ary data 
it and it 
at many 
very far 
pects of 
h video 
Travail 
  
Kondo, Hiroshi 
COMPUTER AIDED DIAGNOSIS FOR PNEUMOCONIOSIS 
RADIOGRAPHS’ USING NEURAL NETWORK 
Hiroshi KONDO, Lifeng ZHANG, and Takaharu KODA 
Kyushu Institute of Technology JAPAN 
kondou@ele.kyutech.ac.jp 
KEY WORDS: X-ray photo, Pneumoconiosis, Neural Network. 
ABSTRACT 
Computer aided diagnosis for pneumoconiosis using neural network is presented. The rounded opacities in the 
pneumoconiosis X-ray photo are picked up by a back propagation (BP) — neural network with several typical training 
patterns. The training patterns from 0.6 mm@ to 30 mm¢ are made by simple circles. The neck problem for an 
automatic pneumoconiosis diagnosis has been to reject the unnecessary part like ribs and vessel's shades. In this paper 
such an unnecessary part is rejected well by adding several output neurons for own presenting neural network . These 
neurons are used only for picking unnecessary parts up. The input for the neural network is 30 ^ 30 pixel image which 
is quarried succeedingly from the bi-level ROI (region of interest) image with the size 500” 500 pixel. The new 
technique called moving normalization is developed here in order to made an appropriate bi-level ROI image. The total 
evaluation is done from the size and figure categorization and density categorization. many simulation examples show 
that the proposed method gives much reliable results than traditional ones. 
1. INTRODUCTION 
In recent years computer-aided diagnosis is very popular in a medical field. The aim of computer-aided diagnosis is to 
alert the radiologist by indicating potential lesions and/or providing quantitative information as a second opinion. Since 
the middle of 1980, a number of computerized schemes for computer-aided diagnosis have been developed for chest 
radiography, mammography, angiography, and bone radiography. Especially in chest radiography, many computerized 
schemes have been applied to the detection and classification of pneumoconiosis, because the quantitative analysis of it 
has been required from the view point of workmen's accident compensation insurance. Pneumoconiosis is a lung 
disease caused by, for example, the long term inhalation of coal dust and the local tissue reaction to the accumulated 
dust particles. The first radiological symptom in the development of simple pneumoconiosis is the appearance of small 
opacities, either rounded or somewhat irregular, in the chest X-rays. According to the profusion of small opacities, 
categories 0-3 have been established to indicate the severity of the disease where category 0 means normal case and 
category 3 means very numerous small opacities. The early studies of computer pneumoconiosis analysis have been 
done by several groups (Morishita, Yanagisawa, Katsuragauia, Doi, MacMabon, Nakamori, Sasaki, and Fennessy. In 
their studies the texture analysis for the X-ray photos has been taken. Recently, however, the study trends toward the 
detection of the small rounded opacity itself because of the extremely development of the computer hard and software. 
They have utilized a special filtering for dropping off the unnecessary part like rib shade in the X-rays. The performance 
of the filtering is not satisfied sometimes due to the vagueness of the X-rays. In this paper the neural network is 
introduced to pick the rounded opacities up from the X-ray photo with no filtering. A neural network is powerful for 
pattern matching. Here a back propagation neural network with three layers is used. 
2. PNEUMOCONIOSIS 
Pneumoconiosis is one of the serious lung occupational disease. Hence the diagnosis result of a medical doctor gives a 
big implication for the workmen's accident compensation insurance. Even such doctor's diagnosis results, however, are 
not often consistent with each other. For this reason it has been required that the quantitative analysis of 
pneumoconiosis is established. According to the classification scheme of the International Labor Office (ILO), there are 
two kind of categories : one is number and area density classification and the other is size-figure one. The former one 
has three ranks from 0 to 3, where 0 means normal case and rank 3 means very serious case. The size-figure 
classification has also three types as P, Q, and R, where P means the equivalent diameter d of the opacity is 
noooooooo0000000000000less than or equal to 1.5 mm, Q means 1.5 mm < des 3.0 mm, and R means 3.0 
< de 10.0 mm. Figure 1 shows allnormalized X-ray photo with 3000“ 3000” 8 bit. The normalization is made by 
setting the minimum gray level 0 (black) and the maximum value 255 (white). The other gray levels are transformed 
linearly between the above two values. Figure 1 is the photo for (3,P) category. Usually in the analysis of a 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 453