Lung Fields Segmentation Algorithm in Chest Radiography - Advances in Image and Graphics Technologies - page 142

the chest radiographs is 20482048 pixels, every pixel is 12 bits, and the scale of
every pixel is 0.1750.175 .

5.2 Parameters Setting

When 1) the number of initial lung boundary candidate points are 30; 2) the coeffi-
cient between gray and shape cost is 10; 3) Right lung: the search region of the
4-16, 27-39 boundary points are the 6060 square region centered at the point,
the remaining boundary points are 100100 square region; Left lung: the
search region of 4-16, 36-39 boundary points are the 6060 square region
centered at the point, the remaining boundary points are 8080 square region; 4)
six feature images are used to search the lung outline. After modifying the lung out-
line, the optimal segmentation performance would be obtained and the average over-
lap of lung segmentation can reach 90.26%. The average overlap is a standard for
measuring similarity between the final segmentation results and the initial marking
outline.

5.3 Experimental Result

Segmentation Results with Different Searching Regions. In test images, the candi-
date points with larger gray cost of each boundary points need to be determined be-
fore searching the optimal lung outline, which were from the search region corres-
ponding to each boundary point. Generally speaking, the gray distribution of different
positions in chest radiographs are different, therefore, we need to classify the search
region according to the different situations. The represent the i boundary point on
right lung and the represent the i boundary point on left lung. There are 42 points
both on left and right lung. Under the same conditions, the segmentation results with
different search regions were given in Table

Ⅰ .

Table 1. Segmentation Results with Different Search Regions

R1-R3,R17-
R26,R40-R42

L1-L3,L17-
L35,L40-L42

Others Overlap (%)

80 120 80 88.56

90 100 70 89.19

100 80 60 90.26

Segmentation Results under Different Conditions. The algorithm proposed in this
paper would generate different segmentation results under different conditions. In this