MVA'98 IAPR Workshop on Machine Vision Applications, Nov. 17-19, 1998. Makuhari, Chiba, Japan
Automatic Judgement of Spinal Deformity from Moire Images
Employing Asymmetry of Local Centroids Location
Kazufumi Ishida t , Seiji Ishikawa ,
Yoshinori Ohtsuka $
t Department of Mechanical and Control Engineering
Kyushu Institute of Technology
$ National Sanatorium Chiba Higashi Hospital
A technique is described for classifying abnormal
cases and normal cases in automatic spinal deformity
analysis by computer based on moire topographic
images of human backs. Displacement of local centroids
is evaluated statistically between the left-hand side and
the right-hand side of the moire images. The technique
was applied to real subjects images in order to draw a
distinction between 60 normal and 60 abnormal cases.
According to the leave-out method, the entire data was
separated into three sets. The linear discriminant
function based on the Mahalanobis distance was defined
on the 2-D feature space employing one of the data sets
containing 40 moire images and classified 80 images in
the remaining two sets. The average classification rate
Spinal deformity is a serious disease for teenagers
and screening has been performed at primary schools
and secondary schools in Kanto, Japan, in order to
realize early check of the illness. For the screening,
moire topographic images of human backs are
conveniently employed, since 3-D asymmetry of a back
caused by spinal deformity is easily observed as an
asymmetric 2-D moire pattern. In practice, orthopaedists
'f Address: Sensuicho 1-1. Tobata, Kitakyushu 804-8550. Japan.
Enuil: firstname.lastname@example.org, ishikawaBis.cntl.kyutech.ac.jp
$Address: Nitonacho 673, Chuou. Chiba 260-0801, Japan.
inspect spinal deformity visually by the moire images
before they proceed to X-ray check.
This inspection is tough work on account of a large
number of moire images yielded by the screening at
schools and automating the inspection by computer
image processing has strongly been requested by
doctors. Since moire imaging is much safer to a subject
than X-ray imaging and it is also easily obtainable by a
moire camera which is commercially available, every
researcher tries to analyze a moire image of a human
back in order to realize automatic diagnosis of spinal
deformity by computer. Reported techniques to date[l-
31 perform 3-D recovery of the undulation of the back
and try to evaluate its geometrical asymmetry. The
result is that they all suffer from their complicated
image processing techniques and impractical processing
time. Nothing worth discussing has been achieved for
the automatic diagnosis of spinal deformity to date.
Unlike these techniques based on 3-D geometric
analysis, the technique proposed in this paper simulates
doctors' 2-D visual asymmetry inspection on a human
moire image. The technique realizes simpler analysis of
the moire image and therefore achieves much shorter
processing time than ever. Asymmetry of the moire
image of a human back is quantified by examining
displacement of local centroids defined on the moire
image. The technique is evaluated its performance
experimentally by real subjects' moire images. Finally
the result is discussed.
lr -1r -1r
2 Asymmetry Representation by Local
Gj (x, ,yl ).
employed for representing asymmetry of the moire
image in R.
The moire images employed are separated into two
groups, i.e., the training set S?(I, l k=1,2,. . ., M,} and
the test set S~(I,, I k=1,2 ,..., Ms}. The feature vectors
( p ,, a t ) obtained from I, E S, are plotted on the 2-D
feature space and the linear discriminant function L is
defined on it employing the Mahalanobis distance. The
images in S, are classified by the line L into normal
cases and abnormal cases of spinal deformity on the
feature space. The leave-out method is employed in the
classification to exclude biased data sampling.
of the values 0,
and the standard
(j=1,2, ..., N) are
2.1 Extracting the area of interest
Let us denote a moire image of a human back by I(x,
y). The origin 0 of the xy-coordinate system is located
at the lower left corner of the image. The ranges of the
coordinates are O l x l x , and OSySy,. The middle line
is defined in the first place on I(x, y). Since the moire
pattern of a human back usually exhibits asymmetry, a
potential symmetry axis is extracted from I(x, y) and
the axis is regarded as the middle line of the back. Let
the middle line be located at x=m.
The area of interest denoted by R is defined in the
second place on I(x, 3 in the following way. Image I(x,
y) is binarized and histogram of the binarized pixels
onto the x-axis is calculated. The locations having the
minimum frequency value are found within O l x l m
and mlx&x,, and two such locations, x=xo and x=x,,
that are the nearest to the middle line are chosen from
the respective ranges. The area of interest R is defined
as the inside area of I(x, y) discriminated by the two
vertical lines x=xo and x=x,. The area R excludes arms
of the subject and takes subject's physical dimensions
into account. See Fig.1 for the procedure, where (a) is a
binarized moire image, (b) histogram of the pixels
(doubled in the direction of the y-axis for easier
observation) and the discriminating two lines, and (c)
the area of interest R.
