Deep learning based prediction of midpalatal suture maturity
By using a deep learning-based method to predict the maturity of the mid-palatal suture, and extracting cervical spine morphological features from lateral cephalometric radiographs using the MSA-Hybrid convolutional neural network, the problem of insufficient accuracy in predicting the maturity of the mid-palatal suture was solved. This method achieves non-invasive and accurate assessment of the maturity of the mid-palatal suture, reducing radiation risks and costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AFFILIATED STOMATOLOGICAL HOSPITAL OF NANJING MEDICAL UNIV
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies lack sufficient accuracy in predicting the maturity of the mid-palatal suture, have insufficient reliability and exhibit gender differences, cannot be widely adopted in primary hospitals, and involve high radiation doses from CBCT scans.
A deep learning-based method for predicting the maturity of the mid-palatal suture was adopted. By acquiring lateral cephalometric images, an MSA-Hybrid convolutional neural network model was constructed to extract cervical spine morphological features and fuse deep feature vectors to predict the maturity of the mid-palatal suture.
It improves the accuracy and consistency of mid-palatal suture maturity prediction, reduces radiation risks and costs, and achieves non-invasive, objective mid-palatal suture maturity assessment, providing a scientific basis for clinical treatment.
Smart Images
Figure CN122156133A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a deep learning-based method for predicting the maturity of the mid-palatal suture, belonging to the field of intelligent medical image analysis technology. Background Technology
[0002] In the field of orthodontics, midpalatal suture maturity has become an important criterion for assessing whether adolescents are suitable for maxillary expansion. Accurate assessment of midpalatal suture maturity is crucial for developing treatment plans for adolescents with insufficient maxillary transverse development. Currently, clinical assessment of midpalatal suture maturity primarily utilizes cone-beam computed tomography (CBCT) to directly observe the structural features of the midpalatal suture. While CBCT offers high accuracy, it delivers a high radiation dose to adolescents, is expensive, and complex to operate, making it difficult to implement in primary care hospitals.
[0003] Recent studies have found a significant correlation between cervical spine bone age and mid-palatal suture maturity, providing a theoretical basis for non-invasive assessment of mid-palatal suture maturity. Traditional statistical methods indirectly infer mid-palatal suture maturity through cervical spine bone age staging (CVM staging), but existing statistical methods lack reliability, cannot cover all stages, and exhibit gender differences in diagnostic efficacy. Therefore, auxiliary diagnostic methods are needed to improve accuracy in clinical practice. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning-based method for predicting the maturity of the mid-palatal suture to address the problems of insufficient reliability and inadequate accuracy in predicting the maturity of the mid-palatal suture in existing technologies.
[0005] The technical solution of this invention is:
[0006] A deep learning-based method for predicting the maturity of the palatal midline includes the following steps:
[0007] S1. Obtain several lateral cephalometric radiograph samples, and label the maturity information of the mid-palatal suture to obtain a lateral cephalometric radiograph dataset.
[0008] S2. Construct a mid-palatal suture maturity prediction model. The mid-palatal suture maturity prediction model includes an image preprocessing module, a feature extraction module, a hybrid multi-scale attention convolutional neural network (MSA-Hybrid convolutional neural network), and a feature fusion module. The image preprocessing module preprocesses the lateral cephalometric images to obtain preprocessed images. The preprocessed images are then input into the MSA-Hybrid convolutional neural network to extract depth feature vectors. The feature extraction module extracts cervical spine morphological parameters from the preprocessed images to obtain morphological feature vectors. The feature fusion module fuses the depth feature vectors and morphological feature vectors to output the prediction results.
[0009] S3. The mid-palatal suture maturity prediction model was trained using a cephalometric radiograph dataset to obtain the trained mid-palatal suture maturity prediction model.
[0010] S4. Input the lateral cephalometric radiograph of the occipital region to be predicted into the trained mid-palatal suture maturity prediction model to obtain the mid-palatal suture maturity prediction result.
[0011] Furthermore, in step S1, the lateral cephalometric image samples are X-ray images, the gender ratio of the lateral cephalometric image samples is not greater than a set threshold, and the palatal midline maturity information includes cervical spine maturity indicators and palatal midline maturity stages, as well as key cervical spine morphological parameter measurements.
