Page 1

MVA '96 IAPR Workshop on Machine Vision Applications. November. 12-14, 1996, Tokyo. Japan

Automatic Spinal Deformity Detection by Two Characteristic Axes

Hiroshi Ueno Seiji Ishikawa

Dept. of Electric, Electronic and Computer Eng.

Kyushu Institute of Technology

Dept. of Civil, Mechanical and Control Eng.

Kyushu Institute of Technology

Yoshinori Otsuka 3

Kiyoshi Kato

National Sanatorium Chiba Higashi Hospital

Dept. of Orthopedic Surgery

Dept. of Electric, Electronic and Computer Eng.

Kyushu Institute of Technology

Abstract

line of human and the other axis is the principal

axis. Difference of their gradients are employed for

judging normal or suspicious with respect to spinal

deformity. The experiment is performed employing

real moire images of children's backs and the result

is shown with discussion.

This paper proposes a technique for judging

spinal deformity from a moire image of a human

back. The middle line and the principal axes of the

back are extracted from the moire image and their

difference is numerically evaluated. For the

extraction of the middle line, the potential

symmetry analysis technique is employed, whereas

the principal axes are obtained from the moment of

inertia matrix defined on the moire image.

Experimental results are given and some issues are

discussed.

1. Introduction

Spinal deformity is a serious problem for

teenagers. Medical doctors inspect moire images of

their backs for primary screening in schools. If a

subject's spine is normal, the moire image is almost

symmetric with respect to the middle line of hislher

back. If a subject has spinal deformity, the moire

image is distorted asymmetrically and the degree of

asymmetry is evaluated visually by doctors. It is

actually crushing effect on medical doctors to

inspect a number of moire images. Therefore,

automating the screening is strongly expected by

medical doctors. There have been some studies on

this automation until now [l, 21 , but none of them

has yet been put into practice.

This paper proposes a technique for judging

possible spinal deformity by extracting two

characteristic axes from a subject's rear moire

image. One of the characteristic axes is the middle

Address: Sensuicho 1-1, Tobata, Kitakyushu 804, Japan.

E-mail: ueno@ishi.cornp.kyutech.ac.jp

* Address: Sensuicho 1-1, Tobata, Kitakyushu 804, Japan.

Address: Nitonacho, Chiba 280, Japan.

2. Characteristic Axes

2.1 Axis of Potential Symmetry

A shape is called to have potential symmetry, if it

makes us associate with a certain kind of symmetry

like a human hand. The shape which has potential

symmetry may provide intrinsic information of the

figure, if it is analyzed taking the original symmetry

type into account. This analysis is originally

realized in [3]. The technique is slightly modified

in this paper.

Let us denote a digital image by F(x,y),

(x,y) E R, and its mirror symmetric image by

Fr(x, y) , (x, y) E Rr . Here R is a region of F

and its reflected region is denoted by Rr . Image

Fr(x,~) is superposed onto ~ ( x ,

translation c(c,,c,) and rotation 8 to find the best

coincidence according to the formula

Y) by parallel

where S = n { ~ n

R'] and

Page 2

Note that the number of the elements of set A is

denoted by n(A). The following restriction should

be taken into account with respect to c, ,c,, and

0 [41;

e

2

c,, = cX tan - .

When the superposed position realizing D of

Eq.(l) is obtained, the normal bisecting the line

segment connecting G , the centroid of F , and Gr ,

the centroid of Fr, is defined as the axis of

potential symmetry. This axis is inclined by 812

with respect to the vertical line. For effective usage

of Eq.(l), the superposed area S needs be more than

a certain threshold value.

2.2 Principal Axis

The extreme value of the moment of inertia gives

the direction of the principal axis of the object

concerned. It coincides with the axis of symmetry,

if the object is exactly symmetric. However, if the

object has even a slight asymmetric part, the

principal axis will never agree with the original

symmetric axis.

Principal axes are calculated from moment of

inertia matrix I defined on an object (or on its part).

It is given by Eq.(3),

Here pm (p,q = 0,1,2,. ..) is the (p + q) th central

moment. By solving the eigen problem of I, the

principal axes are obtained as its eigenvectors.

set the middle line on the moire

image of a human back

I

specify two rectangle areas

for evaluating distortion

I

calculate principal axes

I

I

evaluate deformity indices

Fig.1 Flow chart of the procedure

In the first place, the axis of potential axial

symmetry is extracted as the middle line employing

the technique[3, 41 from a moire image of a

subject's back. Two principal axes are extracted

from upper and lower rectangle areas containing

shoulder blades and the waist, respectively (See

Fig.2). The principal axes are calculated from the

moment of inertia matrix given by Eq.(3) which is

defined by the gray values in each rectangle area.

Since the range of the gray values normally varies

with respective images, their distribution is

normalized so that the mean gray value is 0 and the

standard deviation 1. The rectangle area itself is

specified manually keeping the location symmetric

with respect to the middle line. Its size is , however,

specified by a user at the moment.

3. A Technique of Detecting Distortion

2 1 ',

The moire image of a human back is almost

symmetric with respect to the human middle line, if

hislher spine is normal. However, the moire stripes

are distorted asymmetrically if helshe has spinal

deformity. In order to evaluate this asymmetry in a

numerical way, the proposed technique employs

two characteristic axes obtained from the moire

image of a human back. The flow chart of the entire

procedure is shown in Fig.1.

