Automated cephalometric landmarks localization

AI‐based cephalometric analysis is one of the most common applications of AI in orthodontics. The seminal work on automated cephalometric landmark detection was done in 1984 to detect one landmark (Cohen et al., 1984), followed by numerous methods using computer vision and AI techniques (Leonardi et al., 2008). Usually, supervised ML methods are used for automated identification of cephalometric landmarks (Park et al., 2019). The primary goal is to eliminate or minimize human errors and to reduce time to identify landmarks (Durão et al., 2015). An ML approach using random forest regression voting automatically located 19 landmarks in 24 seconds. The overall average point‐to‐point error was 2.2 ± 0.03 mm (Lindner et al., 2016). Recently, the more advanced ML method of deep learning has been introduced. DL consists of an input layer, multiple hidden layers, and an output layer (Lecun et al., 2015). For automated cephalometry, the radiographs serve as input data and the numerical grayscale values of each pixel serve as individual inputs for neurons of the input layer. The output layer is defined as pairs of X and Y coordinates for each cephalometric landmark (Kunz et al., 2020). Kunz et al. in 2020 showed that a customized convolutional neural network (CNN) DL algorithm could localize anatomical landmarks with precision comparable to that of experienced human examiners, but in a fraction of a second. However, this customized DL algorithm can detect a limited set of landmarks (Arik et al., 2017; Kunz et al., 2020; Nishimoto et al., 2019). To detect more anatomical landmarks, an advanced YOLO (you only look once) model, called YOLO v. 3, showed faster detection and higher accuracy in 1028 cephalograms with less than 0.9 mm in the mean absolute error (MAE) of coordinates when compared to the human (Park et al., 2019). Currently, automatic cephalometric analysis using AI is available in the form of web‐based services. Some of these are open source and several others are commercial (Figure 25.3).

Classification of skeletal maturation stages

Dentofacial orthopedic and orthodontic treatment planning is primarily influenced by the pubertal growth spurt and skeletal maturation. Traditionally, hand–wrist radiographic analysis is considered the gold standard in determining individuals’ growth rate and peak stage of growth. But cervical vertebrae maturation (CVM) staging has become a more favorable alternative method. Lateral cephalometric radiographs are part of the routine orthodontic records set and by avoiding exposing a hand–wrist radiograph we can avoid unnecessarily increased ionizing radiation exposure (McNamara and Franchi, 2018). However, CVM staging is usually assessed visually, which needs an experienced orthodontist. Furthermore, it is not integrated with digital cephalometric software. AI could assist in developing a computer‐based system to determine CVM maturation stages. A semi‐automated method was recently introduced to reduce human error (Amasya et al., 2020). In this method, the orthodontist digitizes 26 landmarks on the cervical vertebrae for manual extraction of the image feature. Then, an ML clinical decision support system uses the numerical values of the features to determine the CVM stage (Amasya et al., 2020).

Recently, DL has been extensively used in the medical field for image pattern recognition and classification. At the University of Illinois Chicago (UIC), a research group developed a fully automated high‐performance system to detect and classify CVM stages. Its system consisted of an aggregate channel features (ACF) object detector that automatically identifies cervical vertebrae in the lateral cephalometric radiograph as the region of interest (ROI). The outputs are fed to a DL custom‐designed CNN with directional filter (CNNDF). The initial layer of the CNN consists of directional filters to emphasize the edges of vertebral bodies in the ROI images. Then the convolution layers extract the information from the images using a two‐dimensional (2D) convolution operation, and the outputs are the images classified according to the CVM maturation stages. The UIC system can classify CVM stages into six classes (CS1–CS6) and the modified five‐stage classification methods (CVMS I–CVMS V) (Figure 25.4; Atici et al., 2022).

Cone beam computed tomography

Cone‐beam computed tomography (CBCT) can reconstruct a three‐dimensional (3D) image of the patient’s head, which is very useful in many orthodontic applications. CBCT is commonly used in 3D diagnosis, virtual treatment planning, and for localization of impacted and supernumerary teeth. Other applications include assessment of bone thickness, of root resorption, and examination of the temporomandibular joint (Scarfe et al., 2017). The first step in the 3D analysis is the localization of anatomical landmarks, which is a tedious and time‐consuming task. Shahidi et al. (2014) proposed an ML algorithm to automatically locate 14 anatomical landmarks on CBCT images. However, the mean deviation (3.40 mm) for all of the automatically identified landmarks was higher than the mean deviation (1.41 mm) for those that were manually detected. Montúfar et al. (2018a, 2018b) proposed two automatic landmark localization systems based on active shape models and a hybrid approach using active shape models followed by a 3D knowledge‐based searching algorithm. However, the mean deviation (2.51 mm) for all of the automatically identified landmarks in the hybrid system was lower than that of the system that only used active shape models (3.64 mm). Developing an automated ML‐based 3D cephalometry analysis framework is challenging because the 3D volume data has greater computational complexity (Lee et al., 2019) than the 2D version of cephalometry. The existing AI systems are not accurate enough for clinical use, but could be used for preliminary localization of orthodontic landmarks followed by manual correction before further orthodontic analyses.

