Multi-modal Root Quantification Analysis Method and System Based on Residual Neural Network

Abstract

The present invention provides a multi-modal root quantification analysis method and system based on a residual neural network. The method obtains multi-modal image samples of the tooth root, and each multi-modal image sample includes a pair of 5T high-field oral MRI images of the same tooth root part, constituting a tooth root 5T MRI and CBCT image data set; constructs and trains a tooth root image registration model including an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module; inputs the to-be-registered 5T MRI image and the to-be-registered CBCT image into the trained tooth root image registration model to obtain the registered 5T MRI and CBCT images; obtains a quantification analysis result; and determines the periodontal health level. The present invention can achieve higher registration accuracy, determines the dental and periodontal health status based on the registered images, realizes an objective quantitative evaluation of the dental and periodontal health status, and improves the accuracy of the judgment of the dental and periodontal health status.

Description

Technical Field

[0001] The present invention relates to a multi-modal root quantification analysis method and system based on a residual neural network, belonging to the technical field of image processing and analysis. Background Art

[0002] In dental medicine, the quantification analysis of tooth roots is crucial for diagnosis and treatment planning. Traditional root analysis methods mainly rely on radiological techniques, such as X-ray films and CT scans. Although these methods are effective, they have radiation risks and lack precision when analyzing root details.

[0003] 5T high-field oral magnetic resonance imaging (5T MRI) and cone beam computed tomography (CBCT) are two commonly used imaging techniques in medical imaging. They have unique advantages in the detection of oral and maxillofacial hard and soft tissue structures respectively. 5T high-field oral magnetic resonance imaging (5T MRI) uses a high magnetic field strength, can provide higher imaging resolution, and has better imaging effects on soft tissues, enabling doctors to observe tiny lesions and structures in the oral cavity, which helps in the early detection and accurate diagnosis of periodontal diseases and pulp diseases.

[0004] With the development of artificial intelligence technology, especially deep learning technology, the multi-modal images of tooth roots are registered and fused, that is, the image data of the two imaging techniques (5T MRI images and CBCT images) are accurately aligned in the same coordinate system. Based on the registered results, the quantification analysis of the tooth body and periodontal conditions is carried out to obtain corresponding information, or a comprehensive evaluation of the oral hard tissues and periodontal soft tissues can be achieved, which helps doctors fully understand the oral health status of patients and provides more reliable support for the auxiliary diagnosis of root diseases.

[0005] Therefore, the present invention proposes a multi-modal root quantification analysis method and system based on a residual neural network for the problem of root quantification analysis, aiming to utilize the image information of 5T MRI and CBCT to achieve accurate registration and quantification analysis of tooth roots, providing an innovative solution for the auxiliary diagnosis and treatment planning of dental diseases. Summary of the Invention

[0006] The object of the present invention is to provide a multi-modal root quantification analysis method and system based on a residual neural network, which can automatically register root images and perform root quantification analysis based on the registration results, including the number of tooth roots, the number of pulp root canals, the thickness of the periodontal ligament, the height of the alveolar bone, and the depth of the periodontal pocket, etc., to solve the problem that the precision of the detail analysis of tooth roots in the prior art needs to be improved.

[0007] The technical solution of the present invention is:

[0008] A multi-modal root quantification analysis method based on a residual neural network, comprising the following steps:

[0009] S1. Obtain multi-modal image samples of the tooth root. Each multi-modal image sample includes a pair of 5T high-field oral MRI images (i.e., 5T MRI images) and CBCT images of the same tooth root part, and label the registration transformation parameters as labels to form a tooth root 5T MRI and CBCT image data set;

[0010] S2. Construct a tooth root image registration model. The tooth root image registration model includes an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module. The image preprocessing module preprocesses the input 5T MRI image and CBCT image to obtain a preprocessed 5T MRI image and a preprocessed CBCT image, and inputs them into the residual neural network. The residual neural network predicts the registration transformation parameters required to register the 5T MRI image to the CBCT image and outputs them to the rigid registration module. The rigid registration module uses a rigid registration method to align the overall structure of the tooth root of the 5T MRI image with the CBCT image to be registered. The spatial transformation module uses a bilinear interpolation method to perform a spatial transformation on the 5T MRI image to obtain the registered 5T MRI and CBCT images;

[0011] S3. Use the tooth root 5T MRI and CBCT image data set in step S1 to train the tooth root image registration model to obtain a trained tooth root image registration model;

[0012] S4. Input the 5T MRI image to be registered and the CBCT image to be registered into the trained tooth root image registration model to obtain the registered 5T MRI and CBCT images;

[0013] S5. Perform a quantification analysis on the registered 5T MRI and CBCT images obtained in step S4, including the analysis of the number of tooth roots, the number of pulp root canals, the thickness of the periodontal ligament, the height of the alveolar bone, and the depth of the periodontal pocket, to obtain a quantification analysis result;

[0014] S6. Based on the quantification analysis result obtained in step S5, determine the periodontal health level.

