Comparison of SUSAN and Sobel Edge Detection in MRI Images for Feature Extraction

Gholamali Rezai-Rad Iran University of Sci. & Tech. Electrical Engineering Department Email. Re ust. ac. ir
Abstract

Majid Aghababaie Iran University of Sci. & Tech. Electrical Engineering Department Email.h
restricted only to "corner-like" cases, but instead taking the concept of "points of localized structure" as an heuristical definition for two-dimensional features. The Plessey feature detector [8] is the most used of this class of algorithms, and leads to a low number of false positives and negatives detected. However, its localization accuracy depends on the feature shape. Since reasonable algorithms for edge and feature detection are already available, the development of new approaches is only justified if they can offer a better performance, a lower computational demand or both. The SUSAN method looks promising in achieving both of these desired characteristics.

A new comparison of SUSAN and Sobel edge detection for MRJ images feature extraction is proposed in this paper. Preprocessing algorithms are important in MRJfeature extraction since they provide the necessary data in the best task. One of these algorithms is edge detection. We introduce two common edge detection algorithms which are SUSAN and Sobel, and then we compare them and explain their properties according to MRJ feature extraction application. By means of experimental results we have made a useful guide for researchers to choose Sobel or SUSAN edge detection in their special MRIfeature extraction problem.

1. Introduction
The preprocessing algorithms, techniques, and operators are used to perform initial processing that makes the primary data reduction and analysis task easier. They include operations related to extracting regions of interest, performing basic algebraic operations on images, enhancing specific image features, and reducing data in both resolution and brightness [1]. Edge detection is one of these algorithms. Edges characterize object boundaries that provide useful information for identifying objects in an image. Detection of edge is a fundamental step for most computer vision applications such as MRI feature extraction and remote sensing [2]. Edge image also can reduce the space and computational time required in further steps for analyzing the image using methods of labeling, Hough transformation, and other image

2. Edge detection
Edge detection operators are based on the idea that edge information in an image is found by looking at the relationship a pixel has with its neighbors [I]. If a pixel's gray-level value is similar to those around it, there is probably not an edge at that point. However, if a pixel has neighbors with widely varying gray levels, it may represent an edge point. In other words, an edge is defined by a discontinuity in gray-level values. Ideally, an edge separates two distinct objects. In practice, apparent edges are caused by changes in color or texture or by the specific lighting conditions present during the image acquisition process. Many of the edge detection operators are implemented with convolution masks, and most are based on discrete approximations to differential operators. Differential operations measure the rate of change in a function, in this case, the image brightness function. A large change in image brightness over a short spatial distance indicates the presence of an edge. Some edge detection operators return orientation information (e.g. Sobel and SUSAN), whereas others only return information about the existence of an edge at each point.

processing techniques [2],[3]. Many types of edge detection methods have been proposed to date [2], [3], [4], and [5]. Derivative-based edge detection approaches are widely used because they
The detection of two-dimensional (2D) image features is even more complex than the detection of edges (which can be considered one-dimensional features). The cause is the diversity and lack of agreement in mathematical definitions of two-dimensional features. As a consequence, many algorithms are adequate only for simple features such as "corners" and more general "points of localized structure" are not correctly reported. Some methods were derived [7,8] without being
0-7803-9521-2/06/$20.00 2006 IEEE.

are easy to use.

3. Sobel edge detection

The Sobel edge detection masks look for edges in both the horizontal and vertical directions and then combine this information into a single metric. The masks are as follows [1]:

1103

RowMask

1 O 1

-2 O 2
-1

-1O , 1
mask =

ColumnMask

-2

0 0

1]
2

(1)

0 0 0 1 1 1 1 1 1 0
-0

-1I

1
At

11 11 11 11 11 11 11 11 01110
if
if

1 1 1 0 1 1 1 1

0 0 1 1 1 1 1 0

(2)

0-

These masks are each convolved with the image. each pixel location we now have two numbers: corresponding to the result from the row mask, and from the column mask. We use these numbers compute the edge magnitude which is equal
2 2

si
s2
to to

The mask is placed at each point in the image and, for each point; the brightness of each pixel within the mask is compared with that of the nucleus (the center point). Originally a simple equation determined this
comparison.

