant parameters.

[0008] Another aspect of this disclosure provides an electronic device, comprising: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, which is then executed by the at least one processor to enable the at least one processor to perform the oral implant planning method according to any embodiment of this disclosure.

[0009] Another aspect of this disclosure provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the oral implant planning method according to any embodiment of this disclosure.

[0010] Another aspect of this disclosure provides a computer program product including a computer program that, when executed by a processor, implements the oral implant planning method according to any embodiment of this disclosure.

[0011] The technical solution of this disclosure constructs a dual-pathway network framework driven by a backbone prediction network and a linear modulation network. This framework segments and locates three-dimensional images of the oral cavity to obtain edentulous feature maps and edentulous tooth position identifiers. Based on these edentulous tooth position identifiers, the linear modulation network introduces feature modulation parameters into the backbone prediction network. This performs affine transformation on the target network layer in the backbone prediction network, enabling the backbone prediction network to have geometric perception capabilities of edentulous sites. The network then outputs customized implant parameters adapted to different tooth positions, ultimately achieving implant selection and matching. This solves the problem of insufficient generalization performance across tooth positions caused by unified modeling of all tooth positions in the oral cavity, reduces prediction bias caused by anatomical differences in different tooth positions, and improves the reliability and accuracy of implant selection planning.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating an oral implant planning method provided in one embodiment of this disclosure; Figure 2 This is a schematic diagram illustrating the construction of a tooth loss feature map according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram illustrating the modulation principle of a linear modulation module provided in one embodiment of the present disclosure; Figure 4 A flowchart illustrating a specific example of an oral implant planning method provided in one embodiment of this disclosure; Figure 5 A flowchart illustrating another oral implant planning method provided in one embodiment of this disclosure; Figure 6 A diagram illustrating the effect of a regularization penalty term provided in an embodiment of this disclosure; Figure 7 This is a schematic diagram of the structure of an oral implant planning device provided in one embodiment of the present disclosure; Figure 8 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present disclosure. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0016] It should be noted that the terms "to be tested," "training," "target," "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises 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 processes, methods, products, or apparatus.

[0017] Figure 1 This is a flowchart illustrating an embodiment of an oral implant planning method provided by this disclosure. This embodiment is applicable to planning the selection of implants for edentulous sites in the oral cavity. The method can be executed by an oral implant planning device, which can be implemented in hardware and / or software and can be configured in a terminal device. Figure 1 As shown, the method includes: S110. Obtain the three-dimensional oral image of the sample to be tested and the trained implant prediction model.

[0018] Specifically, the oral cavity three-dimensional image includes oral anatomical information such as alveolar bone, tooth roots, and nerve canals, and the oral cavity three-dimensional image is acquired by CBCT equipment.

[0019] For example, due to the anisotropic spatial resolution of CBCT equipment, the method further includes: isotropically resampling the original CBCT images acquired by the CBCT equipment to unify the voxel spacing in different directions to a preset value, such as 0.4 mm, to obtain a three-dimensional image of the oral cavity. For example, resampling can be achieved using trilinear interpolation to preserve the continuity and geometric fidelity of the anatomical structure.

[0020] In this embodiment, the implant prediction model is used to predict implant parameters at the edentulous site of the test sample. The implant prediction model consists of a trunk prediction network and a linear modulation network.

[0021] S120. The oral cavity three-dimensional image is segmented and located to obtain a tooth loss feature map and tooth loss position markers.

[0022] Specifically, the segmentation process refers to semantic segmentation of teeth in three-dimensional oral images, accurately identifying the anatomical boundaries and tooth position markers corresponding to each tooth location; the localization process refers to identifying the missing tooth position markers from each tooth position marker, and determining the missing tooth feature map that matches the missing tooth location.

[0023] The missing tooth position identifier is used to uniquely identify the location of the missing tooth site in the oral cavity. For example, the missing tooth position identifier can be coded using a two-digit numbering system provided by the International Dental Federation (FDI), or it can be represented by a spatial vector in the three-dimensional coordinate system of the oral cavity, or it can be other clinically common coding methods, but it is not limited to the examples given above.

[0024] The missing tooth feature map is used to characterize the set of key features within the region of interest of the missing tooth. In an optional embodiment, the missing tooth feature map includes a region of interest mask cropped from the tooth segmentation results, and / or a region of interest image cropped from the three-dimensional oral image.

[0025] Here, the Region of Interest (ROI) refers to a three-dimensional spatial region centered on the edentulous site, defined by a preset shape and size. For example, the preset shape includes, but is not limited to, a sphere, an ellipsoid, or a cube, and the preset size can be predefined or adaptively adjusted based on the tooth size or distribution in the tooth segmentation results.

[0026] Based on the above embodiments, optionally, the missing tooth feature map further includes a distance transformation map, which represents the spatial distance distribution from voxels in the region of interest mask to the nearest tooth voxel.

[0027] In this embodiment, the segmentation and localization of the oral cavity 3D image to obtain a missing tooth feature atlas and missing tooth position markers includes: inputting the oral cavity 3D image into a trained image segmentation model to obtain missing tooth position markers and tooth segmentation results; performing region localization based on the tooth segmentation results to obtain a region of interest (ROI) for the missing tooth, and cropping the tooth segmentation results and the oral cavity 3D image based on the ROI to obtain a ROI mask and a ROI image, respectively; performing distance transformation based on the ROI mask to obtain a distance transformation map, the distance transformation map representing the spatial distance distribution from voxels in the ROI mask to the nearest tooth voxel; and constructing a missing tooth feature atlas based on the ROI mask, the ROI image, and the distance transformation map.

[0028] For example, the network architecture used in the image segmentation model includes, but is not limited to, U-shaped networks, enhanced nested U-shaped networks, full-scale fusion U-shaped networks, or attention mechanism U-shaped networks.