[0015] Further, in step S2, preprocessing the input 5T MRI image and CBCT image to obtain a preprocessed 5T MRI image and a preprocessed CBCT image, specifically:

[0016] Normalize the 5T MRI image and the CBCT image to the same set range;

[0017] Perform data augmentation on the 5T MRI image;

[0018] Resample the 5T MRI image and the CBCT image in the coronal plane using the bilinear interpolation method;

[0019] Perform random cropping on the 5T MRI image and the CBCT image, and make the cropped images contain the root structures, to obtain the preprocessed 5T MRI image and the preprocessed CBCT image.

[0020] Furthermore, in step S2, the residual neural network includes an input layer, a convolutional layer, a pooling layer, a residual module, a global average pooling layer, a fully connected layer, and an output layer.

[0021] Input layer: Used to input the preprocessed 5T MRI image and the preprocessed CBCT image into the convolutional layer;

[0022] Convolutional layer: Adopt a 7x7 convolutional kernel, with a stride of 2, an output channel number of 64, use zero padding, and then perform batch normalization and ReLU activation to obtain the convolved feature map and output it to the max pooling layer;

[0023] Max pooling layer: Use a 3x3 pooling unit with a stride of 2 to perform max pooling operation, which is used to reduce the spatial dimension of the convolved feature map and then output it to the residual block;

[0024] Residual module: Contains 4 cascaded residual units, where the stride of the first residual unit is 1, and the strides of the other 3 residual units are 2, which is used to reduce the size of the feature map and then output it to the global average pooling layer;

[0025] Global average pooling layer: Used to perform global average pooling on the input feature map, take the average of the features in each channel, and obtain a fixed-length 5T MRI image feature vector and a fixed-length CBCT image feature vector;

[0026] Fully connected layer: Used to learn the mapping relationship between the fixed-length 5T MRI image feature vector and the fixed-length CBCT image feature vector and the registration transformation parameters, and predict the obtained registration transformation parameters;

[0027] Output layer: Used to output the registration transformation parameters obtained by the fully connected layer.

[0028] Furthermore, the registration transformation parameters include a translation vector, a rotation angle, and a scaling factor.

[0029] Furthermore, in step S5, perform quantitative analysis based on the registered 5T MRI and CBCT images. Specifically, analyze the number of roots, the number of pulp canals, the root length, the periodontal ligament thickness, the alveolar bone height, and the periodontal pocket depth, and sum the set scores for the periodontal ligament thickness, the alveolar bone height, the periodontal pocket depth, and the pulp health degree respectively to obtain a total score; where:

[0030] ① Analyze the number of tooth roots: Identify the tooth roots in the image and count them;

[0031] ② Analyze the number of pulp root canals: Identify the root canals in the image and count them;

[0032] ③ Analyze the length of the tooth root: Identify the length of the tooth root in the image;

[0033] ④ Analyze the thickness of the periodontal ligament: Identify the thickness of the periodontal ligament in the image;

[0034] ⑤ Analyze the height of the alveolar bone: Identify the height of the alveolar bone in the image;

[0035] ⑥ Analyze the depth of the periodontal pocket: Identify the depth of the periodontal pocket in the image;

[0036] ⑦ Score the thickness of the periodontal ligament: Score 0 if it is > 0.1mm and ≤ 0.4mm, score 1 if it is > 0.4mm or < 0.1mm;

[0037] ⑧ Score the height of the alveolar bone: Score 0 if the alveolar bone height ≥ 2 / 3 of the tooth root length, score 1 if the alveolar bone height > 1 / 2 of the tooth root length and < 2 / 3 of the tooth root length, score 2 if it is < 1 / 2 of the tooth root length or > 1 / 3 of the tooth root length, score 3 if it is < 1 / 3 of the tooth root length;

[0038] ⑨ Score the depth of the periodontal pocket: Score 0 if it is less than 3mm, score 1 if it is > 3mm and ≤ 5mm, score 2 if it is > 5mm;

[0039] ⑩ Score the degree of pulp health: Score 0 if no abnormality is identified, score 1 if inflammation is identified, score 2 if necrosis is identified.

[0040] Specifically, the total score = the score of the periodontal ligament thickness + the score of the alveolar bone height + the score of the periodontal pocket depth + the score of the pulp health degree, where:

[0041] The score in the range of 0 - 2 indicates healthy periodontium;

[0042] The score in the range of 3 - 4 indicates mild periodontal inflammation;

[0043] The score in the range of 5 - 6 indicates moderate periodontal inflammation;

[0044] The score in the range of 7 - 8 indicates severe periodontal inflammation.

