Dental Press Journal of Orthodontics

Tabla de contenido Anterior Actual Siguiente

Tabla de contenido

ORIGINAL ARTICLE • Dental Press J. Orthod. 30 (4) • 2025 • https://doi.org/10.1590/2177-6709.30.4.e2524219.oar linkcopiar

Clinical decisions in Orthodontics using x-ray-based images and artificial intelligence approaches: a scoping review

Autoria SCIMAGO INSTITUTIONS RANKINGS

Introdução: A inteligência artificial (IA) na saúde aumentou suas aplicações nos últimos anos. A grande quantidade de dados disponível, devido ao aprimoramento do armazenamento e dos exames digitais, proporcionou um melhor conhecimento no planejamento do tratamento e abriu novas possibilidades de aplicação da IA no diagnóstico e planejamento ortodôntico.

Objetivo: O objetivo desta revisão de escopo foi examinar quando os modelos de IA aprimoram o processo de tomada de decisão clínica no diagnóstico e planejamento ortodôntico, com foco na utilização de imagens radiográficas.

Métodos: Estratégias eletrônicas de busca individual foram desenvolvidas e realizadas nas bases PubMed/Medline, Scopus, Web of Science, Embase, Lilacs e Cochrane Library (somente artigos em inglês publicados de janeiro de 2000 a 20 de outubro de 2021), buscando estudos relevantes que atendessem aos critérios de elegibilidade.

Resultados: Doze estudos foram incluídos, categorizados em cinco grupos distintos: cirurgia ortognática, osteoartrite da articulação temporomandibular (ATM), padrão esquelético, apneia obstrutiva do sono (AOS) e maturação/desenvolvimento esquelético. A maioria dos modelos de IA utilizados baseava-se em aprendizado profundo (Deep Learning - DL), sendo as radiografias laterais o tipo de imagem mais utilizado.

Conclusão: A integração da IA na prática clínica provavelmente continuará a evoluir, aprimorando o planejamento do tratamento em casos seletivos. As melhores aplicações foram observadas na osteoartrite da ATM, maturação esquelética, classificação esquelética, AOS e necessidade de cirurgia ortognática.

Palavras-chave:
Inteligência artificial; Ortodontia; Raios X

PICOS	Description
Population	X-ray based studies focused on diagnostic or planning
Intervention	Use of AI for diagnostic or planning purposes
Comparator	Expert opinions, human diagnosis or gold standard
Outcome	Performance metrics (accuracy, sensibility, sensitivity...)
Study design	Observational or experimental studies
Question	How does artificial intelligence can be applied to orthodontic diagnosis and planning using x-ray based images?

