Kondo, Hiroshi
pneumoconiosis chest X-ray image the original image like Fig 1 is divided by three blocks from the top to the bottom
shown in Fig 1. We call them high lung field, middle lung field, and low lung field respectively. From these fields the
region of interest (ROI) is Quarried for the analysis (See Fig.1). The size of the ROI is 512^ 512 pixel here. The
opacity figures are almost rounded but sometimes irregular as a complex figure. The rounded opacities in the
pneumoconiosis chest X-ray image appear so often especially in the high and middle lung fields because of the position
near bronchial tubes. Figure 2 shows an example of ROIs. This is the right high lung field from (Q,2) categorized
image. From such ROI image the rounded opacities must be detected. The field evaluation for the classification is done
by calculating the number density and the area density of the rounded opacities, and by comparing those values of the
ILO standard images.
Fig. 8bit Fig. 2]
3. TRADITIONAL METHOD
There are two types of approach for pneumoconiosis radiographs analysis. One is in lung texture analysis and the other
is in a number and area density approach. The texture approach has done by Sasaki, Katsuragawa, and Yanagisawa in
1992. They use physical texture measures obtained from an analysis of the power spectrum of lung textures in digital
chest radiographs. First they select the ROI’s for inter-rib spaces from the original image. After correction of non-
uniform background trend in ROI's two dimensional Fourier transform is performed for filtering by visual system
response. Finally the root mean square variation and the first moment of the power spectrum are calculated for an
evaluation of the pneumoconiosis. The final evaluation is made by a graphical manner with a horizontal axis RMS
variation and a vertical axis the first moment of power spectrum. This method gives us a relatively easy tractability of
categorizing the chest X-ray image. The exactness of the evaluation results, however, is not so reliable because of the
dependence on the ROI size and the trend correction. On the other hand as a density approach for pneumoconiosis
analysis Chen, Hasegawa, and Toriwaki's paper (1989) is excellent. They use a method to detect each rounded opacity.
From the original chest X-ray image, first they extract the opacity candidates by using smoothing filter, 2D circular
difference filter, and a threshold value processing. Then dropping the unnecessary parts like blood vessels and/or rib
shades, they detect rounded opacities. The number density is calculated from the number of the detected opacities. And
the area density is also gotten from the area of the property for a pattern matching. Here we utilize a back
opacities. Finally the evaluation is made from the above propagation NN with three layers. The training patterns
results. They have gotten 71% correspondence rate to are circular figures with various radius like the rounded
doctors evaluation results. And the correspondence rate opacities.
of the detected rounded opacities is 72% (true positive).
The others (28%) are false positive or false negative. It Figure 3 shows a flow chart of the proposed method. In
looks very hard to detect all true opacities for this this figure ROI means a region of interest which is an
method because the 2D circular difference filter is input image with 512^ 512 pixel for this processing.
sensitive not only to rounded opacities but also to rib and Next step is moving normalization of the ROI. Chest X-
vessels shades, although the procedure of this method is ray image has often a big variance in its gray levels.
relatively simple.
«Flow chart»
4. NEW METHOD USING NEURAL NATWORK =
| Moving Normalization |
The correspondence rate of the evaluation results =
depends upon the exactness of a rounded opacity
detection. It is the most important part of the -
pneumoconiosis analysis to detect each rounded
opacities. In this paper a neural network is utilized for >
detecting a rounded opacity. A neural network is mn :
D HOne Pixel i
abbreviated to NN here. An NN has an excellent Shift
No D
454 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.
| NN Input Image (30^ 30 pixel)
