OIIRR recognition system based on deep learning

Abstract

The invention discloses an OIIRR recognition system based on deep learning, and relates to the field of image processing, and the system comprises an obtaining module which is used for obtaining a CBCT image slice after orthodontic treatment of a user; the recognition module is used for inputting the CBCT image slices after orthodontic treatment into an OIIRR recognition model and outputting an OIIRR level; wherein the level of the OIIRR is 0 degree, 1 degree or 2 degrees; 0 degree is the tooth length change 0 mm before and after orthodontics; 1 degree is that the tooth length change before and after orthodontic treatment is less than or equal to 2mm; 2 degrees are that the tooth length change before and after orthodontic treatment is greater than 2mm; the OIIRR recognition model is obtained by training a deep learning network through a training data set; the training data set comprises CBCT image slices after orthodontic treatment of the user sample and corresponding OIIRR levels. According to the invention, the identification efficiency and accuracy of the OIIRR level can be improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image recognition, and particularly to an OIERR recognition system based on deep learning. Background Technique

[0002] Orthodontically-induced external root resorption (OIERR) is a relatively common and unpredictable complication that occurs along with orthodontic tooth movement. It is mainly manifested as a shortening of the root length, with the root apex being oblique, serrated, or abnormally blunt, etc. Clinical studies have shown that the occurrence of OIERR may affect nearly 80% of patients undergoing orthodontic treatment. Severe root resorption may even cause tooth movement or loss, thus affecting the progress and effect of orthodontic treatment. Therefore, the timely diagnosis of OIERR is crucial.

[0003] Orthodontists often use panoramic films or periapical films to examine OIERR. With the popularization of cone-beam computed tomography (CBCT) technology, CBCT has begun to be used clinically to detect OIERR, and most studies have also confirmed that CBCT is a more accurate and reliable method for evaluating root resorption than periapical films and panoramic films. However, since there is currently no single clinical method for diagnosing OIERR, orthodontists must rely on professional knowledge and clinical experience to evaluate OIERR, and this diagnostic decision often has certain subjective biases and instabilities. In addition, the correct evaluation of OIERR requires orthodontists to manually measure the root length before and after orthodontics on the radiograph, which is time-consuming and laborious, and prolongs the patient's visit time. At the same time, it is difficult for young doctors with insufficient clinical experience to quickly and accurately interpret radiographs, and new dentists often need to spend a lot of time studying and participating in radiograph interpretation training.

[0004] In recent years, the progress of artificial intelligence technology has brought about an innovation in medical data analysis and has achieved positive introduction and application in the field of oral medicine. Among them, the convolutional neural network (CNN) model has shown excellent performance in the automatic analysis of oral medical images. As one of the representative algorithms of deep learning, CNN is a type of feedforward neural network that contains convolutional calculations and has a deep structure, and has achieved good results in the detection and classification of various oral diseases, such as dental caries, periapical lesions, periodontitis, etc. Therefore, there is an urgent need for a CBCT image recognition system based on deep learning for orthodontically-induced external root resorption. Summary of the Invention

[0005] The objective of the present invention is to provide an OIERR recognition system based on deep learning, which can improve the recognition efficiency and accuracy of OIERR levels.

[0006] To achieve the above objective, the present invention provides the following solutions:

[0007] An OIERR recognition system based on deep learning, the recognition system includes:

[0008] An acquisition module, configured to acquire CBCT image slices after orthodontic treatment of a user;

[0009] A recognition module, connected to the acquisition module, configured to input the CBCT image slices after orthodontic treatment into an OIERR recognition model and output an OIERR level; wherein, the OIERR level is 0 degree, 1 degree or 2 degrees; 0 degree means that the tooth length change before and after orthodontics is 0 mm; 1 degree means that the tooth length change before and after orthodontics is less than or equal to 2 mm; 2 degree means that the tooth length change before and after orthodontics is greater than 2 mm; the OIERR recognition model is obtained by training a deep learning network with a training data set; the training data set includes CBCT image slices after orthodontic treatment of user samples and corresponding OIERR levels.

[0010] Optionally, the recognition system further includes a construction module;

[0011] The construction module is connected to the recognition module; the construction module is configured to construct the OIERR recognition model and send the OIERR recognition model to the recognition module.