Name	Author, Year, Country	Study Design	Sample (n)	Mean/Range of Age(yr)	Imaging Analysis	Aim	Algorithm Architecture	Performance metrics	Author suggestions/recommendations	Limitations	What was evaluated	Results	Conclusions
AUTOMATED DETECTION OF TMJ OSTEOARTHRITIS BASED ON ARTIFICIAL INTELLIGENCE	Lee et al.,2020, Korea⁴⁴	Retrospective	314 (3,514 images)	mean ± SD age, 39.5 ± 18.2 y	CBCT	to develop a diagnostic tool to detect condylar resorption from CBCT images with AI	A single-shot detector (SSD)-a deep learning framework designed for object detection	Accuracy, precision, recall, and F1 score	additional imaging modalities, such as contrast enhanced MRI and bone scintigraphy, could provide more information with higher accuracy over TMJ's inflammation.	Only 1 observer classified the TMJ images in the training data set.	TMJ osteoarthritis on CBCT on adult patients	The average accuracy, precision, recall, and F1 score over were 0.86, 0.85, 0.84, and 0.84, respectively.	TMJOA from sagittal CBCT images is possible by using a deep neural networks model.
AUTOMATED SKELETAL CLASSIFICATION WITH LATERAL CEPHALOMETRY BASED ON ARTIFICIAL INTELLIGENCE	Yu et al., 2020, Korea⁴³	Retrospective	5890 patients	mean 25.4 y	Lateral radiographs	Propose a novel deep learning system that provides the fully automated, 1-step, end- to-end diagnosis for skeletal classification solely from the lateral cephalogram.	A multimodal convolutional neural network architecture with a modified DenseNet with pretrained weights for the ImageNet data set	Sensitivity, Specificity, AUC and Accuracy	To optimize diagnosis, skeletal components and severity of the discrepancy should be evaluated in future studies. Also, further research comparing this system and the original approach with landmark detection and additional studies on enhancing model performance are needed.	Data was collected from a single organization which may cause a selection bias	Three different models to evaluate sagittal and vertical skeletal pattern of patients	Sensitivity, Specificity, AUC and Accuracy for sagittal discrepancies were 93.55%, 96.77%, 0.965 to 0.991 and 95.70%, respectively. Sensitivity, Specificity, AUC and Accuracy for vertical discrepancies were 94.59%, 97.29%, 0.967 to 0.995 and 96.40%	Skeletal classification can be made with an AI-aided tool for sagittal and vertical discrepancies
MACHINE LEARNING FOR IMAGE-BASED DETECTION OF PATIENTS WITH OBSTRUCTIVE SLEEP APNEA: AN EXPLORATORY STUDY	Tsuiki et al., 2020, Japan⁴⁵	Retrospective	1389 patients (OSA = 867; control = 522)	OSA = 49.7 ± 8.9 Control = 41.2 ± 13.0	Lateral radiographs	test the hypothesis that a machine learning model could be used to differentiate severe OSA and non-OSA by 2-dimensional images	DCNN model called Visual Geometry Group (VGG-19)	Sensitivity, Specificity, Positive and negative likelihood ratio, Positive and negative predictive values and AUC area were evaluated for the three different groups	By the combination of demographic characteristics, anthropometric features and lateral cephalometric images may provide different AI models that could contribute to detect OSA	Data includes only Asian males from a single center	If this DCNN model could be used to detect patients with OSA based on 2-D images using three different areas of interest: Original image, main region and head only	Sensitivity, Specificity, Positive and negative likelihood ratio, Positive and negative predictive values and AUC area were evaluated for the three different groups: - Full image: 0.90, 0.77, 3.88, 0.14, 0.87, 0.82 and 0.89, respectively. - Main Region: 0.84, 0.81, 4.35, 0.20, 0.88, 0.75 and 0.92, respectively. - Head only: 0.71, 0.63, 1.91, 0.46, 0.85, 0.42 and 0.70, respectively	The DCNN accurately identified individuals with severe OSA using lateral cephalometric radiographs
COMPARISON OF DEEP LEARNING MODELS FOR CERVICAL VERTEBRAL MATURATION STAGE CLASSIFICATION ON LATERAL CEPHALOMETRIC RADIOGRAPHS	Seo et al, 2021, Korea⁴²	Observational	600	6-19 years	Lateral radiographs	evaluate and compare the performance of six state-of-the-art CNN-based deep learning models for CVM on lateral cephalometric radiographs, and implement visualization of CVM classification for each model using Grad-CAM technology	Different CNN models (ResNet-18, MobileNet-v2, ResNet-50, ResNet-101, Inception-v3 and Inception-ResNet-v2)	accuracy (1), recall (2), precision (3), F1-score (4), and area under the curve (AUC) values from the ROC curve	Higher quality data and development of better CNN architectures may improve the model's performance. It would be easier to evaluate if cervical vertebrae were segmented from surrounding structures	Small data	If different CNN models can distinguish the CVM classification on lateral cephalometric radiographs	Accuracy was higher than 90% for all models.	Inception-ResNet-v2 showed the best results for all metrics
