Kondo, Hiroshi 
  
Fig.12 Simulation example (3, r) type ROI 
6. SIMULATION 
Figures 11 and 12 are simulation examples. Figure 11 is type 1 (weak) with many p-size rounded opacities. The circle 
which wraps the rounded opacities is a marking for the detection. It is shown that many opacities are well detected. Of 
course, Q-size and r-size of the rounded opacities are also there. But the major size is P-size, so we call this ROI (1, P). 
Figure 12 shows a detecting result with (3, r) type. Many big rounded opacities are detected. But when the figure of the 
opacity is not rounded our NN may not catch such an opacity. One solving approach of this problem is to average and to 
reduce the size. After detecting such opacities we calculate the densities; 
N, 
  
  
D, = — 6 
SUN (6) 
N 
D, = — 7 
4 NM (7) 
where N, and N, is the number area of rounded opacities respectively. N, is the back ground all area. D, is called 
the number density and D, is called the area a density. The classification is done in comparison with the densities of 
ILO standard pneumoconiosis X-ray photo. 
7. CONCLUSION 
We have presented a new automatic diagnosis for pneumoconiosis radiographs using neural network. Rounded opacities 
are caught by NN from tiny to big one. The detecting rate is higher than the several traditional methods. As a pre- 
processing for using NN we have developed a moving normalization method which is very effective for getting a high 
contrast image of a chest X-ray photo. It is important to modify an input suitably for NN. From upper left side to lower 
right side of the ROI the input scan is performed successively by one pixel with 30 ^ 30 pixel size. 
Although NN has a merit in real time processing, our proposed system is not a real time one because it takes a little bit 
much time for pre-processing and so many time repeating of NN processing. Now the processing time is about 5 
minutes for our personal computer (500MHzCPU). 5 minute are not so quick but it seems not to be unrealistic. 
As a left problem is to check a large amount of the chest X-rays and certify the high consistency with the doctor's 
opinion. 
REFERENCES 
Katsuragawa S., Doi K., Macmabon H., Nakamori N., Sasaki Y., and Fennelsy J., Quantitative computer-aided Analysis 
of lung texture in Chest Radiographs, 1988 RSNA annual meeting of the kurt Rossmann Laboratories for Radiologic 
Image Research, Dept. of Radiology. Unv. of Chicago Radio Graphics vol 10, PP257-269. 
Savol.M.A., Li.C.C., and Hoy.R.J, computer-Aided Recognition of Small Rounded Pneumoconiosis Opacities in Chest 
X-rays, 1980 IEEE Trans. on PAMLVOL.PAMI-2, No.5, PP479-482, Morishita J., DoiK., Katsuragawa S., Monnier- 
Cholley L., and MacMahon H., Computer-aided diagnosis for Interstitial infiltrates in chest radiographs: Optical-density 
dependence of texture measures, 1995 Am. Assoc. phys. Med., vol.22, PP1515-1523 
Sasaki Y., Katsuragawa S., and Yanagisawa T., Quantitative Analysis of Pneumoconiosis in Standard Chest 
Radiographs, 1992 Dept. of Radiology, vol.52 — 10, PP1385-1393 
Chen X., Hasegawa J., and Toriwaki J., Recognition of Small Rounded Opacities for Quantitative Diagnosis of 
Pneumoconiosis Radiographs, 1989 IEICE Trans. on D-II vol.J72-D-II, No.6, PP944-953 
  
458 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 
  
T A^