Cephalometric analysis is integral to orthodontic diagnosis and treatment planning. Finding the spatial correlations and angular measurements between the anatomic landmarks requires precisely identifying landmarks on lateral and PA cephalograms. The accuracy of manual annotations of landmarks on tracing sheets or onscreen images depends on the clinician’s visual judgement based on his experience. The advent of digital computerised cephalometric analysis has overcome the cumbersome and time-consuming process of measuring angles, linear and other variables with conventional measuring instruments. However, the critical single step in determining the outcome is related to the accuracy of measurements, which is directly dependent on the precision of landmark identification. The accuracy of landmark identifications could suffer from intra-investigator systematic errors, investigators’ experience, image quality, fatigue factor, the mental state of the operator and other environmental influences.

The process of manually marking the anatomical landmarks by a professional take around 20–25 min and is a laborious and error-prone operation. The advent of automatic landmarking of cephalometric images, which significantly reduces this time, is a welcome relief for orthodontists and maxillofacial radiologists. This automated process not only saves time and ensures precision but also eliminates intra- and inter-investigator errors, providing a reliable and consistent method for cephalometric analysis. This chapter aimed to provide an overview of automated landmark detection techniques, challenges, available datasets, methods proposed by researchers, evaluation metrics to measure the performance of such systems and currently available software for cephalometric analysis.

Automatic landmark detection in cephalometric analysis is a process that utilises computer algorithms to identify anatomical and constructed points on cephalometric radiographs. This is achieved without the need for manual input by an orthodontist or maxillofacial radiologist.

The algorithms are designed to recognise specific patterns and shapes within the radiographs, effectively automating the process of landmark identification.

Template matching is primarily used in image recognition and object detection. The process involves an algorithm that searches for the location of a template (or reference) image within a source/input image. Fig. 34.1 illustrates the generalised working principles of a template-matching algorithm. The red mark indicates the target landmark, for which we create templates of different sizes around the target area, where size refers to the pixel dimensions of the image.

A basic principle of template matching algorithm.

Source: From Vasamsetti S, Sardana V, Kumar P, Kharbanda OP, Sardana HK. Automatic landmark identification in lateral cephalometric images using optimized template matching. J Med Imaging Health Inf. 2015;5(3):458–470. doi:10.1166/jmihi.2015.1426

The figure shows four templates:

• •

  Template (a) is the largest in terms of pixel size,

• •

  Template (b) is smaller than (a),

• •

  Template (c) is smaller than (b), and

• •

  Template (d) is the smallest.

In the first step, template (a) uses a sliding window approach to compare its features with those in the input image. Once a match is found, template (b) repeats the process within the detected region rather than the entire image. This procedure continues with templates (c) and (d).

In the final step, to locate landmarks with maximum accuracy within the smallest area, different machine learning algorithms are employed. This hierarchical approach not only saves time but also significantly improves the precision and efficiency of landmark detection. The proposed automatic algorithm was evaluated against manual marking by three experienced orthodontic specialists. All 24 landmarks (100%) were identified within a 3.0 mm error range of the manual markings. Additionally, 23 landmarks (96%) were detected within a 2.5 mm error range, and 16 landmarks (66.6%) were within a 2.0 mm error range. The optimised template matching (OTM) algorithm shows potential as a promising method for the automatic detection of anatomical landmarks on cephalometric images, instilling confidence in its accuracy.

This section explains the machine learning-based approach for automatic landmark detection on 2D cephalograms using support vector machines (SVMs) and decision trees. SVMs are supervised learning models employed for classification and regression tasks. They operate by identifying the hyperplane that optimally separates the classes within the feature space. SVMs are effective for high-dimensional spaces and have been widely used in image analysis. In this method, the process involves many steps, such as data processing and feature extraction model training, which is either based on SVM or decision trees.

• •

  Data pre-processing : 2D cephalogram images are annotated with predefined landmarks. Pre-processing steps include normalisation to ensure consistent image intensity and data augmentation, as well as to increase variability in the training data. Augmentation techniques such as rotation, translation and scaling help improve model generalisation.

• •

  Feature extraction: For SVMs and decision trees, feature extraction is a crucial step. We use a combination of intensity-based and geometric features to describe the local image regions around each landmark. Intensity-based features include pixel values and gradients, while geometric features capture the spatial relationships between landmarks.

Each square window is segmented into 4 × 4 adjacent smaller windows. For each smaller window, an 8-bin histogram of gradients is generated, covering direction angles from 0 to 2π. These 4 × 4 histograms of gradients are then combined into a single feature vector f (with a dimension of 128) for each image patch. This feature vector f of the image patch P is described as the SIFT descriptor ( Fig. 34.2 ).

SIFT algorithm-based feature extraction.

(A) Original image. (B) Geometry descriptor. (C) The extracted feature vector.

Source: From Wang S, Li H, Li J, Zhang Y, Zou B. Automatic analysis of lateral cephalograms based on multiresolution decision tree regression voting. J Healthc Eng. 2018 Nov 19; 2018:1797502. PMID: 30581546; PMCID: PMC627641. doi:10.1155/2018/1797502

Training model: The SVM model is trained using a set of labelled examples, with each example representing a local image region around a landmark. The objective is to find the hyperplane that maximises the margin between the landmark and non-landmark regions. The decision tree model is trained by recursively splitting the training data based on feature values, creating a tree structure where each node represents a decision based on a specific feature. The goal is to maximise information gain at each split, leading to the most accurate prediction of landmark locations.

