The Process of Automating Tooth Segmentation & Identification from Dental Radiographs

Abstract

Automating the process of tooth segmentation and identification from dental radiographs is essential to perform any further analysis on dental x-rays. Teeth segmentation from dental radiographic images is an essential step for any type of dental image automation. It is one of the most challenging aspect in the image processing of dental radiographs. In this paper, a mathematical morphological approach of teeth segmentation is proposed. Further, proposed is the use of Histogram of Oriented Gradients(HOG’s) and image invariants to use as features for training classifiers. The results using the two approaches as features are compared. The classification of images into molar and premolar has been done on manually cropped images. This paper is an attempt to use both segmentation and classification on segmented periapical x-ray images. The forensic odonatological applications of this approach is wide and of immense benefits in both forensic and biometric identification.Index Terms— Dental image processing, histogram oriented gradients, teeth segmentation, teeth classification.

Introduction

Recent disasters have emphasized the significance of automated dental identification systems. Statistics show that 75% of Tsunami victims in Thailand were similarly identified using dental records, compared to 0.5% identified using DNA[1]. This paper addresses the two important problems of an Automated Dental Identification System (ADIS). This system matches image features extracted from multiple dental radiographic records. Dental radiograph record of an individual usually consists of radiographic films. It is an essential step to accurately segment these films from their constituent dental records in order to extract the dental features and achieve high level of automated postmortem identification. In this paper, an automated approach is proposed to the problem of segmenting films from their radiographs and then to classify the teeth using various features. Challenges include the poor quality of dental radiographs, variety of angles used for the x rays and lack of distinctive features in images. This approach is based on concepts of enhancement, connectivity, mathematical morphology and feature extraction for training and classification. A benchmark for classification of tooth as molar and premolar using the constructed data set is also proposed here.

Materials and Methods

Related works. Research and works on Hu’s moment invariants are discussed in this section. Works on tooth identification has been performed only on manually cropped images. The authors in [10] enhanced the X-ray image with histogram equalization. The teeth were segmented with the assistance of Otsu’s method and Hu’s moment invariants are calculated and are used as the teeth’s features. In their paper tooth recognition was done by feature matching with Euclidian distance. Jindan Zhouc et al. [11] used bitewing images and separated each tooth into crown and root. They developed a method to identify missing teeth areas as well as the shapes of teeth. The authors used adaptive segmentation to extract the teeth contours. The result of this method held a precision of 95% in the top five most similar images. A.K Jain et al in [3] segmented the upper jaw from the lower jaw by detecting the gap valley between them. Afterwards, the technique isolates each tooth from its neighbors in each jaw by detecting the gaps between them using intensity integral projection. This approach is semiautomated since an initial valley gap point is required to detect the gap valley between the upper and lower jaw. A fully automated approach for dental X-ray images is introduced by Abdel-Mottaleb et al [5].

The technique depends on improving the image contrast by applying morphological transformation, and then using the window based adaptive threshold and integral projection to segment the teeth and separate the upper and lower jaw. Nakintorn Pattanachai et al. [2] in 2012,proposed a method for tooth detection using Hu’s moment invariants using two set of x-ray images of a patient. They have manually cropped the dental x-rays to segment each tooth and then classified the tooth using image invariants to match it to a unique patient. In 2008, Guang-Yuan Zhang et al. [12] proposed a real-time eye detection method using Support Vector Machine (SVM) with Hu’s moment invariants. They used the video frame capture to binarize and heuristic rules to screen the contour. Then used them to find the Region Of Interest (ROI). They calculated Hu’s moment invariants of ROI and used SVM model for classification. This method achieved an average successful classification rate of 92.8%. In 2010, Ungkam larujareet et al. [13] used an ‘iris-blob map’ as a new feature for iris identification. The iris texture is enhanced with the Difference of Gaussian (DoG). The iris feature was a map consisting of iris blobs’ bounding rectangles and Hu’s moment invariants of the detected blobs.

Methodology

A data base was constructed using the patient data from the patient records of Riyadh Elm University. The database contains periapical dental radiographs. This was done with ethical permission of the scientific board of Riyadh Elm University. The periapical x-rays were taken with Hellodent plus-Sirona/intraoral X-ray unit.Image segmentation utilizes a combination of two main methods: image enhancement using morphological filters and contour detection.The image is enhanced using adaptive histogram equalization.