[0045] A multi-modal tooth root quantification analysis system based on a residual neural network for implementing the method described in any one of the above, including an image acquisition module, a model construction module, a model training module, an image registration module, a quantification analysis module, and a result determination module,

[0046] Image acquisition module: Obtain multi-modal image samples of tooth roots. Each multi-modal image sample includes a pair of 5T high-field oral MRI images of the same tooth root part, i.e., 5T MRI images and CBCT images, and annotate the registration transformation parameters as labels to form a tooth root 5T MRI and CBCT image dataset;

[0047] Model construction module: Construct a tooth root image registration model. The tooth root image registration model includes an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module. The image preprocessing module preprocesses the input 5T MRI images and CBCT images to obtain preprocessed 5T MRI images and preprocessed CBCT images, and inputs them into the residual neural network. The residual neural network predicts the registration transformation parameters required to register the 5T MRI images to the CBCT images and outputs them to the rigid registration module. The rigid registration module uses a rigid registration method to align the overall structure of the tooth root of the 5T MRI image with the CBCT image to be registered. The spatial transformation module uses bilinear interpolation to perform spatial transformation on the 5T MRI image to obtain registered 5T MRI and CBCT images;

[0048] Model training module: Use the tooth root 5T MRI and CBCT image dataset to train the tooth root image registration model to obtain a trained tooth root image registration model;

[0049] Image registration module: Input the 5T MRI image to be registered and the CBCT image to be registered into the trained tooth root image registration model to obtain registered 5T MRI and CBCT images;

[0050] Quantitative analysis module: Perform quantitative analysis based on the registered 5T MRI and CBCT images, including analysis of the number of tooth roots, the number of pulp root canals, the length of the tooth roots, the thickness of the periodontal ligament, the height of the alveolar bone, and the depth of the periodontal pocket, to obtain quantitative analysis results;

[0051] Result determination module: Determine the periodontal health status based on the quantitative analysis results obtained by the quantitative analysis module.

[0052] The beneficial effects of the present invention are:

[0053] First, compared with the prior art, the multi-modal tooth root quantitative analysis method and system based on a residual neural network can achieve higher registration accuracy, determine the tooth body and periodontal health status based on the registered images, realize the objective quantitative evaluation of the tooth body and periodontal health status, and improve the accuracy of judging the tooth body and periodontal health status.

[0054] Second, since traditional methods may have difficulty capturing sufficient feature information, the present invention uses a residual neural network that can learn high-level feature representations of images, has stronger adaptability, can better handle complex image registration problems, can effectively improve the registration accuracy, and provides more reliable imaging support for the auxiliary diagnosis of root diseases.

[0055] Third, the multi-modal root quantification analysis method and system based on the residual neural network have stronger robustness. This method can learn abstract features in images through the residual neural network and has a certain degree of robustness to noise and interference in images. Therefore, in complex medical imaging scenarios, the present invention can better maintain the stability of registration.

[0056] Fourth, the multi-modal root quantification analysis method and system based on the residual neural network can achieve automation and high efficiency. For traditional methods that require manual selection and marking of feature points and estimation of registration transformation parameters, the method of the present invention based on the residual neural network can automatically learn feature representations and predict registration transformation parameters, so it has a higher degree of automation and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a schematic flowchart of the multi-modal root quantification analysis method based on the residual neural network according to an embodiment of the present invention;

[0058] Figure 2 is a schematic illustration of the root image registration model in the embodiment;

[0059] Figure 3 is a schematic illustration of the residual neural network in the embodiment;

[0060] Figure 4 is a schematic illustration of the set scores of the tooth body and periodontium in the embodiment;

[0061] Figure 5 is a schematic illustration of the multi-modal root quantification analysis system based on the residual neural network according to the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0063] The embodiment provides a multi-modal root quantification analysis method based on a residual neural network, as Figure 1 , including the following steps:

[0064] S1. Obtain root image samples. Each root image sample includes a pair of 5T high-field oral MRI images, i.e., 5T MRI images, and CBCT images of the same root part, and label the registration transformation parameters as labels to form a root 5T MRI and CBCT image dataset.

[0065] In step S1, collect root image samples from a medical image database. Each root image sample contains a pair of 5T MRI images and CBCT images, and label their registration transformation parameters such as translation, rotation, scaling, etc. Ensure that the 5T MRI and CBCT images in the dataset correspond to the same root part of the same patient to ensure the accuracy and reliability of registration. The MRI images are selected for T2-weighted imaging. T2-weighted imaging is sensitive to water, so it is suitable for evaluating the pathological states of pulp inflammation and fluid accumulation. Divide the root 5T MRI and CBCT image dataset into a 70% training set, a 15% validation set, and a 15% test set for model training, tuning, and evaluation.