[0029] In an optional embodiment, the image segmentation model employs a globally context-aware U-shaped convolutional network architecture, where the bottleneck layer integrates a selective state-space model (SSM). The core principle of the selective SSM model is a continuous-time state-space equation, expressed through high-dimensional hidden states. For a one-dimensional input sequence Complete global feature mapping and output sequence. Its continuous-time dynamics are constrained by linear ordinary differential equations, the specific formulas of which are as follows: in, This represents the high-dimensional hidden state corresponding to the current time. This represents the high-dimensional hidden state corresponding to the next time step. This represents the state evolution matrix, used to control the temporal iterative update pattern of the hidden states. and These represent the feature projection parameters, which are responsible for the dimension mapping of the input features and the output mapping of the latent state features, respectively.

[0030] Since continuous-time state-space equations belong to the continuous domain, they cannot be directly adapted to the discrete training and inference logic of deep learning. Therefore, discretization of continuous-time state-space equations is necessary. This disclosure introduces a time scale parameter based on the zero-order preservation principle. The parameter matrix of the continuous domain Transform into a parameter matrix that can be iteratively computed in the discrete domain. The discretized time series recursive formula is expressed as: in, , ,in, Represents the identity matrix. This indicates matrix exponentiation operations.

[0031] Unlike the traditional fixed-parameter SSM model, the selective SSM model... It is not a constant value and can be dynamically and adaptively updated based on the feature information of the input data. This dynamic update mechanism enables the model to adaptively adjust the feature attention at different positions in the sequence according to the input content during the inference process, while maintaining the overall computational complexity at the linear level of the sequence length. ( (Indicates the length of the input sequence). For example, the selective SSM model can be efficiently computed using a parallel correlation scan algorithm, balancing global modeling capability and computational efficiency.

[0032] Specifically, in the image segmentation model, the local feature maps extracted by the encoder are flattened into a one-dimensional long sequence along the spatial dimension. This sequence is then input into the selective SSM model to complete global feature fusion and long-distance dependency modeling. The output one-dimensional long sequence is then reconstructed back into a three-dimensional feature map structure and input into the decoder to complete feature restoration and image segmentation prediction. The parameter matrix of the selective SSM model... It relies on dynamic calculation of local feature maps, thereby adaptively capturing implicit features such as global morphological associations and tissue topological dependencies of the entire dental arch structure.

[0033] Existing technologies for segmenting high-resolution 3D oral images generally face a technical dilemma of balancing performance and efficiency. Traditional convolutional neural networks, limited by their finite receptive field, cannot effectively capture global anatomical contextual features across the dental arch and struggle to fully utilize the overall oral cavity topology. This results in insufficient segmentation accuracy and blurred boundary fitting in key areas such as edentulous sites and alveolar bone boundaries. While the visual Transformer architecture based on self-attention mechanisms possesses powerful global dependency modeling capabilities, its computational complexity increases quadratically with input resolution. When applied to 3D oral volumetric data processing, it generates significant memory consumption and computational power consumption, resulting in a severe computational bottleneck.

[0034] This embodiment integrates the selective SSM model into the bottleneck layer of the U-shaped convolutional network architecture to construct a novel hybrid architecture of selective SSM-U-shaped convolution: while maintaining the strong spatial localization capability of the U-shaped convolutional network, it utilizes the linear complexity advantage of the selective SSM model in modeling long-range dependencies to efficiently capture cross-regional anatomical constraints such as dentition continuity, dental arch symmetry, and bone-tooth interface, thus balancing global context modeling capability and inference efficiency.

[0035] Specifically, the tooth segmentation result is a voxel-level segmentation mask of the tooth in the three-dimensional image of the oral cavity. The region of interest for the missing tooth covers the missing tooth site and the surrounding tooth sites. The region of interest mask contains the binarization features within the region of interest for the missing tooth, and the region of interest image contains the intensity features within the region of interest for the missing tooth.

[0036] In this embodiment, the missing tooth feature map is a three-channel composite input tensor. Its construction method is as follows: in, This indicates a concatenation operation along the channel dimension. Indicates the region of interest in the image. Indicates the region of interest mask. This represents a distance transformation graph.

[0037] For example, suppose For a 3D image domain, for any voxel within the mask of the region of interest The distance transformation value is defined as the minimum Euclidean distance from the voxel to the nearest tooth voxel: in, The set of all tooth voxels within the region of interest mask. express Norm.

[0038] Specifically, the distance transformation map quantifies the tooth adjacency relationship of each voxel in the region of interest of the edentulous space. The smaller the distance, the closer the voxel is to the root of the adjacent tooth, and the higher the risk of implanting an implant at the voxel invading the root of the adjacent tooth. On the other hand, the larger the distance, the more likely the voxel is located in the central region of the edentulous space, with higher implantation safety.

[0039] Figure 2 This is a schematic diagram illustrating the construction of a tooth loss feature map according to an embodiment of the present disclosure. Specifically, Figure 2 Sub-image (A) represents a cross-sectional image of the region of interest for missing teeth obtained from oral CBCT images. Its pixel value represents the intensity value, and a red dashed box marks a missing tooth site. Sub-image (B) represents a two-dimensional segmentation mask of the region of interest for missing teeth obtained from tooth segmentation results. Its pixel value represents whether the pixel belongs to the tooth category. Sub-image (C) represents a distance transformation map calculated based on sub-image (B). Its pixel value represents the Euclidean distance from the pixel to the nearest tooth voxel. Sub-image (D) represents a missing tooth feature map obtained by stitching sub-images (A), (B), and (C) based on the channel dimension.

