The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
824 
2004, Jerebko et al., 2003, Kiraly et al., 2004 and Tarik et. al., 
2006. 
All the above mentioned CAD techniques show different levels 
of accuracy and indicate that future investigations are needed in 
order to obtain a robust technique for polyp detection. In this 
paper we propose two new algorithms for detecting and 
classifying polyp candidates. We also improve the experimental 
results by adding a morphological image cleaning algorithm 
introduced by Richard Alan Peters II, 1995. The presented 
polyp detection technique shows relatively high sensitivity for 
polyps larger than 5 mm. 
2. ALGORITHMS 
2.1 Noise Reduction 
CT images may be considered noisy due to low radiation dose 
requirements and other processing stages, image enhancement 
through noise reduction is a fundamental problem in image 
processing which leads to better looking images to the 
interpreters. Noise reduction is an image restoration problem in 
that it attempts to recover an underlying perfect image from a 
degraded copy. To meet this purpose, we apply the 
Morphological Image Cleaning (MIC) algorithm introduced by 
Richard Alan Peters II, 1995 since it is capable of preserving 
small features while removing noise and scanner artifacts and 
enhancing images. MIC smoothes the image in a number of 
size-bands by computing the pixel wise average of the open- 
close and the close-open of image with disk shaped structuring 
elements of different diameters (OCCO filter). Let I be the 
original image and Z the corresponding structuring element: 
OCCO (/;Z) = y((/°Z)«Z)+|((/*Z)oZ) i 1 ) 
After that, it subtracts these bands out of its previous image to 
create residuals. Let Sj be the result of smoothing I with filters 
of size dj, then Dj is the j'th residual image: 
Dj=S J -S J _ l (2) 
These outputs are signed images. Positive residuals are called 
top hat images and negative ones are called bot hat. Then, it 
segments the residuals into features and noise regions by 
cleaning up top hat and bot hat images. And finally, adds the 
features back to the smoothed version of the original image 
under the following order: bright features are put back in 
smoothed image by adding to it the sum of all the cleaned-up 
top hats and the dark features are put back by subtracting from 
it the sum of all the cleaned-up bot hats. Ideally, this results in 
an image whose edges and other features are as sharp as the 
original yet has smooth regions between them. 
2.2 Segmentation 
The segmentation algorithm includes two separate steps; 
first, extracts the colonic wall applying a region growing 
algorithm (Gonzalez et al., 1993). This idea comes from the fact 
that CT images show high intensity difference between air and 
tissue. Therefore air insufflated colon lumen can be segmented 
applying a simple region growing. In some situations that the 
colon is collapsed due to either residual materials and water or 
insufficient insufflations, we are obliged to use multiple seed 
points for each part. The seeded region growing is done at the 
fixed intensity threshold of -800HU; proposed by Sadleir et al., 
2002. 
We assume the diagnostically region of interest as about five 
pixels outside the colon wall so that no information is lost. Thus 
in the second step we apply a morphological grassfire operation 
proposed by Gokturk et al., 2001 on the image. This algorithm 
finds points that are at equal distance from a layer of points (the 
extracted colon wall pixels).This determines the colon wall 
region within a 5 pixel margin (five pixels outside and five 
pixels inside). But we just need the outside pixels since the 
inside layer may cover the surface candidates. Therefore we can 
mark and subtract the inner added pixels from the result gotten 
before performing grassfire operation. 
2.3 Feature Extraction 
Having colon wall segmented we have to detect polyps on the 
colon surface. Polyp detection algorithms are under 
development to help diagnosis processes. These approaches 
include use of overlapping surface normals (Paik, 2001; Paik et 
al., 2004), curvatures (Summers et al., 2000; Yoshida et al., 
2001), sphere model fitting (Gokturk et al. 2000), vector field 
analysis (Acar et al., 2002) and statistical classification 
techniques such as support vector machines (Gokturk et al., 
2001) or neural network (Jerebko, 2003). Here we present a 
novel Automatic Morphological Polyp Detection (AMPD) 
algorithm. This algorithm marks polyp candidates (potentially 
containing folds) on images and determines their boundaries as 
inputs to the final stage. 
Mathematical morphology is a theoretical model for digital 
images built upon lattice theory and topology. Various image 
processing techniques can be implemented by combining only a 
few simple morphological operations. AMPD algorithm begins 
by eroding the image with a small size (in this work 3) square 
structuring element to reduce very small brighter components 
on darker background and this will effect the image the same in 
all directions because of its symmetric structuring element. Let 
I be the image and H the structuring element. So the erosion of I 
by H is defined as: 
I®H = {x :(//),<=/} 
IQH is composed of points that when H is moved to these 
points, every point of H is contained in I. 
It then operates area opening process which is a filter removing 
the components with area smaller than a definable parameter, 
the connectivity is given by a structuring element. As polyps 
seem like branches connected to colon wall at a perpendicular 
orientation, they can be removed by this procedure considering 
a proper structuring element (SE) and area parameter. 
If I is the image, a the area parameter and B c the structuring 
element, then the area opening of I with respect to Cl and B c is 
defined as: 
I o (a) R = v I o B ( 4 ) 
BeB BCia