AUTOMATIC DATA EXTRACTION IN ODONTOLOGICAL X-RAY

IMAGING

Douglas E. M. de Oliveira

, Gilson A. Giraldi

, Luiz A. Pereira Neves

Adriana G. da Costa

and

Erika C. Kuchler

National Laboratory for Scientiﬁc Computing, Av. Getulio Vargas, 333, Petr

opolis, Brazil

State University of Santa Catarina, Vision Laboratory, S

ao Bento do Sul, Brazil

Federal University of Rio de Janeiro, RJ, Brazil

Keywords:

Thresholding, Mathematical Morphology, PCA, Feature Extraction.

Abstract:

Automating the process of analysis in dental x-ray images is receiving increased attention. In this process,

teeth segmentation from the radiographic images and feature extraction are essential steps. In this paper, we

propose an approach based on thresholding and mathematical morphology for teeth segmentation. First, a

thresholding technique is applied based on the image intensity histogram. Then, mathematical morphology

operators are used to improve the efﬁciency of the teeth segmentation. Finally, we perform the boundary

extraction and apply the Principal Component Analysis (PCA) to get the principal axes of the teeth and some

lengths along it that are useful for dentist diagnosis. The technique is promising and can be extended for other

applications in dental x-ray imaging.

1 INTRODUCTION

Automating the procedure of image analysis for x-ray

dental images is an important tool for diagnosis and

planning of dentistry procedures. From the viewpoint

of image processing, two problems are fundamental

in this process: segmentation and feature extraction.

From a practical point of view, segmentation is

the partition of an image into multiple regions (sets

of pixels) according to some criteria of homogeneity

of features such as color, shape, texture and spatial

relationship (Jain, 1989). These fundamental regions

are disjoint sets of pixels and their union compose

the original whole scene. Approaches in image seg-

mentation can be roughly classiﬁed in: (a) Contour

Based methods, like snakes and active shape models

(Suri et al., 2002; Kass et al., 1988); (b) Region based

techniques (Suri et al., 2005); (c) Global optimiza-

tion approaches (Pan, 1994); (d) Clustering methods,

like k-means, Fuzzy C-means, Hierarchical clustering

and EM (Zhu and Yuille, 1996); and (e) Thresholding

methods (Albuquerque et al., 2004).

Among these approaches, thresholding techniques

(compute a global threshold to distinguish objects

from their background) are simple for implementa-

tion, with low computational cost, been effective tools

to separate objects from their backgrounds (Sahoo

et al., 1988). These methods have been successfully

applied for document image analysis, scene process-

ing, quality inspection, and medical imaging. The

common approach to implement a thresholding tech-

nique is based on the image histogram by searching

for its local minima (valleys). Other possibility is to

search for a threshold value constrained to the max-

imization of some information measure or entropy,

like the classical Shannon or generalizations of it (Al-

buquerque et al., 2004). After performed the image

binarization through the obtained threshold, we can

apply mathematical morphology techniques in order

to improve the result (Rodrigues et al., 2006).

Once the geometry of the objects has been ex-

tracted we can proceed the feature extraction. For

instance, geometric features are of special interest in

this project. Contour-based features, like area and cir-

cularity, as well as anatomical features can be used for

information extraction and classiﬁcation (Rodrigues

et al., 2006; Jain and Chen, 2004).

In this paper, we propose an approach for teeth

segmentation and feature extraction which is based on

the following steps. First, a thresholding technique

is applied based on the image intensity histogram.

Then, mathematical morphology techniques are used

141

E. M. de Oliveira D., A. Giraldi G., A. Pereira Neves L., G. da Costa A. and C. Kuchler É. (2009).

AUTOMATIC DATA EXTRACTION IN ODONTOLOGICAL X-RAY IMAGING.

In Proceedings of the First International Conference on Computer Imaging Theory and Applications, pages 141-144

DOI: 10.5220/0001800601410144

 SciTePress

to complete the teeth segmentation. Next, we perform

the feature extraction. The teeth boundary is extracted

from the binarized image and the Principal Compo-

nent Analysis (PCA) (Fukunaga, 1990) is used to get

the principal axes (r) of the teeth. Next, we automat-

ically determine two measures along r: the crown-

body (CB) and root (R) lengths. The former is ob-

tained through the distance from the deepest pit to the

furcation. The latter is the distance from the furca-

tion to the root apex. Finally, we compute the ration

CB/R.