[0012] Furthermore, in the palatal midline maturity prediction model, the image preprocessing module preprocesses the lateral cephalometric images, including grayscale normalization, resolution unification, and image enhancement.
[0013] Furthermore, in the palatal suture maturity prediction model, the MSA-Hybrid convolutional neural network includes an input layer, an initial feature extraction layer, a backbone network, a global feature integration layer, and an output layer.
[0014] Input layer: Used to input the preprocessed image into the initial feature extraction layer;
[0015] Initial feature extraction layer: Several 3×3 convolutional kernels are used to perform preliminary processing on the input preprocessed image to generate an initial feature map and output it to the backbone network;
[0016] Backbone network: Deep feature maps are obtained by deep feature extraction of the initial feature map through a cross-scale feature fusion architecture that integrates medical knowledge, and the deep feature maps are output to the global feature integration layer;
[0017] Global Feature Integration Layer: Deep feature vectors are obtained by fusing the deep feature maps of the future autonomous backbone network;
[0018] Output layer: Outputs the deep feature vector obtained from the global feature integration layer.
[0019] Furthermore, in the MSA-Hybrid convolutional neural network, the backbone network adopts a cross-scale feature fusion architecture and integrates an anatomical region adaptive attention mechanism module and a medical prior knowledge embedding module.
[0020] The cross-scale feature fusion architecture includes four parallel feature extraction paths at different scales;
[0021] The cross-scale feature fusion architecture receives the initial feature map and distributes it to four parallel feature extraction paths;
[0022] Each feature extraction path includes an anatomical region adaptive attention mechanism module and a medical prior knowledge embedding module.
[0023] Anatomical region adaptive attention mechanism module: Generates cervical spine attention features and morphological analysis features from the input feature map, and combines them to obtain an intermediate feature map;
[0024] Medical prior knowledge embedding module: Dynamically incorporates pre-encoded high-dimensional embedding vectors into intermediate feature maps and outputs deep feature maps.
[0025] Furthermore, the anatomical region adaptive attention mechanism module includes parallel cervical spine attention branches and morphological analysis branches, as well as a feature fusion layer.
[0026] Cervical spine attention branch: Using parametric Gaussian positional coding technology, a spatial attention mask is generated by learning the center coordinates and standard deviation parameters of the Gaussian distribution. This spatial attention mask is multiplied with the input feature map to enhance the feature response of the key anatomical region of the second to fourth cervical vertebrae C2-C4, and output cervical spine attention features.
[0027] Morphological analysis branch: Utilizing deformable convolution technology, the offset of sampling points on the feature map is learned through additional convolutional layers. The sampling position of the convolutional kernel is dynamically adjusted according to the offset to adapt to irregular skeletal contours, and morphological analysis features are output.
[0028] Feature fusion layer: The intermediate feature map is obtained by combining the cervical spine focus features output by the cervical spine focus branch and the morphological analysis features output by the morphological analysis branch.
[0029] Furthermore, the medical prior knowledge embedding module includes a static prior embedding submodule and a dynamic prior modulation submodule.
[0030] Static prior embedding submodule: pre-encodes the input cervical spine maturity staging criteria into a high-dimensional embedding vector;
[0031] The dynamic prior modulation submodule is used to calculate the relevance weights, i.e., matching degree, between the embedded vector and the intermediate feature map in the channel dimension, and dynamically modulates the embedded vector onto the intermediate feature map based on the relevance weights, dynamically incorporating expert knowledge into it, and outputting a deep feature map.
[0032] Furthermore, in step S3, the palatal midline maturity prediction model is trained using the following loss function:
[0033] ,
[0034] in, and Using the cross-entropy loss function, The mean squared error function is used, and λ1, λ2, and λ3 are the weights used to balance the various loss terms.
[0035] The beneficial effects of this invention are:
[0036] I. This deep learning-based method for predicting the maturity of the palatal midline, by constructing a model for predicting the maturity of the palatal midline, can extract cervical spine morphological features from conventional lateral cephalometric radiographs. It uses an MSA-Hybrid convolutional neural network to extract anatomical features and perform analysis and prediction, which can reduce subjective differences in human interpretation, improve the accuracy, consistency, and reliability of bone age prediction, and achieve non-invasive, objective, and efficient prediction of the maturity of the palatal midline, providing a scientific basis for clinical arch expansion treatment decisions.
