attern 
he bi- 
Fig.5. 
t it is 
. The 
i mid 
e can 
ig.6. 
inded 
iS the 
a bi- 
‘more 
back 
ining 
Thre 
ayer 
back 
ropa 
first 
are 
idden 
one, 
  
  
  
Si lin 1+ e” 
this gives a value between 0 and 1. 
4.3.2 New Method 
Kondo, Hiroshi 
(5) 
The outputs of NN correspond to 0.6 — 3.0mm ¢ rounded opacities by step 0.2mm. As for the opacities bigger than 
Output Error 
0.3 d 
0.2 
0.1 + 
0.0 | 
| | 1 te [771 { 
0 1710° .2°10* 3" 10° 
Fig.9 Output error-repeating times 
Characteristic curve 
più p12 5-14 ali 
NY 
  
0-0 r- r— +4 r-36 rn r—40 
. x Te « gn « 
4 
  
3mm¢, the ROI image is reduced into 1/5 scale by 
averaging. Hence the size of each output corresponds to 
3.0 — 15.0mm 4 . It means that the same NN can be 
utilized for detection of 3.0 ~ 15.0mm 4 . It is sufficient 
to detect a rounded opacities with the size until 15 
mm because the rounded opacities bigger than 15.0 
mm ¢ is very rare for the pneumoconiosis X-ray photos. 
In this paper the most popular opacities in 
pneumoconiosis Le., small rounded opacities are only 
detected. 
The convergence of the back propagation NN is made 
after 17321 repeating times. The repeating is basically 
done according to a maximum principal. Figure 9 shows 
the output error-repeating times Characteristic curve. 
This is one of 
the typical curve of the back propagation method. The 
input image to the NN is not of bi-level because of the 
averaging for the size bigger than 3mm¢ . Since the 
figure of the rounded opacities is quite different with a 
circle for making a bi-level image multi-level image is 
much better as a neural network input for the bigger size 
3mm¢ ~ 15mm¢. 
5. TRAINIG PATTERN 
Back propagation NN requires training patterns. Figure 
10 shows several training patterns for pneumoconiosis 
rounded opacities with several sizes. The last five 
elliptic figures are for unnecessary parts like vessel and 
rib shade. These are actually not rounded opacities but 
thin and long ones. Hence such training patterns are 
effective for reduction the false positive. The big 
rounded opacities are made by averaging 5 ~ 5 pixel and 
put it into one pixel. 
r mE A pico OS m 
if ^ 5 5 3 ‚A 
pp i . . 
t e * ^ * s as ‘ L \ 
* 5 + LP SS & 
€ * = " % + 
E y -* ® * 
A. ® s - e 
. e S 
* = e n - x 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 457