ottom
ds the
. The
n the
sition
rized
| done
of the
other
wa in
ligital
non-
ystem
or an
RMS
ity of
Xf the
jiosis
acity.
rcular
or rib
, And
back
tterns
inded
d. In
is an
sing.
st X-
Kondo, Hiroshi
( pm, Size, Location O t
Center Pixel=1 ocation Outpu
N 9
Yes
| Exclusive Deleting NN Input Image |
POR-Classification
END
Figure.3 Flow chart of the proposed method
Especihlly tiny area average value is quite different with that of another area. Hence only a simple normalization of the
ROI is|not necessarity sufficient for processing. We introduce here a new technique in order to get a high contrast image
and wg call it a moving normalization that is shown in the following section. From the result of the moving
normalization we can make a bi-level image by using a threshold value. The input of NN is taken from the bi-level
image ‘and its size is 30^ 30 pixel. If the center pixel of the taken input has 0 (black) then it is not needed to input
because only the center image of the input is the object image for NN. If the center pixel of the input has 1 (white) then
other isolate whites are deleted i. e., changing from 1 to 0 and take it as an input to NN. This processing is called here
an exclusive deleting one. It is important to delete the back ground because NN output reacts also for back ground of
the input. The detection of rounded opacities is the next step. This is the main topic of this paper and also shown in the
following section. Finally calculating the number density and area density we get a classification result in comparison
with these value of the ILO standard X-ray images.
4.1 Moving normalization
In the chest X-ray image usually the contrast is low but the local area gray level is often much different with the other
local area one. Hence the full picture normalization in the gray level gives the unsatisfactory results for our processing.
Here we employ special technique for the normalization.
er sn First, the tiny area R, with 32^ 32pixel is taken from
roa the ROI at left top corner. Let the minimum and the
nun 3| g maximum values of the Re f,. and f,, respectively the
512 -] = gray level f 6,j) of R, is changed as follows.
bd M F,
Y g 6J)- (f 67 fun p d (1)
where F,,, is the maximum value; if we employ 8bit for
Fig.4 Procecuurc Urmuvig nurinatzz tion the quantization then Far —255. The result g 6j) is
quantized from zero to 255. After transforming by using
Eq.(1), we leave the central pixel and dispose all other
pixels. And we repeat this procedure by shifting one
pixel to the left. The procedure is done from left to right
and from top to bottom. The marginal 15 pixels of the
ROI are also disposed. If we need such margins then it is
sufficient to take ROI wider by 15 pixels. Figure 5 is one
example of the moving normalization. Very high
contrast image is gotten in comparison with the original.
This image is much suitable for bi-leveling. In a moving
normalization we tried to take 8^ 8, 16^ 16, 32^ 32,
64" 64. And 32^32 is the best one of all
Í
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 455