[0037] Second, compared with the prior art, the present invention can reduce radiation risk and is cost-effective. By analyzing the cervical spine morphological characteristics in conventional lateral cephalometric radiographs, the bone age of the palatine midline can be predicted without additional CBCT scans, which significantly reduces the radiation exposure dose of patients, reduces additional CBCT scan costs, reduces the economic burden on patients, and improves the utilization rate of medical resources.
[0038] Third, this deep learning-based method for predicting the maturity of the palatal midline, by establishing a sex-specific calibration curve, takes into account developmental differences between the sexes, providing patients with more accurate individualized assessment results. At the same time, the model's lightweight design allows it to run on various terminal devices, facilitating its widespread application in medical institutions at all levels, especially in primary care hospitals. Attached Figure Description
[0039] Figure 1 This is a flowchart illustrating the deep learning-based method for predicting the maturity of the palatal midline according to an embodiment of the present invention.
[0040] Figure 2 This is a schematic diagram illustrating the palatal midline maturity prediction model in the embodiment;
[0041] Figure 3 This is a schematic diagram illustrating the MSA-Hybrid convolutional neural network in the embodiment;
[0042] Figure 4 This is a schematic diagram illustrating the backbone network in the embodiment. Detailed Implementation
[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0044] The embodiment provides a deep learning-based method for predicting the maturity of the palatal midline, including the following steps, such as... Figure 1 :
[0045] S1. Obtain several lateral cephalometric radiographs, and label the maturity information of the mid-palatal suture to obtain a lateral cephalometric radiograph dataset.
[0046] In step S1, the lateral cephalometric radiograph samples are X-ray images, and the sex ratio of the lateral cephalometric radiograph samples is not greater than a set threshold. The palatal midline maturity information includes cervical spine maturity indicators and palatal midline maturity stages, as well as key cervical spine morphological parameter measurements.
[0047] In step S1, the acquired lateral cephalometric radiographs are X-ray images with complete mid-palatal suture maturity annotation information. To obtain accurate annotation information, the corresponding CBCT images can be used to determine the true maturity stage of the mid-palatal suture, and this stage information is used as a label to annotate the corresponding lateral cephalometric radiograph samples. The total sample size is no less than 500 cases to ensure a balanced male-to-female ratio and avoid gender bias. The data annotation content covers cervical spine maturity indicators (CVM stages I-VI), mid-palatal suture maturity stage (AE stage), and key cervical spine morphological parameter measurements. Key cervical spine morphological parameter measurements refer to the measurement and recording of a set number of cervical spine morphological parameters, such as 32, for each image using standard cephalometric methods, which serve as the raw data for subsequent feature extraction modules.
[0048] S2. Construct a model for predicting the maturity of the mid-palatal raphe, such as Figure 2 The palatal midline maturity prediction model includes an image preprocessing module, a feature extraction module, a hybrid multi-scale attention convolutional neural network (MSA-Hybrid convolutional neural network), and a feature fusion module. The image preprocessing module preprocesses the lateral cephalometric images to obtain preprocessed images. The preprocessed images are then input into the MSA-Hybrid convolutional neural network to extract depth feature vectors. The feature extraction module extracts cervical spine morphological parameters from the preprocessed images to obtain morphological feature vectors. The feature fusion module fuses the depth feature vectors and morphological feature vectors to output the prediction results.
[0049] In the mid-palatal suture maturity prediction model, the image preprocessing module preprocesses the lateral cephalometric images, including grayscale normalization, resolution unification, and image enhancement.
[0050] In the image preprocessing module, grayscale normalization standardizes the grayscale value range to the [0,1] interval to enhance image contrast; resolution unification uses a bilinear interpolation algorithm to uniformly adjust the resolution of all images to 2048×2048 pixels to ensure data quality consistency; image enhancement processing includes applying an adaptive histogram equalization algorithm (CLAHE) to enhance the local contrast of the image, and using Gaussian filtering (kernel size 5×5, σ=1.0) to reduce the impact of noise and improve the accuracy of subsequent feature extraction.