[0012] Optionally, the construction module includes:

[0013] An acquisition sub-module, configured to acquire CBCT image slices before orthodontic treatment of user samples and CBCT image slices after orthodontic treatment of user samples, and calculate the tooth length change amount before and after orthodontics;

[0014] A level determination sub-module, connected to the acquisition sub-module, configured to determine the OIERR level of the user sample according to the tooth length change amount before and after orthodontics;

[0015] A preprocessing sub-module, connected to the acquisition sub-module, configured to preprocess the CBCT image slices after orthodontic treatment of the user sample to obtain an enhanced image after orthodontic treatment;

[0016] The construction sub-module is respectively connected to the level determination sub-module, the preprocessing sub-module and the recognition module, and is used to take the enhanced image after orthodontic treatment as the input, take the OIERR level of the user sample as the output, and based on the transfer learning and fine-tuning strategies, use the five-fold cross-validation method to train the deep learning network to obtain the OIERR recognition model, and send the OIERR recognition model to the recognition module.

[0017] Optionally, the preprocessing sub-module includes:

[0018] The sharpening unit is connected to the acquisition sub-module and is used to perform Gaussian sharpening on the CBCT image slices after orthodontic treatment of the user sample to obtain the sharpened image after orthodontic treatment;

[0019] The pixel adjustment unit is connected to the sharpening unit and is used to adjust the pixels of the sharpened image after orthodontic treatment to the preset pixels;

[0020] The normalization unit is connected to the pixel adjustment unit and is used to perform normalization processing on the image after adjusting the pixels to obtain the normalized image;

[0021] The data augmentation unit is connected to the normalization unit and the construction sub-module and is used to perform data augmentation operations on the normalized image to obtain the enhanced image after orthodontic treatment.

[0022] Optionally, the data augmentation operation includes at least any one of random horizontal flipping, random vertical flipping and random rotation.

[0023] Optionally, the construction module further includes an optimization sub-module;

[0024] The optimization sub-module is connected to the construction sub-module and is used to optimize the weights of the OIERR recognition model by using the adaptive moment estimation optimization algorithm.

[0025] Optionally, the deep learning network is a CNN model.

[0026] Optionally, the deep learning network is EfficientNet-B4.

[0027] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0028] In the present invention, the CBCT image slices after the orthodontic treatment of a user are input into the OIERR recognition model to obtain the OIERR level. The OIERR level is 0 degree, 1 degree or 2 degrees. The OIERR recognition model is obtained by training a deep learning network with the CBCT image slices after the orthodontic treatment of user samples as the input and the OIERR level as the output. By applying the OIERR recognition model, the recognition efficiency and accuracy of the OIERR level are improved. Description of the Drawings

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0030] Figure 1 It is a flowchart of the CBCT slices of the present invention;

[0031] Figure 2 It is a schematic diagram of the 0-level root resorption degree of the present invention;

[0032] Figure 3 It is a schematic diagram of the 1-level root resorption degree of the present invention;

[0033] Figure 4 It is a schematic diagram of the 2-level root resorption degree of the present invention;

[0034] Figure 5 It is a schematic diagram of the training process of the OIERR recognition model of the present invention;

[0035] Figure 6 It is a technical roadmap of the recognition system of the CBCT image of the present invention;

[0036] Figure 7 It is a schematic diagram of the model performance of the first-fold training of the OIERR recognition model of the present invention;

[0037] Figure 8 It is a schematic diagram of the model performance of the second-fold training of the OIERR recognition model of the present invention;

[0038] Figure 9 It is a schematic diagram of the model performance of the third-fold training of the OIERR recognition model of the present invention;

[0039] Figure 10 It is a schematic diagram of the model performance of the fourth-fold training of the OIERR recognition model of the present invention;

[0040] Figure 11 It is a schematic diagram of the model performance of the fifth-fold training of the OIERR recognition model of the present invention;

[0041] Figure 12 This is a schematic diagram of the confusion matrix of the present invention. Specific embodiments

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to 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 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 belong to the scope of protection of the present invention.

[0043] The object of the present invention is to provide an OIERR recognition system based on deep learning, which can improve the classification efficiency of CBCT image slices.

