The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
826 
Patterns with three on pixels (area = 7/8), Patterns with all four 
on pixels (area = 1) (Pratt et. al., 1991). 
After construction of polyp templates, we first detect spherical 
shaped polyps. The window, containing the circle pattern, is 
moved across the entire image. Whenever, the circle pattern is 
located inside a mask of 'on' pixels, the algorithm computes 
Euclidean distances between the points on the template and the 
points on the lesion boundary, at predefined intervals (D). 
Then, for each boundary a 360/D-length vector of distances is 
formed. The standard deviation is calculated for each vector. In 
the case the test circle is centered within a polyp, distances 
between the circle perimeter and lesion boundary points 
become nearly equal. Therefore the standard deviation of the 
corresponding distance vector approaches a small value. Thus, 
spherical polyps are distinguished from other lesions using an 
experimental threshold T on the standard deviations. Any 
cluster having the standard deviation smaller than T is 
considered to be polyp. 
In order to detect ellipsoidal polyps, the template with ellipsoid 
pattern is moved across the image containing remained clusters. 
Whenever a closed boundary is located inside the local window, 
the algorithm computes all Euclidean distances between any 
two pixels located on the lesion boundary and finds maximum 
distance. The direction having the maximum distance is 
assumed to be the major axis of potential ellipse. Orientation of 
such direction is calculated and the pattern ellipse is then 
rotated to get the same orientation. Then, it is possible to 
calculate the Euclidean distances between the points on the 
template and the points on the lesion boundary at symmetric 
intervals. Next, a histogram of number of pixels with a given 
distance versus distance values can be constructed for each 
cluster. If lesion is an elliptical shaped polyp, then the distances 
follow the symmetry property of ellipse. Thus, the standard 
deviation for its corresponding histogram takes a small value. 
The same as the first step, classifying procedure is done using a 
threshold T' on the standard deviations. Any cluster with the 
standard deviation smaller than T' is classified as polyp. 
3. RESULTS 
In this section, we discuss the results that we obtained by 
performing our computer-aided colonic polyp detection system 
when applied to 20 real data sets. First we assess the effect of 
noise reduction on detection process by testing the technique 
not including MIC algorithm. The results are summarized in 
table 1. Next, we examined the complete set of algorithm, 
containing all four steps. Table 2 shows the performance of our 
polyp detection technique. 
As the results express, adding the noise reduction algorithm 
improved the total sensitivity rate by about 10 percents. Main 
table-table 2- shows that technique achieved 100% sensitivity 
for detection of polyps larger than 10 mm which are the most 
important types of polyps to be detected in clinical studies. For 
polyps ranging from 5 to 10 mm there is almost high true 
positive where a sensitivity rate of 81.82% is achieved. Also 
our experiment shows a relative low sensitivity (66.67%) for 
polyps smaller than 5 mm. totally, the experimental data 
indicates that our polyp detection technique shows a sensitivity 
rate of 71.73%. 
Polyp type 
Total number 
of polyps 
True 
positive 
sensitivity 
>10 mm 
3 
3 
100% 
[5 -10 )mm 
11 
8 
72.73% 
<5 mm 
30 
17 
50.67% 
flat 
2 
0 
0% 
total 
46 
28 
60.87% 
Table 1. Results of performing the technique without noise 
reduction 
Polyp type 
Total 
number 
of polyps 
True 
positive 
sensitivity 
> 10 mm 
3 
3 
100% 
[5 -10 )mm 
11 
9 
81.82% 
<5 mm 
30 
20 
66.67% 
flat 
2 
1 
50% 
total 
46 
33 
71.73% 
Table 2. Results of performing the complete technique 
4. CONCLUSION 
We have presented and developed a set of algorithms for 
automatic colonic polyp detection including four stages: noise 
reduction, colon wall segmentation, feature extraction and 
finally polyp/fold classification. The morphological image 
cleaning algorithm smoothes the images while preserving their 
important features. Colon wall segmentation is done to 
determine the colon wall region within a 5 pixels margin. 
Feature extraction is done by AMPD algorithm applying 
morphological operators and our polyp/fold classification 
algorithm (TMA) is a template matching algorithm based on 
Euclidian distance searching. 
The proposed system for colonic polyp detection shows almost 
high sensitivity for medium and large polyps which means 
polyps between 5 to 10 mm and larger than 10 mm. it expressed 
total sensitivity of 71.73% which is about 10 percent higher 
than the case without image cleaning. 
REFRENCES 
Acar B, Beaulieu CF, Gokturk SB, Tomasi C, Paik DS, Jerrey 
RBJ, Yee J, Napel S., 2002. Edge displacement field based 
classification for improved detection of polyps in CT 
colonography. IEEE Transactions on Medical Imaging; 21(12): 
1461-7.