[0040] This embodiment introduces distance transformation into the tooth loss feature atlas, explicitly encoding anatomical prior knowledge into differentiable geometric metric signals. This provides the implant prediction model with explicit prior information on geometric safety distances, enabling the implant prediction model to directly learn risk representations that are highly relevant to implant decisions. This strengthens the end-to-end model's ability to discriminate implant safety boundaries and improves the safety of implant selection planning.

[0041] S130. Input the missing tooth position identifier into the linear modulation network to obtain the output modulation parameter data.

[0042] In this embodiment, the modulation parameter data includes feature modulation parameters corresponding to at least one target network layer in the backbone prediction network.

[0043] Specifically, the backbone prediction network includes an output layer and at least one preset network layer. Each preset network layer is used to extract feature information from the missing tooth feature map to obtain an intermediate feature map. The output layer is used to output target implant parameters based on the intermediate feature map. The target network layer is a pre-selected preset network layer in the backbone prediction network that has significant anatomical discriminative power.

[0044] Specifically, the linear modulation network includes linear modulation modules corresponding one-to-one with the target network layers. These linear modulation modules provide feature modulation parameters for the target network layers in the backbone prediction network. The feature modulation parameters include scaling parameters and offset parameters. The scaling parameters are used to adjust the channel response intensity of the intermediate feature maps, and the offset parameters are used to apply a bias to the channel reference values ​​of the intermediate feature maps. Together, they achieve channel-by-channel feature affine recalibration.

[0045] For example, the linear modulation module may be a lightweight convolution module, a depthwise separable convolution module, a pointwise convolution module, a gated feature mapping module, or a combination thereof, but is not limited to the given example.

[0046] In an optional embodiment, the linear modulation module includes an embedding layer and a feature transformation layer; wherein the embedding layer is used to map the missing tooth position identifier into a high-dimensional embedding vector; and the feature transformation layer is used to map the high-dimensional embedding vector into feature modulation parameters that match the corresponding target network layer.

[0047] In an optional embodiment, the feature transformation layer includes a linear projection layer, a nonlinear activation function, and a channel normalization module connected in sequence; wherein, the linear projection layer maps the high-dimensional embedding vector to a dimension consistent with the number of channels in the target network layer, as initial modulation parameters; the nonlinear activation function is used to enhance the expressive power of the initial modulation parameters and better model the nonlinear distribution differences of different tooth sites; the channel normalization module is used to constrain the modulation parameters output by the nonlinear activation function and output feature modulation parameters that match the corresponding target network layer.

[0048] In another alternative embodiment, the feature transformation layer is a multilayer perceptron.

[0049] S140. Input the missing tooth feature map into the trunk prediction network, and perform an affine transformation on the intermediate feature map output by the corresponding target network layer according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network.

[0050] For example, the network architecture of the backbone prediction network can be a regression prediction network, a feature regression network, or a dilated convolutional network, but it is not limited to the given example.

[0051] In an optional embodiment, the backbone prediction network employs a selective state-space model architecture. Specifically, the selective SSM model used by the backbone prediction network is consistent with the selective SSM model integrated into the bottleneck layer in the aforementioned image segmentation model, i.e., both are based on a selective scanning mechanism to model long-range anatomical dependencies.

[0052] For example, the first The first target network layer The affine transformation corresponding to each channel satisfies the following formula: in, Indicates the first The first target network layer Modulation feature map corresponding to each channel Indicates the first The first target network layer Intermediate feature maps corresponding to each channel Indicates the first The first target network layer Scaling parameters corresponding to each channel Indicates the first The first target network layer The offset parameters corresponding to each channel This indicates the location of a missing tooth.

[0053] Figure 3This is a schematic diagram illustrating the modulation principle of a linear modulation module according to an embodiment of this disclosure. Specifically, the FDI number of the missing tooth site is input into the embedding layer of the linear modulation module, and the high-dimensional embedding vector output from the embedding layer is input into a multilayer perceptron to obtain the output scaling parameters and offset parameters. The intermediate feature map is scaled channel-by-channel according to the scaling parameters to obtain a scaled feature map, and the scaled feature map is offset channel-by-channel according to the offset parameters to obtain a modulated feature map. This represents a multiplication operation based on the channel dimension. This represents the summation operation based on the channel dimension.

[0054] S150. Based on the target implant parameters, determine the target implant model corresponding to the sample to be tested.

[0055] Specifically, the target implant parameters characterize the implant's specifications. In one optional embodiment, the target implant parameters include the implant's diameter and length. The diameter represents the maximum outer diameter of the implant's cross-section, and the length represents the distance from the implant platform reference surface to the implant's root tip along the implant's long axis.

[0056] Figure 4 This is a flowchart illustrating a specific example of an oral implant planning method provided in one embodiment of the present disclosure. The example includes two stages. In stage 1, voxel segmentation and cropping based on a selective SSM model, oral CBCT images are input into a selectively SSM-enhanced U-shaped convolutional model for image segmentation, obtaining tooth segmentation results and missing tooth location markers. Figure 4 The missing tooth position shown is identified as FDI number 35. Based on the tooth segmentation results and the region of interest defined by the missing tooth position, a missing tooth feature map containing CBCT intensity features, binary mask features, and distance transformation features is constructed.

[0057] In Phase 2: Geometric Perception-Based Implant Parameter Prediction, the missing tooth feature map is input into the selective SSM model, and the missing tooth position identifier is input into the linear modulation module. The output feature modulation parameters are sequentially processed through the embedding layer and multilayer perceptron to adaptively modulate the intermediate feature map in the selective SSM model, enhancing the model's ability to perceive the specific geometric features of the missing tooth region. Finally, the diameter and length of the implant are output. Based on the implant's diameter and length, the closest implant model is matched from the implant model library.