[0051] The feature extraction module works in parallel with the MSA-Hybrid convolutional neural network, responsible for screening and extracting the most informative morphological parameters from classic clinical measurements. Specifically, a random forest algorithm combined with analysis of variance is used to automatically select 12 key parameters with the strongest discriminative power for palatine suture maturity from the 32 initial cervical spine morphological parameters labeled in step S1, forming a 12-dimensional morphological feature vector. Emphasis is placed on morphological features of the C2-C4 vertebrae, including key morphological indicators such as the anteroposterior height ratio, vertebral concavity, and vertebral tilt angle.
[0052] In the mid-palatal suture maturity prediction model, the MSA-Hybrid convolutional neural network includes an input layer, an initial feature extraction layer, a backbone network, a global feature integration layer, and an output layer, such as... Figure 3 :
[0053] Input layer: Used to input the preprocessed image into the initial feature extraction layer;
[0054] Initial feature extraction layer: Several 3×3 convolutional kernels are used to perform preliminary processing on the input preprocessed image to generate an initial feature map and output it to the backbone network; the number of 3×3 convolutional kernels can be 32, with a stride of 2.
[0055] Backbone network: Deep feature maps are obtained by deep feature extraction of the initial feature map through a cross-scale feature fusion architecture that integrates medical knowledge, and the deep feature maps are output to the global feature integration layer;
[0056] like Figure 4 The backbone network adopts a cross-scale feature fusion architecture and integrates an anatomical region adaptive attention mechanism module and a medical prior knowledge embedding module.
[0057] The cross-scale feature fusion architecture includes four parallel feature extraction paths at different scales;
[0058] The cross-scale feature fusion architecture receives the initial feature map and distributes it to four parallel feature extraction paths;
[0059] Each feature extraction path includes an anatomical region adaptive attention mechanism module and a medical prior knowledge embedding module.
[0060] Anatomical region adaptive attention mechanism module: Generates cervical spine attention features and morphological analysis features from the input feature map, and combines them to obtain an intermediate feature map;
[0061] The anatomical region adaptive attention mechanism module includes parallel cervical spine attention branches and morphological analysis branches, as well as a feature fusion layer.
[0062] Cervical spine attention branch: Using parametric Gaussian positional coding technology, a spatial attention mask is generated by learning the center coordinates and standard deviation parameters of the Gaussian distribution. This spatial attention mask is multiplied with the input feature map to enhance the feature response of the key anatomical region of the second to fourth cervical vertebrae C2-C4, and output cervical spine attention features.
[0063] Morphological analysis branch: Utilizing deformable convolution technology, the offset of sampling points on the feature map is learned through additional convolutional layers. The sampling position of the convolutional kernel is dynamically adjusted according to the offset to adapt to irregular skeletal contours, and morphological analysis features are output.
[0064] Feature fusion layer: The intermediate feature map is obtained by combining the cervical spine focus features output by the cervical spine focus branch and the morphological analysis features output by the morphological analysis branch.
[0065] The medical prior knowledge embedding module dynamically incorporates pre-encoded high-dimensional embedding vectors into intermediate feature maps and outputs a deep feature map. The medical prior knowledge embedding module is a post-modulation unit in the path. It includes a static prior embedding submodule and a dynamic prior modulation submodule.
[0066] Static prior embedding submodule: The cervical spine maturity staging criteria (such as "the correlation between C3 vertebral body lower margin depression and CVM III stage") are pre-encoded into high-dimensional embedding vectors.
[0067] The dynamic prior modulation submodule is used to calculate the relevance weights, i.e., matching degree, between the embedded vector and the intermediate feature map in the channel dimension, and dynamically modulates the embedded vector onto the intermediate feature map based on the relevance weights, dynamically incorporating expert knowledge into it, and outputting a deep feature map.