DEEP CONVOLUTIONAL NEURAL NETWORKS BASED ANALYSIS OF CEPHALOMETRIC RADIOGRAPHS FOR DIFFERENTIAL DIAGNOSIS OF ORTHOGNATHIC SURGERY INDICATIONS	Lee et al., 2020, Korea²⁵	Retrospective	333 individuals	23.1 ± 5.1	Lateral radiographs	Create a model for the differential diagnosis of orthognathic surgery and orthodontic treatment with a DL algorithm using cephalometric radiographs.	Three DL models were tested: Modified-Alexnet, MobileNet and Resnet50	Sensitivity, Specificity, AUC and Accuracy	Future studies can be obtained when assessed with more data. Also, should analyze the impact of better model construction and their comparison with other models	Comparison of the performance was limited due to the lack of data	Differential diagnosis for orthognathic surgery based on cephalometric radiographs using three different DL models	Modified-Alexnet: 0.852 (Sensitivity), 0.973 (Specificity), 0.969 (AUC) and 0.919 (Accuracy). MobileNet: 0.761 (Sensitivity), 0.931 (Specificity), 0.908 (AUC) and 0.838 (Accuracy). Resnet50: 0.750 (Sensitivity), 0.944 (Specificity), 0.923 (AUC) and 0.838 (Accuracy).	DCNN can be applied to predict orthognathic surgery. Also it was possible to predict the most important diagnosis structures
DEEP LEARNING BASED PREDICTION OF NECESSITY FOR ORTHOGNATHIC SURGERY OF SKELETAL MALOCCLUSION USING CEPHALOGRAM IN KOREAN INDIVIDUALS	Shin et al, 2021, Korea²¹	Observational	840 (Class II:244, Class III: 477, Assimetry: 149) (Male: 461, Female: 379)	23.2y	Posterior and lateral radiographs	Develop a deep learning network to automatically predict the need for orthodontic surgery using cephalogram.	ResNet34, with convolution blocks stacked hierarchically (CNN)	accuracy, sensitivity, and specificity.	More studies should be encouraged to achieve public confidence.	Involves only Korean patients from only one hospital and small number of cases.	Exploited a deep learning framework to automatically predict the need for orthognathic surgery based on the cephalogram of the patients	accuracy, sensitivity, and specificity were 0.954, 0.844, and 0.993, respectively	DL can standardize the decision process and help orthodontists
DETERMINATION OF GROWTH AND DEVELOPMENT PERIODS IN ORTHODONTICS WITH ARTIFICIAL NEURAL NETWORK	Kök et al., 2021, Turkey⁴⁰	Retrospective	419 patients (CVM1:70, CVM2:70, CVM3:69, CVM4:70, CVM5:70, CVM6:70	8 to 17y	Lateral and hand-wrist radiographs	Determine the growth-development periods and gender from the cervical vertebrae using the artificial neural network (ANN).	Twenty-four ANN models were obtained and seven with success level and clinically applicable were selected	Accuracy, Sensitivity, Specificity and F1 score	Using all the information from cephalometric radiograph in addition to linear measurements can provide better results	Hand-wrist age determination is more advanced than cephalometry and cervical vertebra.	If the ANN models could determinate patient's cervical vertebral stage and gender.	The accuracy for the testing set was: 71.4% (ANN-1), 49.2% (ANN-2), 76.1% (ANN-3), 50.7% (ANN-4), 69.8% (ANN-5), 81% (ANN-6), 90.4% (ANN-7), 90.5% (ANN-Gender)	The growth-development periods and gender were determined from the cervical vertebrae by using ANN with satisfactory success
EARLY PREDICTION OF THE NEED FOR ORTHOGNATHIC SURGERY IN PATIENTS WITH REPAIRED UNILATERAL CLEFT LIP AND PALATE USING MACHINE LEARNING AND LONGITUDINAL LATERAL CEPHALOMETRIC ANALYSIS DATA	Lin et al., 2021, Korea²²	Retrospective	56 patients	T0: 6.3y and T1: 16.7y	Lateral radiographs	Determine the cephalo- metric predictors of the future need for orthognathic surgery in patients with repaired unilateral cleft lip and palate (UCLP) using machine learning.	Boruta method to identify the most relevant features out of the cephalometric predictors at T0 and XGBoost algorithm to predict the future need for orthognathic surgery	10-fold cross-validation accuracy with F1-score	Multi-center study to increase the sample.	Small sample size.	The most important features of the cephalogram and the need for orthognathic surgery	The prediction model had 87.4% accuracy and when ANB, PP-FH, CF and FCA variables are inserted	At age of 6 years, it was possible to predict the future need for orthognathic surgery using cephalometric predictors with a good accuracy
