Automatic Cephalometric Landmark Detect.
Cephalometric analyses of lateral x-rays of the skull are an important diagnosis tool in orthodontics. It is based on manually locating specific landmarks. This is a tedious, time-consuming and error prone task. We propose an automated system based on the use of Active Appearance Models (AAMs) and suitable for real applications.
Cephalometric analyses of lateral x-rays of the skull are an important diagnosis tool in orthodontics. It is based on manually locating specific landmarks. This is a tedious, time-consuming and error prone task.
To obtain a fully automated system it would be necessary to detect landmarks automatically, thus, reducing the time required to achieve an analysis, improving the accuracy of landmark localization and decreasing errors due to expert subjectivity.
Due to the lack of standardization of cephalograms, these images present high variability. Among others, it is necessary to consider: the anatomical morphology variability of the human head, the variability of texture depending on image quality, the variability of structures present in a cephalogram (not all the x-rays have the same size and include the same structures), the variability of capturing the x-ray (i.e. presence or absence of cephalostat, double structures appearing because of bad positioning of the head during capture) and the variability of the source (cephalograms to analyze could come from various radiologists and therefore be totally different).
We propose an automated system based on the use of Active Appearance Models (AAMs). To create a model with AAM we need a set of labelled images representative of the real variability of the object we want to segment. This will be annotated by an expert and form the training set of the model. Two main steps need to be achieved: training the model and segmenting new images. Special attention has been paid to clinical validation of our method since previous works on this field used few images, were tested in the training set and/or did not take into account the variability of the images.
Our contribution to previous works consists of generating a consistent training set that overcomes the above described drawbacks. Furthermore, we prove that AAMs is the suitable method for achieving adequate results in real cephalometric applications.
The AAM is trained using 115 hand-annotated images previously selected to take into account all the variability necessary for real systems. We have considered in our study 43 cephalometric landmarks to assess most of the cephalometric analysis. The method is tested with a leave-one-out scheme. Results obtained show an improvement over other techniques with an averaged accuracy of 2.32mm and an error lower than 5mm for 93.86% of the cephalometric points considered. 16.52% of the automatic initializations fail. Our method has proved to be useful to locate a considerable amount of landmarks with high accuracy rate for real cephalometric applications.
In future work we will improve the automatic initialization of the system as well as the accuracy of detecting the cephalometric points.
from 2004-01-01 to 2007-11-01
Sylvia Rueda López