[0044] To make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Embodiment 1

[0046] As Figure 6 shown, the present invention provides an OIERR recognition system based on deep learning, and the recognition system includes:

[0047] An acquisition module, configured to acquire CBCT image slices after orthodontic treatment of a user.

[0048] A recognition module, connected to the acquisition module, configured to input the CBCT image slices after orthodontic treatment into an OIERR recognition model and output an OIERR level; wherein, the OIERR level is 0 degree, 1 degree or 2 degrees; 0 degree means that the tooth length change before and after orthodontics is 0 mm; 1 degree means that the tooth length change before and after orthodontics is less than or equal to 2 mm; 2 degree means that the tooth length change before and after orthodontics is greater than 2 mm; the OIERR recognition model is obtained by training a deep learning network with a training data set; the training data set includes CBCT image slices after orthodontic treatment of user samples and corresponding OIERR levels.

[0049] Specifically, the deep learning network is a CNN model.

[0050] Further, the deep learning network is EfficientNet-B4.

[0051] Wherein, the recognition system further includes a construction module.

[0052] The building module is connected to the recognition module; the building module is used to build the OIERR recognition model and send the OIERR recognition model to the recognition module.

[0053] Specifically, the building module includes:

[0054] An acquisition sub-module, configured to acquire CBCT image slices before orthodontic treatment of a user sample and CBCT image slices after orthodontic treatment of the user sample, and calculate the change in tooth length before and after orthodontics.

[0055] In practical applications, as Figure 1 shown, select eligible cases from the case database and obtain the CBCT image slices taken before and after their orthodontic treatment; on the panoramic view of the CBCT image slices, where the panoramic view is a two-dimensional projection image in the CBCT image, used to display the panoramic view of the oral cavity and jaws and is a component of the CBCT image, capture the sagittal scan slice of the incisor showing the maximum area of the pulp cavity, measure the distance from the incisor point to the apex of the root tip, and calculate the change in tooth length before and after orthodontics. Refer to Levander's 5-degree grading method and combine with the actual clinical situation to divide the root conditions into 3 groups: grade 0 (no root resorption); grade 1 (irregular contour formation at the root tip or micro-root resorption, ≤2 mm); grade 2 (severe root resorption at the root tip, >2 mm). Save the CBCT slices after orthodontics and form a three-classification data set according to the root conditions. Finally, a total of 2146 CBCT image slices were collected, including 826 slices with no root tip resorption, 1051 slices with micro-root tip resorption, and 269 slices with severe root tip resorption. The three types of images are as Figure 2 、 Figure 3 and Figure 4 shown. Divide the data set into a training set and a test set in a ratio of 8:2.

[0056] A level determination sub-module, connected to the acquisition sub-module, configured to determine the OIERR level of the user sample according to the change in tooth length before and after orthodontics.

[0057] A preprocessing sub-module, connected to the acquisition sub-module, configured to preprocess the CBCT image slices after orthodontic treatment of the user sample to obtain an enhanced image after orthodontic treatment.

[0058] In practical applications, perform Gaussian sharpening on the original image to improve image blurring and enhance edge sharpness, as Figure 3 shown; divide the data set into a training set and a test set in a ratio of 8:2; uniformly adjust the image after Gaussian sharpening to 256×256 pixels and normalize the pixel values to the range of 0-1; use the following data augmentation strategy to augment the training set: random horizontal and vertical flipping, random rotation (±10°).

[0059] Further, the preprocessing sub-module includes:

[0060] A sharpening unit, connected to the acquisition sub-module, for performing Gaussian sharpening on the CBCT image slices after orthodontic treatment of the user sample to obtain a sharpened image after orthodontic treatment.

[0061] A pixel adjustment unit, connected to the sharpening unit, for adjusting the pixels of the sharpened image after orthodontic treatment to a preset pixel.

[0062] A normalization unit, connected to the pixel adjustment unit, for performing normalization processing on the image after pixel adjustment to obtain a normalized image.

[0063] A data augmentation unit, connected to the normalization unit and the construction sub-module, for performing data augmentation operations on the normalized image to obtain an enhanced image after orthodontic treatment.

[0064] As a specific implementation manner, the data augmentation operation includes at least any one of random horizontal flipping, random vertical flipping, and random rotation.