INFLUENCE OF THE DEPTH OF THE CONVOLUTIONAL NEURAL NETWORKS ON AN ARTIFICIAL INTELLIGENCE MODEL FOR DIAGNOSIS OF ORTHOGNATHIC SURGERY	Kim et al., 2021, Korea²⁴	Retrospective	960 patients (non-surgical: 640, surgical: 320)	24.6 ± 4.9	Lateral radiographs	investigate the relationship between image patterns in cephalometric radiographs and the need for orthognathic surgery, and report on a method for improving the accuracy of predictive models according to the depth of the neural network	CNN models (ResNet-18, 34, 50 and 101)	AUC, Accuracy, Sensitivity and Specificity	Multi-center data to improve the model's performance.	Single center study.	Predict the need of orthognathic surgery using cephalometric radiographs	AUC: 0.979 (ResNet-18), 0.974 (ResNet-34), 0.945 (ResNet-50), 0.944 (ResNet-101). Accuracy: 0.938 (ResNet-18), 0.936 (ResNet-34), 0.911 (ResNet-50), 0.913 (ResNet-101). Sensitivity: 0.882 (ResNet-18), 0.876 (ResNet-34), 0.806 (ResNet-50), 0.824 (ResNet-101). Specificity: 0.966 (ResNet-18), 0.966 (ResNet-34), 0.964 (ResNet-50), 0.958 (ResNet-101).	ResNet-18 and 34 had better performance than ResNet-50 and 101 models.
ON CONSTRUCTION OF TRANSFER LEARNING FOR FACIAL SYMMETRY ASSESSMENT BEFORE AND AFTER ORTHOGNATHIC SURGERY	Lin et al., 2021, Taiwan²³	Retrospective	71 patients	-	CBCT	A CNN model with a transfer learning approach for facial symmetry assessment based on 3-dimensional (3D) features to assist physicians in enhancing medical treatments.	CNN models (VGGNet16, VGGNet19, ResNet50 and Xception)	Accuracy	Larger number of data and more effective NN architectures can enhance the accuracy	The model can only be used for small data sets and can't be trained by new DL architecture.	Predict facial symmetry scores using the trained model.	The accuracy for VGG16, VGG19, ResNet50 and Xception was: 80%, 86%, 83% and 90%, respectively.	The most suitable model for the transfer learning was Xception model
PREDICTION OF HAND-WRIST MATURATION STAGES BASED ON CERVICAL VERTEBRAE IMAGES USING ARTIFICIAL INTELLIGENCE	Kim et al., 2021, Korea³¹	Retrospective	499 images from 455 patients	9.9 ± 2.6	Lateral and hand-wrist radiographs	Predict the hand-wrist maturation stages based on the CV images observed in lateral cepha- lograms, and to analyse the accuracy of the proposed algorithms.	Various combinations of regression models were used	Mean Absolute Error (MAE), round MAE, Root Mean Square of Error (RMSE) and Accuracy	Further studies with a larger sample with na even distribution of sex, chronological age and skeletal maturation. More features can also improve the prediction accuracy.	Small sample size and imbalanced data with a great number of patients whose SMI stage was 0.	Patients hand-wrist maturation based on CV images and its accuracy.	The final ensemble model consisted of eight machine learning models. The MAE, round MAE and RMSE were 0.90, 0.87 and 1.20, respectively.	Chronological age and sex increased the accuracy. CV images can be used to predict hand-wrist SMI using machine learning.
USAGE AND COMPARISON OF ARTIFICIAL INTELLIGENCE ALGORITHMS FOR DETERMINATION OF GROWTH AND DEVELOPMENT BY CERVICAL VERTEBRAE STAGES IN ORTHODONTICS	Kök et al., 2019, Turkey⁴¹	Retrospective	300 patients	8 to 17y	Lateral radiographs	determine cervical vertebrae stages (CVS) for growth and development periods by the frequently used seven artificial intelligence classifiers, and to compare the performance of these algorithms with each other.	Seven algorithms were selected:k-nearest neighbors (k-NN), Naive Bayes (NB), decision tree, ANN, SVM, Random forest and logistic regression	AUC, Accuracy, Precision, F1 score and Recall	There are no author suggestions or recomendations for future studies.	The author's did not mentioned any limitations over the study	Use seven different algorithms to determine cervical vertebrae stages from cephalometric radiograph and compare their performance	Decision tree had the highest accuracy for CVS1 to CVS4 (97.1%, 90.5%, 73.2%, 58.5%, respectively) and kNN for CVS5 (60.9%) and CVS6 (78.7%). ANN had the second-highest accuracy, except for CVS5.	ANN is the preferred method for determining CVS because it was the most stable algorithm

location_on

Dental Press International Av. Luís Teixeira Mendes, 2712 , 87015-001 - Maringá - PR, Tel: (55 44) 3033-9818 - Maringá - PR - Brazil
E-mail: artigos@dentalpress.com.br

rss_feed Acompañe los números de esta revista en su lector de RSS

Notificar error