[0068] To address the challenge of simultaneously capturing both subtle skeletal features and overall anatomical distribution in lateral cephalometric radiographs, a cross-scale feature fusion architecture serves as the backbone of the network, capable of simultaneously capturing micron-level variations at vertebral margins and cervical spine growth trends. The backbone receives the initial feature map and distributes it to four parallel feature extraction paths at different scales (e.g., path 1 processes high-resolution features, path 4 processes low-resolution semantic features). An anatomical region adaptive attention mechanism module acts as a pre-processing unit within these paths. When the feature map enters a feature extraction path, it is first fed into the anatomical region adaptive attention mechanism module: 1. Cervical spine focus branch: Utilizing parameterized Gaussian positional encoding to generate an attention mask, forcing the network to focus on the critical anatomical regions C2-C4. 2. Morphological analysis branch: Employing deformable learnable convolutions, adaptively adjusting the convolution kernel shape to match irregular skeletal contours. The outputs of the two branches are fused to generate an intermediate feature map. This intermediate feature map contains anatomically significant information. To address the challenges of noise interference caused by non-standard patient positioning and overlapping background tissues during imaging, the anatomical region adaptive attention mechanism module utilizes a parameterized Gaussian mask to dynamically lock onto the C2-C4 region, ensuring high model stability across different imaging angles. For situations where deep learning "black box models" are disconnected from clinical diagnostic standards, the medical prior knowledge embedding module integrates expert staging logic into the feature map, achieving strong alignment between feature extraction and clinical logic. This significantly reduces the dispersion of prediction results and improves the consistency between AI assessment and expert clinical judgment.
[0069] Global Feature Integration Layer: The deep feature maps of the backbone network will be fused to obtain a deep feature vector. Specifically, the global feature integration layer adopts an adaptive weighting mechanism, which merges the deep feature maps extracted by the backbone network through the channel-level concatenation (Concat) operator, and then fuses them into a 1×1×1536-dimensional deep feature vector after global average pooling.
[0070] Output layer: Outputs the deep feature vector obtained from the global feature integration layer.
[0071] The MSA-Hybrid convolutional neural network employs a multi-path hybrid design. The network structure includes: an input layer receiving a 2048×2048 lateral cephalometric image; an initial feature extraction layer using 48 3×3 convolutional kernels with a stride of 2 to perform preliminary image processing and generate an initial feature map; a backbone network responsible for extracting core deep features from the initial feature map; and finally, a global feature integration layer using an adaptive weighting mechanism to fuse the multi-scale features extracted by the backbone network into a 1×1×1536-dimensional deep feature vector.
[0072] S3. The mid-palatal suture maturity prediction model is trained using a cephalometric radiograph dataset to obtain the trained mid-palatal suture maturity prediction model.
[0073] In step S3, the palatal midline maturity prediction model is trained using the following loss function:
[0074] ,
[0075] in, and Using the cross-entropy loss function, The mean squared error function is used, and λ1, λ2, and λ3 are the weights used to balance the various loss terms.
[0076] In step S3, λ1, λ2, and λ3 are 0.3, 0.3, and 0.4, respectively; the optimizer is AdamW, with a learning rate of 0.0001 and a weight decay parameter of 0.00001 to balance training stability and convergence speed; the batch size is determined to be 64, and the complete training cycle is 200 rounds; a two-stage training strategy is adopted, in which the backbone network parameters are fixed in the first stage and only the fully connected layers are trained, and in the second stage, all network parameters are unfrozen and training continues; a learning rate cosine annealing strategy and an early stopping mechanism are applied to ensure that the model learns fully and achieves optimal performance.
[0077] S4. Input the lateral cephalometric radiograph of the occipital region to be predicted into the trained mid-palatal suture maturity prediction model to obtain the mid-palatal suture maturity prediction result.
[0078] This deep learning-based method for predicting the maturity of the mid-palatal suture can extract cervical spine morphological features from conventional lateral cephalometric radiographs by constructing a mid-palatal suture maturity prediction model. It uses an MSA-Hybrid convolutional neural network to extract anatomical features and perform analysis and prediction, which can reduce subjective differences in human interpretation, improve the accuracy, consistency and reliability of bone age prediction, and achieve non-invasive, objective and efficient mid-palatal suture maturity prediction, providing a scientific basis for clinical arch expansion treatment decisions.
[0079] Compared with existing technologies, this invention can reduce radiation risks and is cost-effective. It predicts the bone age of the palatine suture by analyzing the morphological characteristics of the cervical spine in routine lateral cephalometric radiographs without the need for additional CBCT scans, which significantly reduces the radiation exposure dose to patients, reduces additional CBCT scan costs, reduces the economic burden on patients, and improves the utilization rate of medical resources.