Evaluating asymmetry by local
Within the area of interest R, as shown in Fig.2, two
square regions are defined at symmetric locations with
respect to the middle line x=m. Length of its side a is
defined by a= min(x,-m, 111-x,]. Let us denote the square
regions of the left-hand side and the right-hand side at
y=j by Rj and R; , respectively. Here j=1,2,. . ., N.
The center of gravity (or the centroid) of R! and R;
are denoted by G
respectively. The centroid G: (Y: , 7: ) is reflected with
respect to the middle line x=rn into the region Ri and
denoted by G; ( x : , l; ) . The Euclid distance denoted
is calculated between
- r - r
) and Gl(x,, yl ) ,
Fig.1. Extracting the area of interest: (a) Binarized
moire image; (b) histogram on the x-axis and two
discriminating lines; and (c) the area of interest R.
Fig.2. Area of interest R on the subject's back and the regions for specifying the centroids.
Fig.3. Examples of local centroid locations on the moire images: (a) A normal case; and (b) an abnormal case.
Table 1. Obtained classification rates.
3 Experimental Result
The moire images employed in the experiment
contain 60 normal cases and 60 abnormal cases. They
are respectively separated into three sets each of which
contains 20 images and denoted by S,, for normal and
S,, for abnormal (i=1,2,3). The set S, is defined by S,=
S,, U S,, . Chosen S, as a training set, i.e., S+,,
S, ti, k#i) are used as a test set. The area of interest R is
defined at 19 individual positions which are mutually 10
pixels apart vertically. Employed parameters are
therefore as follows: N=19, M,=40, M ~ 8 0 .
Experimental result is shown in Table 1, where
classification rates are given by percentage. Examples
of local centroids location are depicted in Fig.3. The
processing time of a single moire image is 12 seconds in
average by Sparc Station 20.
By the experiment employing real moire images of
human backs, 87.9% of classification rate was achieved
in average. To the best of our knowledge, this is the first
experimental report employing real as well as
substantial image data. The paper therefore holds its
significance in this respect. Although the achieved
classification rate is not yet enough to put the proposed
technique into practice, the result makes us expect
future automation of spinal deformity inspection by
One of the advantages of the present technique is that
feature extraction is simple and therefore computation
time is shorter than any other techniques reported to
date. The computation time can be improved further by
the employment of the latest PC.
On the other hand, the present technique is sensitive
to the location of the middle line, since the local
centroid of the left-hand side square region is reflected
with respect to the middle line in order to calculate the
displacement with the local centroid of the right-hand
side square region. Larger asymmetry in the moire
image might result in inexact position of the middle line.
On this issue, we are planning to settle the middle line
visually on the stage of taking a subject's moire image.
This may contribute to obtaining better results in the
Nine abnormal cases out of 60 were classified as
normal. This is a vital situation compared with the
reverse case that normal cases are classified as abnormal.
All of the nine cases are obviously visible of the
asymmetry of moire patterns and yet extracted local
centroids spread on the image almost in a symmetric
way. This may be because gray values distribution in the
square regions unfortunately operated symmetrically
when the centroids were calculated there. The size and
the location of the square regions need be elaborated
A technique was presented for analyzing spinal
deformity by computer based on moire images of
human backs. Asymmetry of local centroids were
statistically evaluated. The technique was examined by
120 real moire images and 87.9% of the classification
rate was achieved. The technique needs further
improvement to obtain higher classification rates.
[l] Idesawa, M., Yatagai, T., Soma, T.: "Scanning moire
method and automatic
shapes" , Appl. Opt., 16,2152 (1977).
 Barouche, M.: "A knowledge based system for
diagnosing spinal deformations: Moire pattern
analysis and interpretation" , Proc. 11 Int. Conf.
Pattern Recogn., 591 -594 (1 992).
 Ishikawa, S., Takagami, S., Kato, K., Otsuka, Y.
"Analyzing deformity of human backs based on
the 3-D topographic reconstruction from moire
images" , Proc. '95 Korea Automat. Control Conf.,
244-247 (1 995).
 Ishikawa, S., Kosaka, H., Kato,K., Otsuka,Y.: "A
method of analyzing a shape with potential
symmetry and its application to detecting spinal
deformity" , Comput. Vision, Virtual Reality,
Robotics in Med., 465-470, Springer (1995).
measurement of 3-D