Another application of AI in CBCT is the automated segmentation of anatomical structures. Chen et al. (2020) proposed a method for automatic segmentation of individual teeth in dental CBCT images. They used a V‐Net, which is a fully convolutional neural network, to generate a tooth probability map and a tooth surface map. This is input into a marker‐controlled watershed transform for tooth segmentation. They showed that a patch size of 64 is optimal for their segmentation method. They had a small dataset comprising only 25 dental CBCT scans due to the time requirement of making voxel‐level masks (Chen et al., 2020). Ezhov et al. (2019) presented the results of using a DL model V‐Net for segmenting CBCT into 32 teeth plus background. The method consists of two sequential models: the coarse model that outputs 33 classes and the fine model that outputs 2 classes if the data are of a given tooth type or not. They also presented the coarse–fine segmentation pipeline, which uses two types of datasets, coarse and fine. The coarse dataset is created with linearly interpreted bounding boxes of CBCT axial slices. The fine dataset is a per‐voxel mask created manually. The method uses average symmetric surface distance (ASD) as one of the metrics to evaluate the model. ASD is the average distance from all points on the boundary of the machine‐predicted mask to the ground truth mask and vice versa. Their model achieved 0.17 mm ASD. They also assessed the overall tooth localization performance by measuring binary voxel‐wise intersection over union (IoU) between the ground truth volumetric mask and the model prediction. They achieved 0.94 IoU (Ezhov et al., 2019).

Dental panoramic radiographs

The orthodontist uses dental panoramic radiographs as a popular initial tool for diagnosis. It allows for assessing both the presence and position of tooth buds, supernumerary teeth, impacted teeth, missing tooth buds, and temporomandibular joints. Analysis of these images is done by visual inspection (Wirtz et al., 2018). A robust and accurate AI model for automated analysis of these images could be a helpful screening tool. Automated analysis of panoramic radiographs requires building AI models capable of detecting, labeling, segmenting, and classifying the teeth and surrounding structures. Abdi et al. (2015) developed an algorithm for auto‐segmenting the mandible in dental panoramic radiographs. The algorithm was tested on a set of 95 panoramic x‐rays, with the results evaluated against the manual segmentations of three expert dentists. It achieved an average performance of >93% in dice similarity, specificity, and sensitivity. Wirtz et al. (2018) reported a novel method for automatic teeth segmentation of dental panoramic radiographs using a coupled shape model in combination with a neural network, but the wisdom teeth were not included in the model. Vinayahalingam et al. (2019) developed and validated an automated approach to detect and segment the third molars and mandibular nerve by DL using panoramic radiographs. Krois et al. (2019) applied deep CNNs to detect periodontal bone loss (PBL) on panoramic dental radiographs. CNNs trained on a limited number of radiographic image segments showed similar discrimination ability to dentists for assessing PBL on panoramic radiographs.

Digital three‐dimensional dental models

An essential diagnostic step is analysis of the orthodontic models. This analysis includes arch length analysis and relations of the anterior and posterior width of the dental arches, as well as sagittal occlusion of the canine, Angle’s classification of molar occlusion, and anterior vertical and horizontal overlap. Digital 3D dental models have replaced traditional plaster models as they provide an accurate, valid, and clinically useful alternative, and for the diagnostic set‐up too. Furthermore, digital 3D dental models are fundamental for computer‐aided‐design (CAD) dental systems, customized orthodontic appliances, and virtual surgical planning (Elnagar et al., 2017). Tooth segmentation and labeling are the most critical components of digital model analysis and CAD orthodontics; however, manual segmentation of the 3D digital models is a time‐consuming and tedious task (Elnagar et al., 2020). Automated segmentation of digital models would save time, reduce human error, and streamline the digital workflow in contemporary orthodontics practice.

Xu et al. (2019) proposed a data‐driven method for 3D dental mesh segmentation by using a deep CNN model. It receives a 3D dental model as input and outputs the label list of each triangle face. They suggested a label‐free mesh simplification method for preserving teeth boundary information and significantly improving the efficiency of the networks. The authors claim that their model achieves 99.06% accuracy for the upper dental model and 98.79% for the lower dental model.

Another DL AI method called MeshSegNet has been proposed to automatically label individual teeth on raw dental surfaces acquired by 3D intraoral scanners. MeshSegNet achieves segmentation results of 0.938, 0.946, and 0.934 for Dice similarity coefficient, sensitivity, and positive predictive value, respectively (Lian et al., 2020). To solve the lack of accurate scan data in narrow interdental spaces, Kim et al. (2020) used generative neural networks (GANs), consisting of a generator and a critic, for 3D digital model segmentation. The generator generates synthetic images, and the critic has to discriminate whether each is a real or a synthetic image. This leads to a feedback loop where the synthetic images and the critic improve. For the tooth segmentation task, they obtain results that prove an average improvement of 0.04 mm (Kim et al., 2020).