[0066] S2. Construct a root image registration model, such as Figure 2 , the root image registration model includes an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module. The image preprocessing module preprocesses the input 5T MRI image and CBCT image to obtain the preprocessed 5T MRI image and the preprocessed CBCT image, and inputs them into the residual neural network. The residual neural network predicts the registration transformation parameters required to register the 5T MRI image to the CBCT image and outputs them to the rigid registration module. The rigid registration module uses a rigid registration method to align the overall structure of the root of the 5T MRI image with the CBCT image to be registered. The spatial transformation module uses bilinear interpolation to perform spatial transformation on the 5T MRI image to obtain the registered 5T MRI and CBCT images.

[0067] In step S2, preprocess the 5T MRI image and the CBCT image to obtain the preprocessed 5T MRI image and the preprocessed CBCT image. Specifically,

[0068] 1) Normalize the 5T MRI image and the CBCT image to the same set range. Specifically, divide the intensity value of the CBCT image by 1000 HU (Hounsfield unit), and truncate the values outside the range of [-1, 1]. Perform linear rescaling on the 5T MRI image and also convert it to the range of [-1, 1]. It can handle different intensity values generated by different imaging techniques and ensure the consistency of data during the training process;

[0069] 2) Perform data augmentation on the 5T MRI images; apply random color jitter (randomization of brightness and contrast: 0.2) to the 5T MRI images to simulate the brightness and contrast changes in real-world imaging and enhance the robustness of the model;

[0070] 3) Use the bilinear interpolation method to resample the paired images in the coronal plane to a uniform spatial resolution of 1x1 mm, with a slice thickness of 2.5 - 3.5 mm. Slice the volume into 2D coronal images and delete the slices that are not segmented;

[0071] 4) Perform random cropping on the 5T MRI images and CBCT images, resize the images to 1024x1024 pixels, and make the cropped images contain the root structure of the teeth, ensuring that the root structure in the cropped images is clearly visible and there is no irrelevant background or noise, to obtain the preprocessed 5T MRI images and preprocessed CBCT images.

[0072] Preprocess the paired 5T MRI images and CBCT images respectively. The preprocessing methods include image data normalization, image enhancement, image resampling, image cropping, etc. These operations can reduce the differences between the images, provide better inputs for the subsequent registration process, and improve the registration effect.

[0073] In step S2, select ResNet-50 as the residual neural network. The residual neural network includes an input layer, a convolutional layer, a pooling layer, a residual module, a global average pooling layer, a fully connected layer, and an output layer, as Figure 3 :

[0074] Input layer: used to input the preprocessed 5T MRI images and preprocessed CBCT images into the convolutional layer;

[0075] Convolutional layer: adopt a 7x7 convolutional kernel, with a stride of 2, an output channel number of 64, use zero padding, and then perform batch normalization and rectified linear unit ReLU activation to obtain the convolved feature map and output it to the max pooling layer; among them, the rectified linear unit ReLU maps negative input values to zero and keeps positive input values unchanged, and its mathematical expression is: f(x) = max(0, x), where x is the input value.

[0076] Max pooling layer: perform max pooling operation using a 3x3 pooling unit with a stride of 2 to reduce the spatial dimension of the convolved feature map and output it to the residual module;

[0077] Residual module: contains 4 cascaded residual units, and each residual unit contains several convolutional layers and batch normalization layers. The stride of the first residual unit is 1, and the strides of the other 3 residual units are 2, which are used to reduce the size of the feature map and output it to the global average pooling layer;

[0078] Global average pooling layer: It is used to perform global average pooling on the input feature map, take the average of the features within each channel, and obtain a fixed-length 5T MRI image feature vector and a fixed-length CBCT image feature vector;

[0079] Fully connected layer: It is used to learn the mapping relationship between the fixed-length 5T MRI image feature vector and the fixed-length CBCT image feature vector and the registration transformation parameters, and predict the registration transformation parameters;

[0080] Output layer: It is used to output the registration transformation parameters obtained by the fully connected layer. It contains three nodes, corresponding to the translation vector, rotation angle, and scaling factor respectively, and is used to map the 5T MRI image to the spatial position and direction of the CBCT image. These parameters can more comprehensively describe the spatial transformation relationship between images, thereby further improving the accuracy and stability of registration.