Fig4 Two rectangle areas

Page 3

In order to discriminate normal cases from

suspicious cases, an index given by Eq.(4) is

calculated from the two characteristic axes;

cases. The entire classification rate then becomes

90%.

Here a, is the angle in degrees of the middle line,

and a, is the angle in degrees of the principal axis.

The values of ad at the upper rectangle area and

the lower rectangle area are denoted by aduppr and

adlnwer, respectively. They are called characteristic

angles.

4. Experimental Results

Experiment is performed employing the moire

images of junior-high school students' backs.

Photographs of the images are digitized into 256 by

256 pixels images with 256 gray levels by an image

scanner connected to a personal computer and they

are fed into a workstation (Sparc Stationlo) via

network lines. The proposed technique is corded

into a program by C language and it is implemented

on a workstation to analyze specified moire images.

Two of the experimental results are shown in

Fig.3. On the left-hand side, the original moire

image with the detected middle line is shown, while

on the right-hand side, the moire image with the

obtained principal axes and the values of ad are

given. Figure 3a is a normal case and Fig.3b is a

suspicious case which is referred to as an alternate

pattern. The average elapsed time for extracting the

two axes is approximately 2 minutes by Sparc

Station 10.

Forty data, 20 normal and 20 suspicious, are

employed in the experiment. A feature space of the

characteristic angles is shown in Fig.4. The

horizontal axis is the values of adupper,

vertical axis is the values of adlwe,. As shown in

Fig.4, the horizontal axis does not contribute to

differing normal cases from suspicious cases in this

experiment. Therefore the vertical axis is solely

employed in the experiment for separating the both

cases. The result is given in Table 1 where two

respective thresholds are used to divide the

scattered data into two classes on the ad,,,

In Table la, the threshold for adlowe, is chosen so

that all the suspicious cases are correctly classified.

Then the total classification rate is 80%. In Table

lb, on the other hand, the threshold is chosen to

achieve 100% of correct classification of normal

and the

axis.

ad,,,,,

ad,,,,

= 1.62

= -8.30

Fig3

Experimental results: (a) a normal pattern,

and (b) an alternate pattern.

A * -A

A suspicious

- 8 ,

."

--

Fig.4

Characteristic angles

Page 4

Table 1 Classification rates

separation among these suspicious cases is also

within the scope of this study.

threshold

1.98

suspicious

20( 100%)

normal

12(60%)

total

80%

(a)

5. Discussion

The proposed technique was examined its

performance by an experiment and 90% of correct

classification rate was achieved. The experiment is

still under way employing a number of data to find

the best threshold value for the classification.

Employment of two characteristic axes seem to

work well, since the gradient of the principal axis is

highly sensitive to asymmetric distortion of the

moire image concerned, while detection of the

middle line is likely to be stable because the moire

image analyzed is not very asymmetric. The

rectangle areas need careful choice, however, since

the gradient of the principal axis depends on its

location and size. One may need to specifL them

taking account of a subject's height, the girth of

hisher breast and that of hisher waist.

Although the variable aduppr was not used for

classification in the performed experiment, it

should be employed for hture experiments, since

the area containing shoulder blades are actually

taken into consideration in medical doctors' visual

inspection. Effective use of adupp, largely depends

on appropriate choice of the rectangle area near

shoulder blades.

In the performed experiment, classification

between normal and suspicious was a primary

concern. The suspicious cases are, however,

separated into five typical patterns and they are

likely to be recognized by the gradients of the

principal axes at the two rectangle areas. Automatic

6. Conclusion

A technique was proposed for detecting spinal

deformity automatically from moire images of

subjects' backs. Two characteristic axes were

employed, i.e., the middle line on a subject's back

approximated by the axis of potential symmetry and

the principal axes calculated from the moment of

inertia matrix. Their angular differences were

employed for discriminating suspicious cases from

normal cases and some experimental results were

shown. Automatic rectangle areas specification and

the employment of the pattern information of the

area surrounding the shoulder blades need to be

investigated. Separating possible spinal deformity

into individual cases also remains to be studied.

References

[I] Batouche, M. : "A knowledge based system for

diagnosing spinal deformations: Moire pattern

analysis and interpretation", Proc. I1 Int. Con$

. Putt. Recogn., 59 1 -594(1992).

[2] Ishikawa, S., Ueno, H., Kato, K., Otsuka, Y. :

"Automatic analysis of spinal deformity from

moire images", Proc. Int. Con$ Automat., 2 13-

2 16(1995).

[3] Takeda, T. Ishikawa, S.,

"Employing symmetric subsets for identifLing

asymmetry of human skulls", Proc. IAPR

Workshop on Machine Vision Applications,

338-341(1994).

[4] Ishikawa, S., Kosaka, H., Kato, K., Otsuka,

Y. : "A method of analyzing a shape with

potential symmetry and its application to

detecting spinal deformity", Computer Vision,

Virtual Reality and Robotics in Medicine, 465-

470, Springer(l995).

Kato, K. :