[0058] Specifically, the target implant model is obtained by searching a predefined implant model library using a lookup table or differentiable interpolation method based on the target implant parameters. The implant model library represents a set of key-value pairs with the preset implant model as the key and the implant parameters as the values. For example, the search method may employ nearest neighbor matching, Euclidean distance matching, search threshold interval matching, or K-nearest neighbor classification matching, but is not limited to the given examples.

[0059] Optionally, based on the above embodiments, the method further includes: automating the configuration and planning of the implant spatial posture according to the target implant model using an oral intelligent agent model. The decision knowledge base of the oral intelligent agent model includes clinical guidelines from the International Society for Implantology and / or a professional implant knowledge base built based on the "3A2B golden rule".

[0060] Among them, the International Society for Implant Dentistry's Clinical Guidelines are based on massive global clinical data and industry consensus. They are recognized as standardized and regulated diagnostic and treatment guidelines in the field of dental implantology, providing clinical principles or references for implantation planning in core dimensions such as implantation location, angle, depth, bone condition matching, and surgical risk avoidance. The "3A2B Golden Principle" can be used to reduce related implantation risks.

[0061] The technical solution of this embodiment constructs a dual-pathway network framework driven by a backbone prediction network and a linear modulation network. This framework segments and locates 3D images of the oral cavity to obtain edentulous feature maps and edentulous tooth position identifiers. Based on these edentulous tooth position identifiers, the linear modulation network introduces feature modulation parameters into the backbone prediction network. This performs affine transformation on the target network layer in the backbone prediction network, enabling the backbone prediction network to have geometric perception capabilities of edentulous sites. The network then outputs customized implant parameters adapted to different tooth positions, ultimately achieving implant selection and matching. This solution addresses the problem of insufficient generalization performance across tooth positions caused by unified modeling of all tooth positions in the oral cavity, reduces prediction bias caused by anatomical differences in different tooth positions, and improves the reliability and accuracy of implant selection planning.

[0062] Figure 5 This is a flowchart illustrating another oral implant planning method provided in one embodiment of this disclosure. This embodiment further refines the training method of the "implant prediction model" in the above embodiment. For example... Figure 5 As shown, the method includes: S210. Obtain the training dataset.

[0063] In this embodiment, the training dataset contains training missing tooth data corresponding to multiple training samples. The training missing tooth data includes a missing tooth feature map and a missing tooth position identifier. Specifically, the method for determining the training missing tooth data of the training samples is the same as or similar to the method for determining the missing tooth feature map and missing tooth position identifier of the test samples provided in the above embodiment, and will not be repeated in this embodiment.

[0064] S220. For each training sample, the training tooth loss data of the training sample is input into the untrained implant prediction model to obtain the output predicted implant parameters, and the supervised loss value corresponding to the training sample is determined based on the predicted implant parameters.

[0065] In an optional embodiment, the supervised loss value consists of a classification loss term and / or a regression loss term, wherein the classification loss term is used to measure the category matching error between the predicted implant model corresponding to the predicted implant parameters and the real implant model corresponding to the training samples; and the regression loss term is used to quantify the numerical deviation between the predicted implant parameters and the real implant parameters of the training samples.

[0066] In one specific embodiment, determining the supervised loss value corresponding to the training sample based on the predicted implant parameters includes: determining the predicted implant type corresponding to the training sample based on the predicted implant parameters, and determining a classification loss value based on the predicted implant type and the actual implant type corresponding to the training sample; determining a regression loss value based on the predicted implant parameters and the actual implant parameters of the training sample; and performing a weighted summation of the classification loss value and the regression loss value to obtain the supervised loss value.

[0067] For example, the classification loss term may be the cross-entropy loss function, and the regression loss term may be the smoothed L1 loss function or the L2 loss function, but is not limited to the given example.

[0068] The regression loss value Satisfy the following formula: in, This represents the actual implant parameters. This indicates the predicted implant parameters. This represents the deviation threshold, such as a deviation threshold of 0.3.

[0069] Specifically, when the numerical deviation between the actual implant parameters and the predicted implant parameters is small, the smooth L1 loss function exhibits a smooth loss, ensuring the stability of loss convergence; when the numerical deviation is large, the smooth L1 loss function exhibits a linear loss, avoiding the severe interference of outliers on gradient updates.

[0070] S230. Based on the supervised loss values, iteratively optimize the network parameters of the implant prediction model until the training termination condition is met, and obtain the trained implant prediction model.

[0071] In one specific embodiment, the mean of each supervised loss value is calculated to obtain an average loss value. Based on the average loss value, backpropagation is performed using a stochastic gradient descent or adaptive moment estimation optimizer to update the network parameters of the implant prediction model.

[0072] In another optional embodiment, the step of iteratively optimizing the network parameters of the implant prediction model based on each of the supervised loss values ​​includes: for each training sample, determining a regularization penalty value for the training sample based on the real implant parameter set and the predicted implant parameters of the training sample; determining a joint loss value based on each of the supervised loss values ​​and each of the regularization penalty values; and iteratively optimizing the network parameters of the implant prediction model based on the joint loss value.

[0073] In this embodiment, the actual implant parameter set represents the range of possible implant sizes, and the actual implant parameter set includes a set of size specifications corresponding to each size parameter in the predicted implant parameters. In an optional embodiment, the actual implant parameter set includes a set of diameter specifications and / or a set of length specifications.

[0074] For example, taking the Straumann BLT (Straumann Bone Level Tapered Implant) system as an example, the diameter specification set S dia ={3.3mm, 4.1mm, 4.8mm}, the length specification set S len ={8mm, 10mm, 12mm}.

[0075] The regularization penalty value is used to constrain the deviation between the predicted implant parameters and the range of available implant sizes, so that the output of the implant prediction model strictly falls within the clinically available discretized size center, avoiding the generation of unfit or non-standard implant parameters.