[0065] A construction sub-module, respectively connected to the level determination sub-module, the preprocessing sub-module, and the recognition module, for taking the enhanced image after orthodontic treatment as input and the OIERR level of the user sample as output, and training a deep learning network using the five-fold cross-validation method based on transfer learning and fine-tuning strategies to obtain an OIERR recognition model, and sending the OIERR recognition model to the recognition module.

[0066] In addition, the construction module further includes an optimization sub-module.

[0067] The optimization sub-module, connected to the construction sub-module, for optimizing the weights of the OIERR recognition model using the adaptive moment estimation optimization algorithm.

[0068] In practical applications, as Figure 5 shown, based on transfer learning and fine-tuning strategies, using the five-fold cross-validation method, a CNN model is trained and verified on the preprocessed training set to develop an OIERR recognition model.

[0069] Transfer learning and fine-tuning strategy: For transfer learning, the pre-trained weights on the ImageNet dataset are used to transfer the initial model. Meanwhile, based on the original weights, the parameters of the entire model are trained and optimized using the training set. That is, the pre-trained version of the EfficientNet-B4 model is called from the pytorch library without freezing its convolutional layers, enabling the CNN to perform weight fine-tuning on the OIERR dataset.

[0070] The CNN model selects EfficientNet-B4. During training, the number of epochs is 60, the batch size is 16, the initial learning rate is 0.001, the number of steps for learning rate decay is 500, and the weight decay ratio is 1e-4. The adaptive moment estimation (ADAM) optimization algorithm is used to reduce the value of the loss function and update the network weight parameters.

[0071] The model is tested on the application test set, and the diagnostic performance of the OIERR recognition model is evaluated by calculating relevant indicators such as accuracy, precision, sensitivity, specificity, F1-score, and AUC.

[0072] The training results of the OIERR recognition model are as Figure 7 shown, the test results of the OIERR recognition model are shown in Table 1, and the confusion matrix is as Figure 12 shown.

[0073] Table 1 Statistical table of model test results

[0074]

[0075]

[0076] Since the five-fold cross-validation method is adopted in the model training of the present invention, the training process includes 5 folds. In the first fold, the data is divided into 5 parts, the first part is taken as the validation set, and the remaining 4 parts are taken as the training set, and then the model is trained and stopped after 40 epochs of iteration. Next, the second part of the data is used as the validation set, and the remaining 4 parts of the data are used as the training set, and the model is trained again until it stops. And so on, 5 times of training and validation are carried out, and finally 5 models and the performance indicators on 5 validation sets are obtained, and the average value is obtained to get the final model performance. Figures 7 - 11 In it, Fold 1 - Fold 5 respectively represent the model performance trained in the first fold to the fifth fold, that is, the relationship between the model training accuracy and validation accuracy and the number of iterations Epoch.

[0077] In Figure 12 , the confusion matrix is the result of model testing, Figure 12The results in [Figure 0] show that this figure is the confusion matrix of EfficientNet-B4, which can reflect the correspondence between the prediction results of the classification model and the actual situation, and visually display the prediction accuracy and misclassification of the model. The OIERR recognition model developed based on EfficientNet-B4 presents excellent OIERR level recognition effects, showing obvious correct predictions and fewer misclassifications.

[0078] The present invention has the following advantages:

[0079] 1. Taking the prior art as a comparative example to illustrate that the solution of the present invention solves the problems of the prior art.

[0080] It is proposed for the first time to automatically diagnose OIERR on CBCT image slices through a deep convolutional neural network model, which effectively improves the diagnosis speed of orthodontists, provides diagnostic references for inexperienced new clinical dentists, and solves the problems of time-consuming, strong subjectivity and decision-making uncertainty in manual film reading.

[0081] Application prospect: This study developed an artificial intelligence model based on CNN to automatically identify OIERR. Subsequently, the algorithm model can be integrated into the existing CBCT film reading software for algorithm implementation to achieve end-to-end intelligent diagnosis of root resorption.

[0082] 2. Taking a control experiment as a comparative example, such as the reaction temperature is 20 - 50 °C, and the experimental schemes corresponding to reaction temperatures outside the range of 20 - 50 can be provided.