The results show that the technique is promising

and can be extended for other applications in dental

x-ray imaging.

2 PROPOSED METHOD

In this section we describe the main steps of our tech-

nique for segmentation and information extraction in

dental x-ray images. The segmentation step is based

on a thresholding method and mathematical morphol-

ogy operators. Information extraction is performed

by contour detection and PCA. The Figure 1 pictures

a generic teeth boundary and the target features: prin-

cipal axes (r), deepest pit (C), furcation (B) and root

apex (R).

Figure 1: Teeth features: principal axes (r), deepest pit (C),

furcation (B) and root apex (R).

In the section 2.1 we describe the pipeline for

boundary extration and in section 2.2 the information

extraction step is presented.

2.1 Mathematical Morphology and

Boundary Extraction

A typical histogram for the dental images is pic-

tured on Figure 2. We have observed that the second

and third local maxima gives the intensity range that

roughly covers the structure of interest. For instance,

Figure 2 shows the histogram for Figure 3.a and the

Figure 3.b shows the thresholding result.

Figure 2: Typical histogram for the data base images.

We observe some undersegmentation in this ﬁg-

ure. Then, we invert the image and starts the pro-

cess to generate a mask in order to discard wrong pix-

els. A simple search method is used to get the mask

(Figure 3.d). This procedure is done using gray scale

histogram thresholding. Thus, it is possible to iden-

tify the tooth crowns and remove the main tooth. So,

a XOR operation is performed with the inverted im-

age, giving the result pictured on Figure 3.e. Finally,

an opening operation extracts the boundary pixels, as

shown in Figure 3.f. This operation is done using

cross structuring element with a single iteration.

(a) (b)

(e) (f)

Figure 3: Morphological Chain: (a) Original image. (b)

Thresholding result. (c) Inverted image. (d) Mask. (e) XOR

between Figures 3.c and 3.d. (f) Boundary pixels.

IMAGAPP 2009 - International Conference on Imaging Theory and Applications

142

2.2 Information Extraction

Once performed the boundary extraction, we must

consider geometric features. The main axes of the

teeth is the ﬁrst target in this step. As already known

in the literature (Fukunaga, 1990), it can be obtained

by the Principal Component Analysis (PCA), also

called Karhunen-Loeve, or KL method (Jain, 1989).

Thus, let us suppose that the data consists of N tuples

or data vectors, from a n-dimensional space. Then,

PCA searches for k n-dimensional orthonormal vec-

tors that can best be used to represent the data, where

k ≤ n, in the sense of data compression. Figure 4 pic-

ture this idea using a bidimensional representation. If

we suppose that the data points S =

{

,...,u

}

are distributed over the teeth boundary, it follows that

the coordinate system



X,Y



, shown in Figure 4, is

more suitable for representing the data set. The PCA

technique gives this coordinate system through the

following steps (Fukunaga, 1990):

1. Compute the centroid of the data set:

∑

i=1

. (1)

2. Subtract the centroid C

from the data points:

= u

− C

,i = 1, 2, · · ·, N. (2)

3. Compute the covariance matrix R :

R =

∑

i=1

∗T

. (3)

4. Find the matrix

Φ = [Φ

,Φ

,· · ·,Φ

n−1

,Φ

] (4)

where Φ

is the i-th eigenvector of R, sorted in de-

creasing order of the corresponding eigenvalues.

The columns of Φ gives the new basis vectors.

The operation:

= Φ

, i = 1, 2, ..., n, (5)

computes the data representation in the new basis.

Therefore, we take the obtained curve (boundary), ap-

ply steps (1)-(4) to get Φ, a 2 × 2 matrix in our case

(n = 2). Then, Φ

is taken as the principal axes of the

teeth (Figure 4). Expression (5) is applied for chang-

ing coordinates in order to simplify the searching for

the deepest pit, the furcation and the root apex (Figure

1).

Figure 4: (a)Original dataset. (b) Extraction of the principal

component.

3 DISCUSSION

We have a data base composed by 150 panoramic ra-

diographs, with resolution of 300 dpi, in gray scale

that will be used to test the method. The obtained re-

sults will be compared with manual evaluations per-

formed by the radiologists of our team. So, some

statistics are going to be generated in order to quantify

the precision of the proposed method.

