Analysis of Dental Images using Artificial Immune Systems Zhou Ji 1, Dipankar Dasgupta 1, Zhiling Yang 2 & Hongmei Teng 1 1: The University of Memphis.

Analysis of Dental Images using Artificial Immune Systems

Zhou Ji1, Dipankar Dasgupta1, Zhiling Yang2 & Hongmei Teng1

1: The University of Memphis2: Yinchuan Stomatological Hospital, China

CEC 2006. Vancouver, BC, Canada. July 17, 2006.

outline

Application background AIS method Data preprocessing Preliminary results

Application background

Occlusion: alignment of teeth/jaw Malocclusion

Abnormal occlusion Diagnosis using X-ray

malocclusion

Different types: I (normal bite), II (overbite), and III (underbite)

Mild or severe (functional)

lateral view skull X-ray

Normal case Example of malocclusion

conventional diagnosis methodAngle’s classification: angle ANB (3 in the picture)

N

A

B

AIS method

Negative selection algorithmsA detector set is generated from normal

samples and used to detect abnormal cases. One-class classification: classification

between two classes using samples from one class to train the systemanomaly detection

V-detector

A new negative selection algorithmMaximized detection size of detectorsCoverage estimate

Data preprocessing method-feature extraction

Using brightness distribution instead of traditional feature extraction (identification of entities or anatomical parts)

Binarization at multiple thresholds Description of each binary image with four

real numbers

remove artificial parts

binarization using multiple thresholds

choose thresholds

T0 = Vmax,

T1 = Vmax − (Vmax − Vmin)/n , ..., Tn-1 = Vmax − (n − 1)(Vmax − Vmin)/n ,

Thresholds are decided by the actual values of the image.

decide reference point

Binarized at the highest threshold

extract four featuresat each threshold (or for each binary image)

1. Horizontal displacementx = xwhite − x0,

(xwhite is the mean of x of white pixels)

1. Vertical displacementy = ywhite − y0,

(ywhite is the mean of y of white pixels)

1. Displacement distancer = mean of distances between white pixels to (x0, y0)

1. Area massA = total number of white pixel/width · height

Steps to represent an image as a real-valued vector

A grayscale image n binary images Each binary image 4 real values

1. Clean up 2. Binarization3. Calculate 4 features4. Normalization

Result: a grayscale image is represented by a 4n-dimensional vector over [0,1]4n

Preliminary experiment results

compare with SVM

Using half of normal data to train

SVM result

V-detector result

Summary

V-detector shows some potentials. A novel feature extraction is proposed.

Key idea: general shapes instead of anatomical part Issues:

Other possible feature representations more normal data are desired.

Thank you!

Questions?