TEnh ’(g) is greater than gamma for g in The image after enhancement has dark background pixels and bright bone and tooth filters. Top-hat and bottom-hat morphological filters are applied to this enhanced image. A top-hat filter performs enhances the white parts of the image i.e. teeth and bones and the bottom hat filter enhance the background. Top hat filtered image is added and bottom hat is subtracted to the contrast enhanced image to further enhance the image and highlight the region of interest. The post processed image is converted to binary image using a grayscale threshold obtained using the image histogram which clearly separates the background from the image. More morphological operations are performed on this image for separation of the background and bone pixels. ’clean’ , ’majoity’ , ’hbreak’ morphological fitters. Figure 2 shows the enhanced image ,top hat filtered result ,bottom hat filtered image and the enhanced image along with the top hat -bottom hat filtered result. Figure 3 depicts the binary conversion used. Figure 2: a) Enhanced image b)Top hat filter c)Bottom hat filter d)Enhanced image+top hat -bottom hat filter.Figure 3.Binary conversion used.The local morphological filters used and performed are the following functions on an 8 pixel neighborhood for every pixel clean: sets isolated pixels to 0

• hbreak: Removes H-connected pixels.
• majrity: Sets a pixel to 1 if five or more pixels in its 3-by-3 neighborhood are 1s; otherwise, it sets the pixel to 0.A connected component analysis is performed on this filtered image to identify different connected components from the image.

The segments which have more than 4000 pixels are experimentally identified as tooth and are stored as new images. This is shown in figure 4.Figure 4: Connected Component AnalysisUsing the segmented images obtained from the previous step tooth segments were extracted and manually labelled as molars and premolars. A total of 329 teeth (molar/premolars only) were segmented out of which 164 were molars and 165 premolars. These images were appropriately labelled.Histogram of oriented gradients were extracted from the images and used as a feature set for training linear SVM (Support Vector Machine). Support vector machine is a kind of learning model which look for the best decision boundary. This boundary is located between the two classes to classify. In support vector machines the decision boundary is chosen to be the one for which the margin is maximized. This margin is the shortest distance from the decision boundary and any train point [7]. The technique counts occurrences of gradient orientation in localized portions of an image.

2.1.2. Classification using Hu’s InvariantsThis paper proposes to extract feature by using Hu’s moment invariants [6]. Hu’s moment invariants can be classified as a shape descriptor which is used in computer vision.[8]. It is based on the theory of algebraic invariants and derives to seven invariants. The basic idea is to describe objects by a set of measurable quantities called invariants and its invariant features on image translation, scaling and rotation. Hu’s moments are shown in 5. Hu’s invariant features were used to train linear.

Results and DiscussionThe approach of Hu’s invariants is as follows in the figure 5, and a result after segmentation is shown in figure 6. The results from segmentation are good enough to semi automate the procedure. Figure 5: Hu’s Invarients.Figure 6: Segmentation Results The main problem in segmentation error is the poor quality of some x-ray images. In some cases when the x rays had very good contrast teeth were segmented with 100% accuracy[9].Hu’s invariants provided much higher accuracy compared to histogram of oriented gradients. Another advantage of using moment invariants is the fact that it utilizes only 7 features for training as compared to 81 in HOG. Data is manually split into training and testing set and average accuracy is reported. A comparison between the two approaches in summarized in 1. 100 training images were used in each case and 64 and 65 test images respectively were used for molars and premolars.

Conclusions

The segmentation accuracy obtained is nominal but not optimal for complete automation of the segmentation process. Table 1. Shows the results of comparison. Further, the results using Hu’s invariants are promising and this can be used for unique tooth identification. Simultaneous work on multiple x rays of the same patient and train this classifier on the extracted features to uniquely identify the patient from the information is under progress. This work forms the basis of forensic odontology which has a lot of potential in human identification and related works.This should clearly explain the main conclusions of the work highlighting its importance and relevance.AcknowledgmentsI sincerely acknowledge the consistent support received from Riyadh Elm University(REU), Riyadh, Saudi Arabia All acknowledgments (if any) should be included at the very end of the paper before the references and may include supporting grants, presentations, and so forth.