Region based and contour based approaches could

be also considered to perform the segmentation tasks.

However, region based methods depends on some

characterization of the intensity patterns inside the re-

gion of interest (Suri et al., 2005; Zhu and Yuille,

1996). It is not a simple task due to inhomogeneities

of the intensity ﬁeld as well as texture patterns in-

side the image, as we can observe in Figure 3.a. On

the other hand, contour based approaches, like snakes

or level set models (Shah et al., 2006; Keyhaninejad

et al., 2006), must be properly parameterized in or-

der to converge to the desired boundary without stop-

ping on local minima. By considering that the teeth

boundary follows some pattern, a deformable model

incorporating shape information (shape model) could

be more efﬁcient than a traditional (free) snake model

(Buchaillard et al., 2007).

4 CONCLUSIONS

The proposed method is a combination of mathe-

matical concepts in morphology and shape analysis

AUTOMATIC DATA EXTRACTION IN ODONTOLOGICAL X-RAY IMAGING

143

(PCA), as well as algorithms to automatically com-

pute the ratio CB/BR in x-ray images. The computer

implementation has been developed using Object Ori-

ented best practices.

REFERENCES

Albuquerque, M. P., Albuquerque, M. P., Esquef, I., and

Mello, A. (2004). Image thresholding using tsallis en-

tropy. Pattern Recognition Letters, 25:1059–1065.

Buchaillard, S. I., Ong, S. H., Payan, Y., and Foong, K.

(2007). 3d statistical models for tooth surface recon-

struction. Comput. Biol. Med., 37(10):1461–1471.

Fukunaga, K. (1990). Introduction to statistical patterns

recognition. Academic Press, New York., 18(8):831–

836.

Jain, A. K. (1989). Fundamentals of Digital Image Process-

ing. Prentice-Hall, Inc.

Jain, A. K. and Chen, H. (2004). Matching of dental x-ray

images for human identiﬁcation. Pattern Recognition,

37(7):1519–1532.

Kass, M., Witkin, A., and Terzopoulos, D. (1988). Snakes:

Active contour models. International Journal of Com-

puter Vision, 1(4):321–331.

Keyhaninejad, S., Zorooﬁ, R. A., Setarehdan, S. K., and

Shirani, G. (2006). Automated segmentation of teeth

in multi-slice ct images. In Visual Information Engi-

neering, pages 339–344.

Pan, H. (1994). Two-level global optimization for image

segmentation. ISPRS journal of photogrammetry and

remote sensing, 49(2):21–32.

Rodrigues, P. S., Giraldi, G. A., Chang, R. F., and Suri,

J. (2006). Non-extensive entropy for cad systems of

breast cancer images. In Society, I. C., editor, Pro-

ceedings of Brazilian Simposium on Computer Graph-

ics and Image Processing, pages 121–128, Belo Hori-

zonte, Brazil.

Sahoo, P. K., Soltani, S., and Wong, A. K. C. (1988). A sur-

vey of thresholding techniques. Comput. Vis. Graph-

ics Image Process, 41:233–260.

Shah, S., Abaza, A., Ross, A., and Ammar, H. (2006). Auto-

matic tooth segmentation using active contour without

edges. In Proc. of the Biom. Consortium Conference,

pages 1–6.

Suri, J. S., Liu, K., Singh, S., Laxminarayan, S., Zeng,

X., and Reden, L. (2002). Shape recovery algorithms

using level sets in 2-d/3-d medical imagery: a state-

of-the-art review. IEEE Transactions on Information

Technology in Biomedicine, 6(1):8–28.

Suri, J. S., Wilson, D., and Laxminarayan, S. (2005). Hand-

book of Biomedical Image Analysis: Volume 3: Reg-

istration Models (International Topics in Biomedical

Engineering). Springer-Verlag New York, Inc.

Zhu, S. C. and Yuille, A. (1996). Region competition:

Unifying snakes, region growing, and bayes/mdl for

multibang image segmentation. IEEE Trans. Pattern

Anal. Mach. Intell., 18(9):884–900.

IMAGAPP 2009 - International Conference on Imaging Theory and Applications

144