Deep learning, especially CNNs, has become a potent tool for automating this process, offering potential improvements in accuracy and efficiency. CNNs are the most commonly used deep learning architectures for image-related tasks, including landmark detection in cephalograms. Key studies have developed various CNN-based models tailored for this purpose:

• 1.

  Single-stage CNNs : These models directly predict landmark coordinates from the input cephalogram in a single pass. Arik et al. proposed a deep CNN that learns to regress landmark positions from raw pixel values, achieving significant accuracy improvements over traditional methods.

• 2.

  Two-stage CNNs : In this approach, the first stage identifies potential regions of interest (ROIs), and the second stage refines landmark predictions within these ROIs. This method enhances accuracy by focusing on smaller, more manageable sections of the image.

• 3.

  Heatmap-based CNNs : These models predict heatmaps corresponding to the likelihood of landmark positions. The peaks in these heat maps indicate the most probable locations of the landmarks. Payer et al. demonstrated that using heatmaps improves the robustness of landmark detection against variations in image quality and patient anatomy.

• 4.

  U-Net : This architecture, initially developed for biomedical image segmentation, has been adapted for landmark detection. It uses an encoder-decoder structure with skip connections to preserve spatial information, improving the precision of landmark localisation.

• 5.

  Residual networks (ResNets) : Adding residual connections helps alleviate the vanishing gradient problem, allowing for the training of deeper networks. These deeper networks can capture more complex features necessary for accurate landmark detection.

• 6.

  Ensemble methods : Combining multiple models can enhance prediction accuracy. Ensemble approaches average the outputs of several CNNs to reduce the impact of individual model errors.

A fully automatic landmark detection model utilising a CNN is depicted in Fig. 34.3 . This deep learning model features a two-step structure consisting of a ROI machine and a detection machine ( Fig. 34.3 A). Each CNN is composed of eight convolutional layers, five pooling layers and two fully connected layers ( Fig. 34.3 B).

An architecture for fully automatic landmark detection model using a convolutional neural network (CNN).

(A) The workflow of the two-step machine architecture is built on the cropped region of interest (ROI) and the detection network. (B) The structure of the CNN model. ROI , Region of interest; PNS , posterior nasal spine; ELU , exponential linear units; conv , convolution layer, pooling , pooling layer, fc , fully connected layer. 20210077

Source: From Kim YH, Lee C, Ha EG, Choi YJ, Han SS. A fully deep learning model for the automatic identification of cephalometric landmarks. Imaging Sci Dent. 2021 Sep;51(3):299–306. doi: https://doi.org/10.5624/isd.

Highlights of the deep learning architecture/algorithm used for the identification of landmarks on 2D cephalograms from the year 2017 to 2020 on the different datasets are summarised in Table 34.1 . ^, Schwendicke et al. reviewed the accuracy of existing methods for landmark prediction error. Through meta-analysis, they reported that landmark prediction errors were centred around a 2 mm threshold, with a mean of −0.581 mm and a 95% confidence interval ranging from −1.264 mm to 0.102 mm. The proportion of landmarks detected within this 2-mm threshold was 0.799, with a confidence interval of 0.770–0.824.

TABLE 34.1

Data sets and deep learning algorithm used for identification of landmarks from year 2017 to 2020 ^,

First author	Data source used	Architecture/modelling framework
Arik 2017	IEEE ISBI 2015	A combination of CNN and shape model.
Chen 2019	IEEE ISBI 2015	Utilizes VGG-19, ResNet20 and Inception architectures with a tailored attentive feature pyramid fusion module.
Gilmour 2020	IEEE ISBI 2015	Adapted ResNet34 combined with a unique image pyramid strategy (emphasising spatial features).
Huang 2020	Training data: CQ500 CTs test data: IEEE ISBI 2015	LeNet-5 is used for ROI patches detection and ResNet50 is used for landmark detection within the patches.
Hwang 2020	Private dataset	Modified YOLO V3.
Kim 2020	Private data set and IEEE ISBI 2015	Stacked hourglass network architecture.
Lee 2020	IEEE ISBI 2015	Custom CNN for ROI and custom Bayesian CNN for landmark detection.
Lee 2019	Private dataset	Integration of customised CNNs model for classification of ROI and landmarks detection.
Lee 2019	Private dataset	VGG-19
Ma	Private dataset	A customised CNNs architecture for both classification as well as for regression task.
Muraev 2020	Unspecified	Multiclass FPN and ResNeXt-50 with Squeeze-and-Excitation modules.
Noothout 2020	IEEE ISBI 2015	Customized fully convolutional networks (FCNs) derived from ResNet34.
O’Neil 2018	Private dataset	Custom FCN paired with Atlas Correction.
Oh 2020	IEEE ISBI 2015	DACFL, a redesigned FCN model used as a combined for local feature extraction and anatomical structure detection and segmentation loss.
Park 2020	Private dataset	Employs YOLO V3 and SSD.
Qian 2020	IEEE ISBI 2015	A designed Cepha-NN.
Song 2020	IEEE ISBI 2015	ROI extraction utilizing ResNet50.
Yun 2020	Private dataset	Customised CNNs model.
Zhong 2020	IEEE ISBI 2015	Two-stage U-Net models (global and local).

Only gold members can continue reading. Log In or Register to continue

Share this:

• Click to share on Twitter (Opens in new window)
• Click to share on Facebook (Opens in new window)

Related

Related posts:

1. Psychological implications of malocclusion and orthodontic treatment
2. Limits of orthodontic treatment modalities: 3D perspective
3. Classification and methods of recording malocclusion
4. Downs’ analysis
5. Prenatal development of the foetus concerning the craniofacial region
6. Clinical and functional evaluation

Stay updated, free dental videos. Join our Telegram channel

Join