[0076] In an optional embodiment, determining the regularization penalty value of the training sample based on the real implant parameter set and the predicted implant parameters of the training sample includes: determining the absolute deviation between the predicted implant parameters and the nearest neighbor specification center of the real implant parameter set in each size dimension based on the real implant parameter set; and performing a weighted summation of the absolute deviations in each size dimension to obtain the regularization penalty value.

[0077] In one specific embodiment, each of the absolute deviations includes an absolute deviation in the diameter dimension and an absolute deviation in the length dimension, wherein the absolute deviation represents the absolute value of the difference between the predicted diameter / length value and the nearest available specification center of the corresponding size dimension, and the regularization penalty value represents the degree of deviation of the predicted implant parameter from the nearest neighbor specification center.

[0078] In another alternative embodiment, the regularization penalty value consists of a soft rounding constraint and / or a decision boundary penalty.

[0079] In an optional embodiment, determining the regularization penalty value of the training sample based on the real implant parameter set and the predicted implant parameters of the training sample includes: determining the specification deviation data of the predicted implant parameters based on the real implant parameter set, and determining the regularization penalty value of the training sample based on the specification deviation data.

[0080] In another optional embodiment, determining the regularization penalty value of the training sample based on the real implant parameter set and the predicted implant parameters of the training sample includes: determining a decision boundary set based on the real implant parameter set; determining the narrow margin data of the predicted implant parameters based on the decision boundary set; and determining the regularization penalty value of the training sample based on the narrow margin data.

[0081] In another optional embodiment, determining the regularization penalty value of the training sample based on the real implant parameter set and the predicted implant parameters of the training sample includes: determining the specification deviation data of the predicted implant parameters based on the real implant parameter set, and determining a soft rounding constraint value based on the specification deviation data; determining a decision boundary set based on the real implant parameter set; determining the narrow margin data of the predicted implant parameters based on the decision boundary set, and determining a decision boundary penalty value based on the narrow margin data; and determining the regularization penalty value of the training sample based on the soft rounding constraint value and the decision boundary penalty value.

[0082] The specification deviation data characterizes the overall degree of deviation of the predicted implant parameters from the actual implant parameter set.

[0083] Specifically, the specification deviation data includes the distance between the predicted diameter in the predicted implant parameters and the center of each diameter specification in the diameter specification set, and the distance between the predicted length in the predicted implant parameters and the center of each length specification in the length specification set.

[0084] For example, the soft rounding constraint value Satisfying the formula: in, Indicates the first of the predicted implant parameters One predicted size parameter, Represents the first in the actual implant parameter set A set of size specifications, , representing the temperature coefficient, is used to control the tightness of the constraint. For example, when the first... When the predicted size parameter is the predicted diameter When the first When the predicted dimension parameter is the predicted length However, this is not limited to the examples given.

[0085] Specifically, during backpropagation, the soft rounding constraint term generates a differentiable gradient guiding force for each predicted size parameter, pointing towards the nearest discrete specification center. The closer the predicted size parameter is to a discrete specification center, the more dominant the exponential term of that discrete specification center is in the summation term, and the stronger its corresponding gradient guiding force. When the predicted size parameter falls closer to the middle of two sets of discrete specification centers, the magnitudes of the gradient guiding forces corresponding to the two sets of discrete specification centers are similar, but their directions are mutually exclusive. The soft rounding constraint mechanism achieves dynamic gradient allocation through exponential weighted summation, continuously pulling the predicted size parameter towards the closer discrete specification center.

[0086] The decision boundary set consists of transitional implant parameters between two adjacent real implant parameters. Specifically, the transitional implant parameters represent the linear interpolation points between two adjacent real implant parameters; for example, the linear interpolation coefficient can be 0.5.

[0087] In an optional embodiment, determining the decision boundary set based on the actual implant parameter set includes: obtaining the target size specification set in the actual implant parameter set, and determining the decision boundary set corresponding to the target size specification set.

[0088] The target size specification set is either the set of size specifications corresponding to any size parameter in the actual implant parameter set, or the target size specification set is the set of size specifications in the actual implant parameter set that satisfies a preset distribution condition. Specifically, the preset distribution condition includes at least one of the following: the number of size specifications in the size specification set is greater than a specification discrete threshold, the specification spacing between two adjacent size specifications is uniformly distributed, and the specification spacing is greater than a preset spacing threshold.

[0089] The above implant length set S len and implant diameter set S dia For example, the target size specification set is the implant length set S.len The corresponding decision boundary set is B = {9mm, 11mm}.

[0090] For a set of size specifications with a small number of specifications or unevenly distributed spacing, the risk of predicted implant parameters oscillating in the transition region between adjacent specifications is relatively low. The advantage of setting preset distribution conditions in this embodiment is that it avoids introducing redundant gradient interference to the size parameters. Forcing them to converge to a center that is too large or too small will cause regression bias and prediction distortion.

[0091] In this embodiment, the narrow margin data characterizes the degree of closeness between the predicted implant parameters and each transition implant parameter.

[0092] For example, the decision boundary penalty value Satisfying the formula: in, Indicates the width of the narrow margin. This indicates the prediction length. Taking the decision boundary set B above as an example, the narrow margin width... It can be set to 0.2mm.

[0093] Specifically, the decision boundary penalty term will generate a large penalty gradient when the predicted length falls into the narrow margin range defined by the transition length, so as to push the predicted length away from the narrow margin range and move it closer to the center of the adjacent length specification, thereby achieving a smooth approximation of the discrete length specification.