[0081] Residual neural network is a deep learning architecture that solves the problem of difficult training as the network depth increases by introducing the concept of "residual learning". In traditional convolutional neural networks, as the number of network layers increases, the problem of gradient vanishing or explosion becomes more serious, making it difficult to train deep networks. Residual neural networks solve this problem by adding skip connections or shortcut connections, allowing gradients to flow directly to the previous layers. In the field of dental root image registration, using a residual network as a feature extractor can improve the training efficiency and performance of the network, allow the construction of deeper network structures, and avoid the problem of gradient vanishing or explosion.

[0082] In step S2, the rigid registration module uses the affine transformation method to perform a rigid transformation on the 5T MRI image according to the registration transformation parameters. The 5T MRI image is aligned to the CBCT image through translation, rotation, and scaling operations. Among them, the translation operation is used to move the image in space, the rotation operation is used to adjust the rotation angle of the image, and the scaling operation is used to adjust the size of the image to better align with the target image. After rigid registration, the spatial transformation module uses the bilinear interpolation method to perform a spatial transformation on the 5T MRI image to ensure the quality and accuracy of the registered image. The registered 5T MRI and CBCT images are subjected to quality evaluation to ensure the accuracy and stability of registration. The root mean square error RMSE is used to calculate the gap between the predicted parameters and the labeled parameters to evaluate the registration error. The smaller the RMSE, the smaller the gap between the predicted parameters and the labeled parameters, the smaller the registration error, and the higher the accuracy and stability of registration. The registered 5T MRI and CBCT images are used for applications such as medical diagnosis and surgical planning. The registered images can be further processed, such as image segmentation and feature extraction, to obtain more diagnostic information.

[0083] S3. Use the root 5TMRI and CBCT image datasets in step S1 to train the root image registration model, and obtain the trained root image registration model.

[0084] In step S3, the mean square error (MSE) is used as the loss function to measure the difference between the output of the residual neural network in the root image registration model and the labeled registration transformation parameters. The Adam optimizer is used to train the residual neural network of the root image registration model, with the learning rate set to 0.001 and the batch size set to 32.

[0085] S4. Pass the to-be-registered 5TMRI image and the to-be-registered CBCT image through the trained root image registration model to obtain the registered 5TMRI and CBCT images.

[0086] In step S4, quantitative analysis is performed based on the registered 5TMRI and CBCT images. Specifically, it includes the analysis of the number of roots, the number of pulp root canals, the root length, the periodontal ligament thickness, the alveolar bone height, and the periodontal pocket depth, and the total score is obtained by summing the set scores for the periodontal ligament thickness, the alveolar bone height, the periodontal pocket depth, and the pulp health degree respectively; where:

[0087] ① Analyze the number of roots: Identify the roots in the image and count them.

[0088] ② Analyze the number of pulp root canals: Identify the root canals in the image and count them.

[0089] ③ Analyze the root length: Identify the root length in the image.

[0090] ④ Analyze the periodontal ligament thickness: Identify the periodontal ligament thickness in the image.

[0091] ⑤ Analyze the alveolar bone height: Identify the alveolar bone height in the image.

[0092] ⑥ Analyze the periodontal pocket depth: Identify the periodontal pocket depth in the image.

[0093] ⑦ Score the periodontal ligament thickness: Score 0 if it is > 0.1mm and ≤ 0.4mm, and score 1 if it is > 0.4mm or < 0.1mm.

[0094] ⑧ Score the alveolar bone height: Score 0 if the alveolar bone height ≥ 2 / 3 of the root length, score 1 if the alveolar bone height > 1 / 2 of the root length and < 2 / 3 of the root length, score 2 if it is < 1 / 2 of the root length or > 1 / 3 of the root length, and score 3 if it is < 1 / 3 of the root length.

[0095] ⑨ Score the periodontal pocket depth: Score 0 if it is less than 3mm, score 1 if it is > 3mm and ≤ 5mm, and score 2 if it is > 5mm.

[0096] ⑩Score the dental pulp health level: If no abnormality is identified, score 0; if inflammation is identified, score 1; if necrosis is identified, score 2.

[0097] Evaluate the health status of the dental pulp through the registered 5T MRI and CBCT images to determine whether there is inflammation, necrosis or normal status:

[0098] Normal dental pulp: The signal intensity is greater than the first threshold and not greater than the second threshold, and the difference in adjacent signal intensities is less than the set threshold, that is, medium to high signal is shown, the signal is uniform, and there are no obvious bright or dark spots.

[0099] Dental pulp inflammation: When the dental pulp becomes inflamed, due to the infiltration of edema and inflammatory cells, the water content in the dental pulp tissue increases. The signal intensity is greater than the third threshold, that is, acute inflammation shows a high signal (bright), indicating fluid accumulation; the signal intensity is greater than the second threshold and not greater than the third threshold, that is, the chronic inflammation signal is higher but not as significant as the acute phase.

