Lung Fields Segmentation Algorithm in Chest Radiography - Advances in Image and Graphics Technologies - page 141

3.3 Segmentation by Use of Dynamic Programming

For the boundary point i in test images, once the m candidate points with smaller gray
cost were selected, a nm gray cost matrix M was constructed. Therefore, searching
for the optimal outline thus translated in finding an optimal path through M. That
means selecting an element from each row of the matrix M to constitute the optimal
path [12]. In the process of finding the optimal path, the sum of gray and shape cost
must be the largest:

(4)

J ∑

Where γ is the coefficient between gray and shape cost. To make the gray and shape
cost played roughly the same role in searching the boundary points searching,
the γ need to be adjusted properly.

4 Lung Outline Correction Based on Active Shape Model

In some images, the initial position may be too far away from the actual lung outline.
When segmenting the lung outline by use of gray and shape similarity information,
the lung outline may be not covered by search region. Therefore we need to adjust the
lung outline, and improve the situation of some points getting stuck in a local opti-
mum to obtain the better segmentation results.
In this phase, the lung outline of first segmentation was modified by the Active
Shape Model. Firstly, we need to find the shape parameters to adjust the lung outline,
and then the new lung outline was obtained, where the shape parameters can be ob-
tained according to the ASM. Then, the degree of similarity of gray distribution in
gradient direction between the new boundary points in all feature images from test
sample and the boundary points from training sample must be considered. The larger
degree of similarity indicated that the possibility being the optimal lung outline is
great. These above two phases need to be carried out iteratively until obtain the most
optimal lung outline.

5 Experimental Results and Analysis

5.1 Experimental Data

In our study, the experimental data used herein is the public database from the Japa-
nese Society of Radiological Technology. It includes 247 Posterior-Anterior images,
of which 93 images are normal, of which 154 images with nodules. The size of