[0094] Figure 6 This is a diagram illustrating the effect of a regularization penalty term provided in one embodiment of the present disclosure. Specifically, Figure 6 In the diagram, 8mm, 10mm, and 12mm represent the center of the length specification, 9mm and 11mm represent the center of the length transition, the grayscale box centered on the center of the length transition represents the narrow margin range, and the arrows indicate the gradient direction. Figure 6 Point A in the decision boundary does not fall within the narrow margin range, resulting in a penalty value. Zero, soft rounding constraint value A gradient force is applied to point A, moving it towards 8mm; point B lies within a narrow margin range between 8mm and 10mm, and the decision boundary penalty value is... A significant repulsive gradient is applied to point B, forcing it to move closer to the nearest discrete size center of 10 mm. Point C lies within a narrow margin range between 10 mm and 12 mm, and the decision boundary penalty value is... A significant repulsive gradient is applied to point C, forcing it to move closer to the nearest discrete size center (10 mm). This is compared to the decision boundary penalty value. soft rounding constraint value The gradient forces generated at points B and C are relatively weak, and mainly play an auxiliary guiding role.

[0095] In an optional embodiment, determining the regularization penalty value of the training sample based on the soft rounding constraint value and the decision boundary penalty value includes: performing a weighted summation of the soft rounding constraint value and the decision boundary penalty value to obtain the regularization penalty value of the training sample.

[0096] Specifically, the weight coefficients corresponding to the soft rounding constraint and the decision boundary penalty are fixed values ​​or dynamically adjusted values ​​related to the model training rounds.

[0097] In one optional embodiment, the weight coefficient of the soft rounding constraint term increases with each training iteration of the model, and the weight coefficient of the decision boundary penalty term also increases with each training iteration. Specifically, in the early stages of model training, the soft rounding constraint term guides the model to quickly learn the overall distribution pattern of implant parameters. In the later stages of model training, the weight of the decision boundary penalty term is increased to reduce the oscillation of implant parameters during the specification transition period, thereby improving the convergence stability of model training.

[0098] Specifically, based on the supervised loss value and regularization penalty value of each training sample, the sample loss value of the training sample is determined, and the average of the sample loss values ​​is calculated to obtain the joint loss value.

[0099] For example, the sample loss value of the training samples Satisfy the following formula: in, Represents the classification loss value. This represents the regression loss value. Indicates the number of training rounds of the model. and These represent the weighting coefficients of the soft rounding constraint and the decision boundary penalty, respectively.

[0100] For example, the training termination conditions include, but are not limited to, loss value convergence, the number of training epochs reaching a preset epoch threshold, and the accuracy of the validation dataset reaching a preset accuracy threshold.

[0101] S240. Obtain the three-dimensional oral image of the sample to be tested and the trained implant prediction model.

[0102] S250. The oral cavity three-dimensional image is segmented and located to obtain a tooth loss feature map and tooth loss position markers.

[0103] S260. Input the missing tooth position identifier into the linear modulation network to obtain the output modulation parameter data.

[0104] S270. Input the missing tooth feature map into the trunk prediction network, and perform an affine transformation on the intermediate feature map output by the corresponding target network layer according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network.

[0105] S280. Based on the target implant parameters, determine the target implant model corresponding to the sample to be tested.

[0106] S240-S280 in this embodiment are the same as those in the above embodiment. Figure 1 The S110-S150 shown are the same or similar, and will not be described again in this embodiment.

[0107] Traditional implant parameters are treated as continuous values ​​for regression prediction, ignoring the fact that implant parameters exist only in a limited and discrete range of manufacturer specifications. This leads to a large deviation between the predicted implant parameters and the actual available specifications, resulting in low accuracy and practicality of implant selection planning.

[0108] The technical solution of this embodiment introduces a boundary soft constraint mechanism and a decision boundary penalty mechanism during the model training stage, and iteratively optimizes the network parameters of the implant prediction model in combination with a supervised loss term to complete the training process of the implant prediction model. This solves the problem of mismatch between the predicted implant parameters and the actual available specifications, reduces the deviation between the predicted results and the actual available specifications, and thus improves the accuracy and practicality of implant selection planning.

[0109] It should be noted that the collection, use, storage, sharing and transfer of user personal information involved in the technical solution disclosed herein all comply with the provisions of relevant laws and regulations, and require notification to users and obtaining their consent or authorization. Where applicable, user personal information has been subjected to de-identification and / or anonymization and / or encryption technical processing.

[0110] The following are embodiments of the oral implant planning device provided in this disclosure. This device and the oral implant planning method described above belong to the same inventive concept. For details not described in detail in the embodiments of the oral implant planning device, please refer to the content of the oral implant planning method in the above embodiments.

[0111] Figure 7 This is a schematic diagram of the structure of an oral implant planning device provided in one embodiment of the present disclosure. Figure 7 As shown, the device includes: an implant prediction model acquisition module 310, a missing tooth position identification module 320, a modulation parameter data determination module 330, a target implant parameter output module 340, and a target implant model determination module 350.

[0112] The implant prediction model acquisition module 310 is used to acquire the oral three-dimensional image of the sample to be tested and the trained implant prediction model, which consists of a backbone prediction network and a linear modulation network. The missing tooth position identification module 320 is used to segment and locate the oral cavity three-dimensional image to obtain a missing tooth feature map and missing tooth position identification; The modulation parameter data determination module 330 is used to input the missing tooth position identifier into the linear modulation network to obtain the output modulation parameter data, wherein the modulation parameter data includes feature modulation parameters corresponding to at least one target network layer in the backbone prediction network. The target implant parameter output module 340 is used to input the missing tooth feature map into the trunk prediction network, and perform affine transformation on the intermediate feature map output by the corresponding target network layer according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network. The target implant model determination module 350 is used to determine the target implant model corresponding to the sample to be tested based on the target implant parameters.