[0100] Dental pulp necrosis: The signal intensity is not greater than the first threshold, that is, it shows a low signal (dark), especially when it is completely necrotic, the signal intensity is significantly reduced.

[0101] S5. Perform quantitative analysis on the registered 5T MRI and CBCT images obtained in step S4, including the analysis of the number of tooth roots, the number of dental pulp root canals, the root length, the periodontal ligament thickness, the alveolar bone height, and the periodontal pocket depth, to obtain the quantitative analysis results.

[0102] S6. Determine the periodontal health level based on the quantitative analysis results obtained in step S5.

[0103] In step S6, by scoring the root images, the health status of the hard tissues of the oral teeth and the periodontal soft tissues can be quantitatively evaluated. The higher the calculated total score, the worse the health status.

[0104] Specifically, the total score = the score of the periodontal ligament thickness + the score of the alveolar bone height + the score of the periodontal pocket depth + the score of the dental pulp health level, where:

[0105] The score in the range of 0-2 indicates healthy periodontium;

[0106] The score in the range of 3-4 indicates mild periodontal inflammation;

[0107] The score in the range of 5-6 indicates moderate periodontal inflammation;

[0108] The score in the range of 7-8 indicates severe periodontal inflammation.

[0109] Compared with the prior art, the multimodal root quantification analysis method based on the residual neural network can achieve higher registration accuracy. Based on the registered images, the dental and periodontal health status is determined, realizing the objective quantitative evaluation of the dental and periodontal health status and improving the accuracy of the judgment of the dental and periodontal health status.

[0110] The multimodal root quantification analysis method based on the residual neural network uses the residual neural network to predict the registration transformation parameters required to register 5T MRI images to CBCT images. It has better feature learning ability, can better capture the feature information in the images, helps to improve the accuracy and robustness of registration, can better optimize the entire registration process, reduces the error propagation and complexity in the traditional methods, and can significantly improve the registration accuracy. This method uses a residual network as the feature extractor. By introducing a residual learning framework, the residual network allows the network to learn feature representations beyond the identity mapping, which helps to capture deeper features, including complex local structures and global context information. Aiming at the problems of noise and artifacts that may exist in medical images, the residual module consists of multiple residual units, and each residual unit contains several layers. The layers within the residual unit learn the residual between the input and the output, that is, the part that the network hopes to predict, making it have better robustness to noise and small changes in the images and suitable for medical image registration.

[0111] Such as Figure 4 , the embodiment also provides a multimodal root quantification analysis method based on the residual neural network for implementing the method described in any one of the above, including an image acquisition module, a model construction module, a model training module, an image registration module, a quantification analysis module, and a result determination module.

[0112] Image acquisition module: Obtain root image samples. Each root image sample includes a pair of 5T high-field oral MRI images (i.e., 5T MRI images) and CBCT images of the same root part, and the registration transformation parameters are marked as labels to form a root 5T MRI and CBCT image dataset;

[0113] Model construction module: Construct a root image registration model. The root image registration model includes an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module. The image preprocessing module preprocesses the input 5T MRI images and CBCT images to obtain preprocessed 5T MRI images and preprocessed CBCT images, which are then input into the residual neural network. The residual neural network predicts the registration transformation parameters required to register the 5T MRI images to the CBCT images and outputs them to the rigid registration module. The rigid registration module uses a rigid registration method to align the overall structure of the roots in the 5T MRI images with the CBCT images to be registered. The spatial transformation module performs a spatial transformation on the 5T MRI images using the bilinear interpolation method to obtain the registered 5T MRI and CBCT images;

[0114] Model training module: Use the root 5T MRI and CBCT image datasets to train the root image registration model to obtain a trained root image registration model;

[0115] Image registration module: Input the 5T MRI image to be registered and the CBCT image to be registered into the trained root image registration model to obtain the registered 5T MRI and CBCT images.

[0116] Quantitative analysis module: Perform quantitative analysis based on the registered 5T MRI and CBCT images, including the analysis of the number of roots, the number of pulp canals, the periodontal ligament thickness, the alveolar bone height, and the periodontal pocket depth, to obtain the quantitative analysis results;

[0117] Result determination module: Determine the periodontal health status based on the quantitative analysis results obtained by the quantitative analysis module.

[0118] In view of the fact that traditional methods may be difficult to capture sufficient feature information, the present invention adopts a residual neural network that can learn high-level feature representations of images, has stronger adaptability, can better handle complex image registration problems, and can effectively improve the registration accuracy.

