Lung Fields Segmentation Algorithm in Chest Radiography - Advances in Image and Graphics Technologies - page 139

2.2 Aligning the Training Set

In order to compare the points in the same position from different training images,
which need to be aligned to one another. The lung outlines of training images need to
be aligned closely by scaling, rotation and translation operations.

2.3 Establishing an Outline Model

After aligning the shape of lung outlines in the training set, Principle Component
Analysis was employed to determine statistical information of shape variations. Then
an outline model was built, which can improve the efficiency of the algorithm.
The average shape of all the training images was assumed to

. The
covariance matrix between the average shape and the training images after aligning
was defined as S

∑

∑

. The value of N was the number of fea-
ture images. The eigenvalues and eigenvectors of the covariance matrix were calcu-
lated by S . Then the eigenvalues need to be sorted. The
tors correspond to the k largest eigenvalues .
The t principal eigenvectors need to be selected to form a new spindle system .
Then any shape belongs to the shape domain can be approximated by an average
shape and weighted spindle system:

xp

(1)

Where is a matrix composed with t

tors, is the weight vector, ensured the percentage of target object
deformation determined by t eigenvalues accounting for the target object deformation
determined by all eigenvalue s is not less tha n V (V is general 0.98 ), and should be
limited to the condition 3 3 .

3 The Segmentation of Lung Fields Based on Gray and Shape

Cost

In this paper, we used gray and shape cost simultaneously to make the gray and shape
information of lung outline similar to the training images.

3.1 Feature Image

The variation of the image gray is prominent in feature image, therefore, the feature
image was used to obtain the candidate points and the gray cost of each candidate
point. Before obtaining the feature image of the original image, the Gaussian filter
was required to reduce the effect of noise on segmentation results.