[0080] This deep learning-based method for predicting mid-palatal suture maturity constructs a mid-palatal suture maturity prediction model, establishing a direct correlation between cervical spine morphology and mid-palatal suture maturity. It is the first to apply deep learning technology to establish an accurate predictive relationship between cervical spine morphological features and mid-palatal suture maturity, enabling non-invasive assessment of mid-palatal suture maturity from routine cephalometric radiographs. A dedicated MSA-Hybrid network architecture is designed for cervical spine-mid-palatal suture morphological correlation analysis, possessing adaptive extraction capabilities for medical anatomical features. A three-branch parallel output network architecture is constructed, incorporating CVM staging, mid-palatal suture staging, and numerical maturity assessment, fully leveraging the correlation between tasks to improve prediction accuracy. Knowledge distillation technology is employed to ensure usability on mobile devices and in primary healthcare institutions.
[0081] This deep learning-based method for predicting the maturity of the mid-palatal suture, compared with existing technologies, applies deep learning technology designed specifically for medical anatomical structure analysis to the correlation analysis between cervical spine morphology and mid-palatal suture maturity. Its core technologies include an anatomical region adaptive attention mechanism, cross-scale feature fusion path, and embedding of medical prior knowledge.
[0082] This deep learning-based method for predicting the maturity of the palatal midline employs an innovative MSA-Hybrid convolutional neural network, which possesses the ability to adaptively extract anatomical features and fuse multi-scale features. It can effectively extract cervical spine morphological feature information, ensuring the professionalism and efficiency of the network.
[0083] This deep learning-based method for predicting the maturity of the midpalatal suture requires only standard lateral cephalometric radiographs of the patient, without additional radiation exposure, to quickly output assessment results of the midpalatal suture maturity. This provides support for medical decision-making and offers individualized treatment recommendations based on the assessment results: when the predicted CVM < 4 and the midpalatal suture stage is AC, the system suggests "the bone is not fully mature, and arch expansion treatment can be considered"; when the predicted CVM ≥ 4 or the midpalatal suture stage is DE, the system suggests "the bone is basically mature, and conventional arch expansion treatment is not recommended; surgical-assisted arch expansion or other treatment options can be considered"; for borderline cases, the system provides detailed probability information and suggests that doctors make a comprehensive judgment based on the patient's clinical presentation. Borderline cases refer to situations where the model's prediction results are at a critical dividing point in clinical decision-making and have low confidence. For example: ① The model-predicted CVM stage is close to the clinical decision threshold of stage 4, and the model's prediction probabilities for stages 3 and 4 are very close; ② The model-predicted mid-palatal suture stage is in the transitional phase between stage C (can expand the palate) and stage D (not recommended), and the model's prediction probabilities for these two stages are similar. Detailed probability information refers to the specific probability values given by the model for all possible stages. For example, the system will display "CVM stage prediction results: {I:5%, II: 15%, III: 45%, IV: 32%, V: 3%, VI: 0%}".
[0084] Those skilled in the art should understand that variations can be implemented by combining existing technology with the above embodiments, which will not be elaborated here. Such variations do not affect the essence of the present invention, and will not be elaborated here either.
[0085] The above are merely embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention without creative effort should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.
Claims
1. A deep learning-based method for predicting the maturity of the palatal midline, characterized in that: Includes the following steps, S1. Obtain several lateral cephalometric radiograph samples, and label the maturity information of the mid-palatal suture to obtain a lateral cephalometric radiograph dataset. S2. Construct a mid-palatal suture maturity prediction model. The mid-palatal suture maturity prediction model includes an image preprocessing module, a feature extraction module, a hybrid multi-scale attention convolutional neural network (MSA-Hybrid convolutional neural network), and a feature fusion module. The image preprocessing module preprocesses the lateral cephalometric images to obtain preprocessed images. The preprocessed images are then input into the MSA-Hybrid convolutional neural network to extract depth feature vectors. The feature extraction module extracts cervical spine morphological parameters from the preprocessed images to obtain morphological feature vectors. The feature fusion module fuses the depth feature vectors and morphological feature vectors to output the prediction results. S3. The mid-palatal suture maturity prediction model was trained using a cephalometric radiograph dataset to obtain the trained mid-palatal suture maturity prediction model. S4. Input the lateral cephalometric radiograph of the occipital region to be predicted into the trained mid-palatal suture maturity prediction model to obtain the mid-palatal suture maturity prediction result.