[0113] The technical solution of this embodiment constructs a dual-pathway network framework driven by a backbone prediction network and a linear modulation network. This framework segments and locates 3D images of the oral cavity to obtain edentulous feature maps and edentulous tooth position identifiers. Based on these edentulous tooth position identifiers, the linear modulation network introduces feature modulation parameters into the backbone prediction network. This performs affine transformation on the target network layer in the backbone prediction network, enabling the backbone prediction network to have geometric perception capabilities of edentulous sites. The network then outputs customized implant parameters adapted to different tooth positions, ultimately achieving implant selection and matching. This solution addresses the problem of insufficient generalization performance across tooth positions caused by unified modeling of all tooth positions in the oral cavity, reduces prediction bias caused by anatomical differences in different tooth positions, and improves the reliability and accuracy of implant selection planning.

[0114] In an optional embodiment, the linear modulation network includes linear modulation modules that correspond one-to-one with the target network layers, and the linear modulation module includes an embedding layer and a feature transformation layer; The embedding layer is used to map the missing tooth position identifier into a high-dimensional embedding vector; the feature transformation layer is used to map the high-dimensional embedding vector into feature modulation parameters that match the corresponding target network layer.

[0115] In one alternative embodiment, the backbone prediction network adopts a selective state-space model architecture, and the feature transformation layer is a multilayer perceptron.

[0116] In an optional embodiment, the device further includes: The training dataset acquisition module is used to acquire the training dataset, which contains training missing tooth data corresponding to multiple training samples. The training missing tooth data includes missing tooth feature maps and missing tooth position identifiers. The supervised loss value determination module is used to input the training tooth loss data of the training sample into the untrained implant prediction model for each training sample, obtain the output predicted implant parameters, and determine the supervised loss value corresponding to the training sample based on the predicted implant parameters. The implant prediction model training module is used to iteratively optimize the network parameters of the implant prediction model based on the supervised loss values ​​until the training termination condition is met, thereby obtaining the trained implant prediction model.

[0117] In one optional embodiment, the implant prediction model training module includes: The regularization penalty value determination unit is used to determine the regularization penalty value of each training sample based on the real implant parameter set and the predicted implant parameters of the training sample. The joint loss value determination unit is used to determine the joint loss value based on each of the supervision loss values ​​and each of the regularization penalty values, and to iteratively optimize the network parameters of the implant prediction model based on the joint loss value. The actual implant parameter set includes a diameter specification set and / or a length specification set, and the actual implant parameter set represents the range of selectable implant specifications.

[0118] In an optional embodiment, the regularization penalty value is composed of a soft rounding constraint term and / or a decision boundary penalty term, and the regularization penalty value determination unit includes: The soft rounding constraint value determination subunit is used to determine the specification deviation data of the predicted implant parameters based on the actual implant parameter set, and to determine the soft rounding constraint value based on the specification deviation data; wherein, the specification deviation data characterizes the overall deviation degree of the predicted implant parameters relative to the actual implant parameter set; The decision boundary set determination subunit is used to determine the decision boundary set based on the actual implant parameter set, wherein the decision boundary set is composed of the transition implant parameters between two adjacent actual implant parameters; The decision boundary penalty value determination subunit is used to determine the narrow margin data of the predicted implant parameters based on the decision boundary set, and to determine the decision boundary penalty value based on the narrow margin data; wherein, the narrow margin data characterizes the degree of proximity between the predicted implant parameters and each transition implant parameter; The regularization penalty value determination subunit is used to determine the regularization penalty value of the training sample based on the soft rounding constraint value and the decision boundary penalty value.

[0119] In an optional embodiment, the regularization penalty value determination subunit is specifically used for: The regularization penalty value of the training sample is obtained by weighted summation of the soft rounding constraint value and the decision boundary penalty value; The weight coefficients corresponding to the soft rounding constraint and the decision boundary penalty increase with the number of training rounds of the model.

[0120] In one optional embodiment, the missing tooth position identification module 320 is specifically used for: The oral cavity 3D image is input into the trained image segmentation model to obtain the missing tooth position identification and tooth segmentation results; Based on the tooth segmentation results, region localization is performed to obtain the region of interest for missing teeth. Based on the region of interest for missing teeth, the tooth segmentation results and the three-dimensional oral cavity image are respectively cropped to obtain the region of interest mask and the region of interest image. A distance transformation is performed on the region of interest mask to obtain a distance transformation map, which represents the spatial distance distribution from the voxel in the region of interest mask to the nearest tooth voxel. Based on the region of interest mask, the region of interest image, and the distance transformation map, a tooth loss feature map is constructed.

[0121] In an alternative embodiment, the image segmentation model employs a globally context-aware U-shaped convolutional network architecture, the bottleneck of which integrates a selective state-space model.

[0122] The oral implant planning device provided in this disclosure can execute the oral implant planning method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects for executing the method.

[0123] Figure 8 This is a schematic diagram of an electronic device provided according to one embodiment of the present disclosure. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0124] like Figure 8 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor 11. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0125] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information or data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0126] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the dental implant planning method provided in the above embodiments.

[0127] In some embodiments, the oral implant planning method provided in the above embodiments can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the oral implant planning method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the oral implant planning method by any other suitable means (e.g., by means of firmware).

[0128] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication unit 19, or installed from storage unit 18, or installed from ROM 12. When the computer program is executed by processor 11, it performs the functions defined in the methods of embodiments of this disclosure.

[0129] Various embodiments of the systems and techniques described above can be implemented in the following systems or combinations thereof: digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), system-on-chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0130] Computer programs for implementing the oral implant planning methods of this disclosure can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0131] In the context of this application, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable storage medium. Examples of machine-readable storage media include, based on an electrical connection of at least one wire, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0132] To provide interaction with a user, the systems and techniques described herein can be implemented on a terminal device having: a display device for displaying information to the user (e.g., a cathode-ray tube (CRT) or liquid crystal display (LCD) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the terminal device. Other types of devices can also provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including audio input or tactile input).

