The accuracy of a designed software for automated localization of craniofacial landmarks on CBCT images

Abstract

BackgroundTwo-dimensional projection radiographs have been traditionally considered the modality of choice for cephalometric analysis. To overcome the shortcomings of two-dimensional images, three-dimensional computed tomography (CT) has been used to evaluate craniofacial structures. However, manual landmark detection depends on medical expertise, and the process is time-consuming. The present study was designed to produce software capable of automated localization of craniofacial landmarks on cone beam (CB) CT images based on image registration and to evaluate its accuracy.MethodsThe software was designed using MATLAB programming language. The technique was a combination of feature-based (principal axes registration) and voxel similarity-based methods for image registration. A total of 8 CBCT images were selected as our reference images for creating a head atlas. Then, 20 CBCT images were randomly selected as the test images for evaluating the method. Three experts twice located 14 landmarks in all 28 CBCT images during two examinations set 6 weeks apart. The differences in the distances of coordinates of each landmark on each image between manual and automated detection methods were calculated and reported as mean errors.ResultsThe combined intraclass correlation coefficient for intraobserver reliability was 0.89 and for interobserver reliability 0.87 (95% confidence interval, 0.82 to 0.93). The mean errors of all 14 landmarks were <4 mm. Additionally, 63.57% of landmarks had a mean error of <3 mm compared with manual detection (gold standard method).ConclusionThe accuracy of our approach for automated localization of craniofacial landmarks, which was based on combining feature-based and voxel similarity-based methods for image registration, was acceptable. Nevertheless we recommend repetition of this study using other techniques, such as intensity-based methods.

Automatic Craniomaxillofacial Landmark Digitization via Segmentation-Guided Partially-Joint Regression Forest Model and Multiscale Statistical Features

“…The mean error of our approach is 1.44 mm , which is significantly better than the mean errors of 3.15 mm , 2.41 mm and 3.40 mm in [10], [42] and [7], respectively. Although the datasets were different, the significantly reduced error implies the effectiveness of our method, compared with the state-of-the-art.…”

“…Finally, we qualitatively compared our results with CBCT landmark digitization methods published in [10], [42] and [7]. The mean error of our approach is 1.44 mm , which is significantly better than the mean errors of 3.15 mm , 2.41 mm and 3.40 mm in [10], [42] and [7], respectively.…”

“…Although the datasets were different, the significantly reduced error implies the effectiveness of our method, compared with the state-of-the-art. In order to provide quantitative comparison with other methods, we implemented a multi-atlas-based landmark digitization approach which is similar to reference [7], where the landmarks are mapped from the corresponding positions in the non-linearly aligned atlases with the averaging strategy. As shown in Fig.…”

“…To summarize, there are three mainstreams: 1) Interest point detection [2], [3], 2) Atlas-based landmark detection [4], [5], [6], [7], and 3) Machine-learning-based landmark detection.…”

Joint Craniomaxillofacial Bone Segmentation and Landmark Digitization by Context-Guided Fully Convolutional Networks

“…Current published reports regarding automated bone segmentation and landmark digitization can be generally divided into (1) multi-atlas (MA) based methods [1] and (2) learning based methods [2,3]. In the multi-atlas based methods , the segmentation and landmark digitization can be completed by transferring the labeled regions and landmarks from multi-atlas images to the target image via image registration [4].…”