![](bg2.png)
The structure of this presentation is organized as
follows: in section2, we introduce why and how the
ROI is selected. In section3, the ROI-based LUT
construction method is presented. Section4 reports
the performance comparison result of the current
method and of baseline method. Finally, the
conclusion is drawn in Section 5.
2 AUTOMATED ROI SELECTION
After doing a background segmentation based on
ICU’s image histogram and difference histogram
(Kuhn, 1999), we get an appropriate threshold for
removing the background. A region-labeling
operation can be done to prevent over-segmentation.
Then we perform the automated ROI selection.
In ICU images, the position of the parts
necessary for the clinicians’ diagnosis varies. In
some cased, they will only take up a little part of the
image. The basic principle of automated ROI
selection is to identify the RIO in each image
automatically and adjust the image contrast values
within the ROI to a suitable range for each image, so
that comparison of one image to another is feasible.
Figure 1: Automated region of interest selection; this is an
example of selecting similar regions of interest for two
images of the same patient.
ROI identification located key features (lung line,
spine line) in an image and allows the correlation of
two or more images accordingly. Figures1(c) and (f)
show two chest X-ray images of the same patient
with two automated regions of interest (ROI)
selected.
First we use a median filter to resize the image,
then a Gaussian filter for noise removal. Next, the
locations of the spine line and lung line are detected
(Amit and Mark, 2005). Fig.1 (a) and (d) show the
spine and lung line detection. We search for the
highest/lowest mean column value row by row.
Connecting these points, we validate the lung line
step (Fig.1 (b) and (e)), and combine and validate
similar lung line parts based on gray-level and
position.
With the approximate lung line and spine line
determined, a spine-line-fitting step can be executed.
This is performed by doing an iterative of the spine-
line-fitting step. We search all the rows between the
top and bottom of the lung lines. We then choose the
fitting result that has the lower mean residual form
these two. We then can get a trapezoid ROI for all
the images of the same patient based on the spine
line and the distance of the spine line to the lung
line.
3 ROI-BASED LUT
CONSTRUCTION
Once one or more ROIs have been identified, we can
do the ROI-based LUT construction step.
First we identify the primary area o the image
from the histogram data that is related only to the
ROI. Points lp and rp represent left and right points,
respectively, of the histogram data that is from the
main range (2.5%-95%) in the ROI. After that, for
each image, left points lp1 and lp2, and right points
rp1 and rp2, are obtained. The goal of next few steps
is to remap left points lp1 and lp2, and right points
rp1 and rp2, to the corresponding points A1 and A2,
in order to form consistent images in the output
images.
Figure 2: Lookup table construction.
Figure 2 shows how various portions of the image
are remapped for consistent rendering. We can map
the right point rp, obtained from the ROI of each
input image, to the same value Ar in the output
image that has been determined for the same patient.
However, to accommodate the difference in patient
position between two images of same patient, we
proposed to use Ar for each image. Here, the
A IMAGE PROCESSING METHOD FOR COMPARISON OF MULTIPLE RADIOGRAPHS
229