[0119] In this multi-modal root quantitative analysis method and system based on a residual neural network, in the root image registration model, after preprocessing the 5T MRI images and CBCT images, the residual neural network uses residual blocks to extract features and then a fully connected layer is used for parameter prediction. The residual blocks are used as feature extractors, which can extract and represent features from the input 5T MRI images and CBCT images, can directly use the original images for registration, and can reduce the complexity of intermediate steps and data conversion.

[0120] The multi-modal root quantification analysis method and system based on the residual neural network have stronger robustness. Through the residual neural network, this method can learn the abstract features in the image and has a certain robustness to the noise and interference in the image. Therefore, in complex medical imaging scenarios, the present invention can better maintain the stability of registration.

[0121] The multi-modal root quantification analysis method and system based on the residual neural network can achieve automation and high efficiency. For the traditional method that requires manual selection and marking of feature points and estimation of registration transformation parameters, the present invention can automatically learn feature representations and predict registration transformation parameters, so it has a higher degree of automation and efficiency.

[0122] Those skilled in the art should understand that those skilled in the art can implement variations in combination with the prior art and 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.

[0123] The above are only the implementation manners of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be thought of by those skilled in the art within the technical scope disclosed by the present invention without creative labor should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope defined by the claims.

Claims

1. A multi-modal root quantification analysis method based on a residual neural network, characterized in that: including the following steps, S1. Obtain multimodal image samples of tooth roots. Each multimodal image sample includes a pair of 5T high-field oral MRI images (i.e., 5T MRI images) and CBCT images of the same tooth root site, and label the registration transformation parameters as labels to form a tooth root 5T MRI and CBCT image dataset; S2. Construct a tooth root image registration model. The tooth root image registration model includes an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module. The image preprocessing module preprocesses the input 5T MRI images and CBCT images to obtain preprocessed 5T MRI images and preprocessed CBCT images and inputs them into the residual neural network. The residual neural network predicts the registration transformation parameters required to register the 5T MRI images to the CBCT images and outputs them to the rigid registration module. The rigid registration module uses a rigid registration method to align the overall structure of the tooth roots in the 5T MRI images with the CBCT images to be registered. The spatial transformation module uses bilinear interpolation to perform spatial transformation on the 5T MRI images to obtain registered 5T MRI and CBCT images. In step S2, when preprocessing the input 5T MRI images and CBCT images to obtain preprocessed 5T MRI images and preprocessed CBCT images, specifically, normalize the 5T MRI images and CBCT images to the same set range; perform data augmentation on the 5T MRI images; use bilinear interpolation to resample the 5T MRI images and CBCT images in the coronal plane; perform random cropping on the 5T MRI images and CBCT images, and ensure that the cropped images contain the structures of the tooth roots to obtain preprocessed 5T MRI images and preprocessed CBCT images; S3. Use the tooth root 5T MRI and CBCT image dataset in step S1 to train the tooth root image registration model to obtain a trained tooth root image registration model; S4. Pass the 5T MRI images to be registered and the CBCT images to be registered through the trained tooth root image registration model to obtain registered 5T MRI and CBCT images; S5. Perform quantitative analysis on the registered 5T MRI and CBCT images obtained in step S4, including the analysis of the number of tooth roots, the number of pulp root canals, the thickness of the periodontal ligament, the height of the alveolar bone, and the depth of the periodontal pocket, to obtain quantitative analysis results; S6. Determine the periodontal health status based on the quantitative analysis results obtained in step S5.

2. The multi-modal root quantification analysis method based on a residual neural network according to claim 1, wherein: In step S2, the residual neural network includes an input layer, a convolutional layer, a pooling layer, a residual module, a global average pooling layer, a fully connected layer, and an output layer. Input layer: used to input the preprocessed 5T MRI images and preprocessed CBCT images into the convolutional layer; Convolutional layer: uses a 7x7 convolutional kernel, a stride of 2, an output channel number of 64, uses zero padding, and then performs batch normalization and rectified linear unit ReLU activation to obtain a convolutional feature map and output it to the max pooling layer; Max pooling layer: Perform max pooling operation using 3x3 pooling units with a stride of 2 to reduce the spatial dimension of the feature map after convolution and output it to the residual block; Residual module: Consists of 4 cascaded residual units. The stride of the first residual unit is 1, and the strides of the other 3 residual units are 2. It is used to reduce the size of the feature map and output it to the global average pooling layer; Global average pooling layer: Used to perform global average pooling on the input feature map, taking the average of the features within each channel to obtain a fixed-length 5T MRI image feature vector and a fixed-length CBCT image feature vector; Fully connected layer: Used to learn the mapping relationship between the fixed-length 5T MRI image feature vector and the fixed-length CBCT image feature vector and the registration transformation parameters, and predict the registration transformation parameters; Output layer: Used to output the registration transformation parameters obtained by the fully connected layer.