[0083] This study is based on 6 classic CNN pre-training frameworks, namely EfficientNet-B1, EfficientNet-B2, EfficientNet-B3, EfficientNet-B4, EfficientNet-B5, and MobileNet-V3. Using simple cross-validation and K-fold (K = 5, 7, 10) cross-validation methods respectively, model training and validation are carried out on the preprocessed root resorption dataset, with a view to developing the optimal OIERR intelligent recognition model. And the diagnostic performance of the model was evaluated on the test set. According to the test results, it was found that EfficientNet-B4 trained under the 5-fold cross-validation method showed the best performance.

[0084] Table 2 Statistical table of test results of different CNN pre-training frameworks

[0085]

[0086] 3. The OIERR recognition system based on deep learning provided by the present invention is expected to provide an objective and reliable technical means for orthodontists to assist in diagnosing OIERR, and can improve the diagnosis efficiency and accuracy of OIERR.

[0087] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other.

[0088] In this article, specific examples are used to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention. At the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation on the present invention.

Claims

1. An OIERR recognition system based on deep learning, characterized in that, The recognition system includes: An acquisition module, configured to acquire CBCT image slices after orthodontic treatment of a user; A recognition module, connected to the acquisition module, configured to input the CBCT image slices after orthodontic treatment into an OIERR recognition model and output an OIERR level; wherein, the OIERR level is 0 degree, 1 degree or 2 degrees; 0 degree means that the tooth length change before and after orthodontics is 0 mm; 1 degree means that the tooth length change before and after orthodontics is less than or equal to 2 mm; 2 degree means that the tooth length change before and after orthodontics is greater than 2 mm; the OIERR recognition model is obtained by training a deep learning network with a training dataset; the training dataset includes CBCT image slices after orthodontic treatment of user samples and the corresponding OIERR levels.

2. The OIERR recognition system based on deep learning according to claim 1, characterized in that, The recognition system further includes a construction module; The construction module is connected to the recognition module; the construction module is configured to construct the OIERR recognition model and send the OIERR recognition model to the recognition module.

3. The OIERR recognition system based on deep learning according to claim 2, characterized in that, The construction module includes: An acquisition sub-module, configured to acquire CBCT image slices before orthodontic treatment of a user sample and CBCT image slices after orthodontic treatment of the user sample, and calculate the tooth length change amount before and after orthodontics; A level determination sub-module, connected to the acquisition sub-module, configured to determine the OIERR level of the user sample according to the tooth length change amount before and after orthodontics; A preprocessing sub-module, connected to the acquisition sub-module, configured to preprocess the CBCT image slices after orthodontic treatment of the user sample to obtain an enhanced image after orthodontic treatment; A construction sub-module, respectively connected to the level determination sub-module, the preprocessing sub-module and the recognition module, configured to use the enhanced image after orthodontic treatment as an input and the OIERR level of the user sample as an output, and train a deep learning network by using a five-fold cross-validation method based on transfer learning and fine-tuning strategies to obtain an OIERR recognition model, and send the OIERR recognition model to the recognition module.

4. The OIERR recognition system based on deep learning according to claim 3, characterized in that, The preprocessing sub-module includes: A sharpening unit, connected to the acquisition sub-module, configured to perform Gaussian sharpening on the CBCT image slices after orthodontic treatment of the user sample to obtain a sharpened image after orthodontic treatment; A pixel adjustment unit, connected to the sharpening unit, configured to adjust the pixels of the sharpened image after orthodontic treatment to a preset pixel; A normalization unit, connected to the pixel adjustment unit, configured to perform normalization processing on the image after pixel adjustment to obtain a normalized image; A data augmentation unit, connected to the normalization unit and the construction sub-module, configured to perform data augmentation operations on the normalized image to obtain an enhanced image after orthodontic treatment.

5. The OIERR recognition system based on deep learning according to claim 4, characterized in that, The data augmentation operations include at least any one of random horizontal flipping, random vertical flipping and random rotation.

6. The OIERR recognition system based on deep learning according to claim 3, characterized in that, The construction module further includes an optimization sub-module; The optimization sub-module, connected to the construction sub-module, is configured to optimize the weights of the OIERR recognition model by using an adaptive moment estimation optimization algorithm.

7. The OIERR recognition system based on deep learning according to claim 1, characterized in that The deep learning network is a CNN model.

8. The OIERR recognition system based on deep learning according to claim 1, characterized in that, The deep learning network is EfficientNet-B4.

Images