[0133] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0134] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0135] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0136] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for planning dental implant procedures, characterized in that, include: Acquire three-dimensional images of the oral cavity of the sample to be tested and a trained implant prediction model, wherein the implant prediction model consists of a backbone prediction network and a linear modulation network; The oral cavity three-dimensional image is segmented and localized to obtain a tooth loss feature atlas and tooth loss position markers; The missing tooth position identifier is input into the linear modulation network to obtain the output modulation parameter data, which includes feature modulation parameters corresponding to at least one target network layer in the backbone prediction network. The missing tooth feature map is input into the trunk prediction network, and the intermediate feature map output by the corresponding target network layer is subjected to affine transformation according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network. Based on the target implant parameters, determine the target implant model corresponding to the sample to be tested.

2. The method according to claim 1, characterized in that, The linear modulation network includes linear modulation modules that correspond one-to-one with the target network layers. Each linear modulation module includes an embedding layer and a feature transformation layer. The embedding layer is used to map the missing tooth position identifier into a high-dimensional embedding vector; the feature transformation layer is used to map the high-dimensional embedding vector into feature modulation parameters that match the corresponding target network layer.

3. The method according to claim 2, characterized in that, The backbone prediction network adopts a selective state-space model architecture, and the feature transformation layer is a multilayer perceptron.

4. The method according to claim 1, characterized in that, The training method for the implant prediction model is as follows: Obtain a training dataset, which contains training missing tooth data corresponding to multiple training samples, including missing tooth feature maps and missing tooth position identifiers; For each training sample, the training tooth loss data of the training sample is input into the untrained implant prediction model to obtain the output predicted implant parameters, and the supervision loss value corresponding to the training sample is determined based on the predicted implant parameters. Based on the supervised loss values, the network parameters of the implant prediction model are iteratively optimized until the training termination condition is met, thus obtaining the trained implant prediction model.

5. The method according to claim 4, characterized in that, The step of iteratively optimizing the network parameters of the implant prediction model based on each of the supervised loss values ​​includes: For each training sample, a regularization penalty value is determined based on the real implant parameter set and the predicted implant parameters of the training sample. Based on the supervision loss value and the regularization penalty value, a joint loss value is determined, and the network parameters of the implant prediction model are iteratively optimized based on the joint loss value. The actual implant parameter set includes a diameter specification set and / or a length specification set, and the actual implant parameter set represents the range of selectable implant specifications.

6. The method according to claim 5, characterized in that, The regularization penalty value consists of a soft rounding constraint term and / or a decision boundary penalty term. Determining the regularization penalty value of the training samples based on the actual implant parameter set and the predicted implant parameters of the training samples includes: Based on the actual implant parameter set, the specification deviation data of the predicted implant parameters is determined, and based on the specification deviation data, a soft rounding constraint value is determined; wherein, the specification deviation data characterizes the overall degree of deviation of the predicted implant parameters relative to the actual implant parameter set; Based on the actual implant parameter set, a decision boundary set is determined, which is composed of transition implant parameters between two adjacent actual implant parameters; Based on the decision boundary set, the narrow margin data of the predicted implant parameters are determined, and the decision boundary penalty value is determined based on the narrow margin data; wherein, the narrow margin data characterizes the degree of closeness between the predicted implant parameters and each transition implant parameter; The regularization penalty value of the training sample is determined based on the soft rounding constraint value and the decision boundary penalty value.

7. The method according to claim 6, characterized in that, Determining the regularization penalty value of the training sample based on the soft rounding constraint value and the decision boundary penalty value includes: The regularization penalty value of the training sample is obtained by weighted summation of the soft rounding constraint value and the decision boundary penalty value; The weight coefficients corresponding to the soft rounding constraint and the decision boundary penalty increase with the number of training rounds of the model.

8. The method according to claim 1, characterized in that, The segmentation and localization of the oral cavity three-dimensional image to obtain a tooth loss feature atlas and tooth loss position markers includes: The oral cavity 3D image is input into the trained image segmentation model to obtain the missing tooth position identification and tooth segmentation results; Based on the tooth segmentation results, region localization is performed to obtain the region of interest for missing teeth. Based on the region of interest for missing teeth, the tooth segmentation results and the three-dimensional oral cavity image are respectively cropped to obtain the region of interest mask and the region of interest image. A distance transformation is performed on the region of interest mask to obtain a distance transformation map, which represents the spatial distance distribution from the voxel in the region of interest mask to the nearest tooth voxel. Based on the region of interest mask, the region of interest image, and the distance transformation map, a tooth loss feature map is constructed.

9. The method according to claim 8, characterized in that, The image segmentation model adopts a U-shaped convolutional network architecture based on global context awareness, and the bottleneck of the U-shaped convolutional network architecture is integrated with a selective state-space model.

10. A dental implant planning device, characterized in that, include: The implant prediction model acquisition module is used to acquire the oral three-dimensional image of the sample to be tested and the trained implant prediction model, which consists of a backbone prediction network and a linear modulation network. The missing tooth position identification module is used to segment and locate the oral cavity three-dimensional image to obtain a missing tooth feature map and missing tooth position identification; The modulation parameter data determination module is used to input the missing tooth position identifier into the linear modulation network to obtain the output modulation parameter data, wherein the modulation parameter data includes feature modulation parameters corresponding to at least one target network layer in the backbone prediction network. The target implant parameter output module is used to input the missing tooth feature map into the trunk prediction network, and perform affine transformation on the intermediate feature map output by the corresponding target network layer according to the feature modulation parameters to obtain the target implant parameters output by the trunk prediction network. The target implant model determination module is used to determine the target implant model corresponding to the sample to be tested based on the target implant parameters.

11. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the oral implant planning method according to any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the oral implant planning method according to any one of claims 1-9.

13. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the oral implant planning method according to any one of claims 1-9.

Images