3. The multimodal root quantification analysis method based on a residual neural network according to claim 2, wherein: The registration transformation parameters include translation vector, rotation angle, and scaling factor.

4. The multi-modal root quantification analysis method based on a residual neural network according to any one of claims 1-2, characterized in that: In step S5, perform quantitative analysis based on the registered 5T MRI and CBCT images. Specifically, analyze the number of tooth roots, the number of pulp root canals, the length of the tooth roots, the thickness of the periodontal ligament, the height of the alveolar bone, and the depth of the periodontal pocket, and sum the set scores for the thickness of the periodontal ligament, the height of the alveolar bone, the depth of the periodontal pocket, and the degree of pulp health to obtain a total score; where: ① Analyze the number of tooth roots: Identify the tooth roots in the image and count them; ② Analyze the number of pulp root canals: Identify the root canals in the image and count them; ③ Analyze the length of the tooth roots: Identify the length of the tooth roots in the image; ④ Analyze the thickness of the periodontal ligament: Identify the thickness of the periodontal ligament in the image; ⑤ Analyze the height of the alveolar bone: Identify the height of the alveolar bone in the image; ⑥ Analyze the depth of the periodontal pocket: Identify the depth of the periodontal pocket in the image; ⑦ Score the thickness of the periodontal ligament: Score 0 if it is > 0.1mm and ≤ 0.4mm, score 1 if it is > 0.4mm or < 0.1mm; ⑧ Score the height of the alveolar bone: Score 0 if the alveolar bone height ≥ 2 / 3 of the tooth root length, score 1 if the alveolar bone height > 1 / 2 of the tooth root length and < 2 / 3 of the tooth root length, score 2 if it is < 1 / 2 of the tooth root length or > 1 / 3 of the tooth root length, score 3 if it is < 1 / 3 of the tooth root length; ⑨ Score the depth of the periodontal pocket: Score 0 if it is less than 3 mm, score 1 if it is > 3mm and ≤ 5mm, score 2 if it is > 5mm; ⑩ Score the degree of pulp health: Score 0 if no abnormality is identified, score 1 if inflammation is identified, score 2 if necrosis is identified.

5. The multi-modal root quantification analysis method based on a residual neural network according to claim 4, characterized in that: Total score = score of the periodontal ligament thickness + score of the alveolar bone height + score of the periodontal pocket depth + score of the degree of pulp health, where: The score in the range of 0 - 2 indicates healthy periodontium; The score in the range of 3 - 4 indicates mild periodontal inflammation; The score in the range of 5 - 6 indicates moderate periodontal inflammation; The score in the range of 7 - 8 indicates severe periodontal inflammation.

6. A multi-modal root quantification analysis system based on a residual neural network for implementing the method according to any one of claims 1-5, characterized in that: Includes an image acquisition module, a model construction module, a model training module, an image registration module, a quantitative analysis module, and a result determination module. Image acquisition module: Obtain multi-modal image samples of tooth roots. Each multi-modal image sample includes a pair of 5T high-field oral MRI images of the same tooth root part, i.e., 5T MRI images and CBCT images, and annotate the registration transformation parameters as labels to form a tooth root 5T MRI and CBCT image dataset; Model construction module: Construct a tooth root image registration model. The tooth root image registration model includes an image preprocessing module, a residual neural network, a rigid registration module, and a spatial transformation module. The image preprocessing module preprocesses the input 5T MRI images and CBCT images to obtain preprocessed 5T MRI images and preprocessed CBCT images and inputs them into the residual neural network. The residual neural network predicts the registration transformation parameters required to register the 5T MRI images to the CBCT images and outputs them to the rigid registration module. The rigid registration module uses a rigid registration method to align the overall structure of the tooth roots in the 5T MRI images with the CBCT images to be registered. The spatial transformation module uses bilinear interpolation to perform spatial transformation on the 5T MRI images to obtain registered 5T MRI and CBCT images; Model training module: Use the tooth root 5T MRI and CBCT image dataset to train the tooth root image registration model to obtain a trained tooth root image registration model; Image registration module: Pass the 5T MRI image to be registered and the CBCT image to be registered through the trained tooth root image registration model to obtain registered 5T MRI and CBCT images; Quantitative analysis module: Perform quantitative analysis based on the registered 5T MRI and CBCT images, including the analysis of the number of tooth roots, the number of pulp canals, the length of the tooth roots, the thickness of the periodontal ligament, the height of the alveolar bone, and the depth of the periodontal pocket, to obtain quantitative analysis results; Result determination module: Determine the periodontal health status based on the quantitative analysis results obtained by the quantitative analysis module.

Images