2. The deep learning-based method for predicting the maturity of the palatal midline as described in claim 1, characterized in that: In step S1, the lateral cephalometric radiograph samples are X-ray images, and the sex ratio of the lateral cephalometric radiograph samples is not greater than a set threshold. The palatal midline maturity information includes cervical spine maturity indicators and palatal midline maturity stages, as well as key cervical spine morphological parameter measurements.
3. The deep learning-based method for predicting the maturity of the palatal midline as described in claim 1, characterized in that: In the mid-palatal suture maturity prediction model, the image preprocessing module preprocesses the lateral cephalometric images, including grayscale normalization, resolution unification, and image enhancement.
4. The deep learning-based method for predicting the maturity of the palatal midline as described in claim 1, characterized in that: In the mid-palatal suture maturity prediction model, the MSA-Hybrid convolutional neural network includes an input layer, an initial feature extraction layer, a backbone network, a global feature integration layer, and an output layer. Input layer: Used to input the preprocessed image into the initial feature extraction layer; Initial feature extraction layer: Several 3×3 convolutional kernels are used to perform preliminary processing on the input preprocessed image to generate an initial feature map and output it to the backbone network; Backbone network: Deep feature maps are obtained by deep feature extraction of the initial feature map through a cross-scale feature fusion architecture that integrates medical knowledge, and the deep feature maps are output to the global feature integration layer; Global Feature Integration Layer: Deep feature vectors are obtained by fusing the deep feature maps of the future autonomous backbone network; Output layer: Outputs the deep feature vector obtained from the global feature integration layer.
5. The deep learning-based method for predicting the maturity of the palatal midline as described in claim 4, characterized in that: In the MSA-Hybrid convolutional neural network, the backbone network adopts a cross-scale feature fusion architecture and integrates an anatomical region adaptive attention mechanism module and a medical prior knowledge embedding module. The cross-scale feature fusion architecture includes four parallel feature extraction paths at different scales; The cross-scale feature fusion architecture receives the initial feature map and distributes it to four parallel feature extraction paths; Each feature extraction path includes an anatomical region adaptive attention mechanism module and a medical prior knowledge embedding module. Anatomical region adaptive attention mechanism module: Generates cervical spine attention features and morphological analysis features from the input feature map, and combines them to obtain an intermediate feature map; Medical prior knowledge embedding module: Dynamically incorporates pre-encoded high-dimensional embedding vectors into intermediate feature maps and outputs deep feature maps.
6. The deep learning-based method for predicting the maturity of the palatal midline as described in claim 5, characterized in that: The anatomical region adaptive attention mechanism module includes parallel cervical spine attention branches and morphological analysis branches, as well as a feature fusion layer. Cervical spine attention branch: Using parametric Gaussian positional coding technology, a spatial attention mask is generated by learning the center coordinates and standard deviation parameters of the Gaussian distribution. This spatial attention mask is multiplied with the input feature map to enhance the feature response of the key anatomical region of the second to fourth cervical vertebrae C2-C4, and output cervical spine attention features. Morphological analysis branch: Utilizing deformable convolution technology, the offset of sampling points on the feature map is learned through additional convolutional layers. The sampling position of the convolutional kernel is dynamically adjusted according to the offset to adapt to irregular skeletal contours, and morphological analysis features are output. Feature fusion layer: The intermediate feature map is obtained by combining the cervical spine focus features output by the cervical spine focus branch and the morphological analysis features output by the morphological analysis branch.
7. The deep learning-based method for predicting the maturity of the palatal midline as described in claim 5, characterized in that: The medical prior knowledge embedding module includes a static prior embedding submodule and a dynamic prior modulation submodule. Static prior embedding submodule: pre-encodes the input cervical spine maturity staging criteria into a high-dimensional embedding vector; The dynamic prior modulation submodule is used to calculate the relevance weights, i.e., matching degree, between the embedded vector and the intermediate feature map in the channel dimension, and dynamically modulates the embedded vector onto the intermediate feature map based on the relevance weights, dynamically incorporating expert knowledge into it, and outputting a deep feature map.
8. The deep learning-based method for predicting the maturity of the palatal midline as described in any one of claims 1-7, characterized in that: In step S3, the palatal midline maturity prediction model is trained using the following loss function: ， in, and Using the cross-entropy loss function, The mean squared error function is used, and λ1, λ2, and λ3 are the weights used to balance the various loss terms.