~i

+ 52

-> ->

IQrJ-IQroJ .t
I
Qr

J-jI

(3)
>- t

4. SUSAN edge detection

Although much of the previously discussed methods include at least one non-linear stage, the SUSAN approach represents a somewhat different method for edge and feature extraction, since it is almost entirely based on non-linear filtering. The basic idea of the SUSAN method is to associate to each pixel of the image a small area of neighbor pixels with similar brightness to this center pixel. This small area is called the "USAN" (for "Univalue Segment Assimilating Nucleus"). From the size, centroid and axis of symmetry of these areas, edges and more general "localized features" are located. According to the SUSAN principle, "an image processed to give as output inverted USAN areas has edges and two-dimensional features strongly enhanced, with the two dimensional features more strongly enhanced than edges". The method then resumes in the search for local minima of the USAN areas, hence the acronym SUSAN, for "Smallest Univalue Segment Assimilating Nucleus". The SUSAN edge finder has been implemented using circular masks (sometimes known as windows or kernels) to give isotropic responses. Digital approximations to circles have been used, either with constant weighting within them or with Gaussian weighting. The usual radius is 3.4 pixels (giving a mask of 37 pixels), and the smallest mask considered is the traditional three by three mask. The 37 pixel circular mask is used in all feature detection experiments unless otherwise stated. This mask has the following form:
Where ro is the position of the nucleus in the two

dimensional image,

is the position of any other point

within the mask, I(rJ is the brightness of any pixel, t

is the brightness difference threshold and c is the output of the comparison. This comparison is done for each pixel within the mask, and a running total, n, of the outputs is made:
n

),
r

c r(

(4)

This total n is just the number of pixels in the USAN, i.e. it gives the USAN's area. The parameter t determines the minimum contrast of features which will be detected and also the maximum amount of noise which will be ignored. Next, n is compared with a fixed threshold g (the
3nmax geometric threshold), which is set to
,

where

initial edge response following rule:

nmax is the maximum value which

take. The is then created by using the

n can

R(r0) = ig

( )
0

{r>g

(5)

otherwise

0-7803-9521-2/06/$20.00 2006 IEEE.

1104

Where R

roj is the initial edge

response. This is

clearly a simple formulation of the SUSAN principle, i.e., the smaller the USAN area, the larger the edge response. When non-maximum suppression has been performed the edge enhancement is complete.

5. MRI feature extraction

One of the most useful medical imaging modalities is magnetic resonance imaging (MRI). The image gray levels in MRI depend on several issue parameters including: spin density (SD), spain-lattice (TI) and spinspin (T2) relaxation times, flow velocity (v), and chemical shift (6). The first three parameters are most important in clinical imaging [6]. An MRI scene sequence shows spatial location of different tissues, with different contrast in each image. It can therefore be considered as spatial domain representation of tissues. In a spatial domain representation, pixels corresponding to specific tissues are locally connected but may be distributed over different sections of the image. Image analysis can be accomplished using an appropriate feature space method. Feature space methods can be useful for all three steps of image analysis: 1) Identification of objects; 2) Segmentation of objects; and 3) Quantitative on objects, to obtain information that can be used in decision making (diagnosis, treatment planning, and evaluation of treatment). In a computerized image analysis system, the input image is first preprocessed. Preprocessing may involve restoration, enhancement, or just proper representation of the data. Then certain features are extracted for segmentation of the image into its components. The segmented image is fed into a classifier or an image understanding system. Feature extraction is a crucial step for MRI segmentation. Rather than using all the information in the images at once, feature extraction and selection breaks down the problem of segmentation to the grouping of feature vectors. Features can be pixel intensities themselves, feature calculated from the pixel intensities, or edge and texture features.

connectivity. Also edges with SUSAN are one pixel thick 2) The results show the SUSAN edge detector to be accurate, stable and very fast on both test and real images. It has given very good results in all the tests. 3) The speed of SUSAN edge detection in MRI feature extraction is much better than Sobel edge detection. The Sobel edge detection is approximately 10 times slower than SUSAN edge detection. 4) In the SUSAN edge detection, we can find the orientation of the edge easier than Sobel edge detection. 5) SUSAN edge detection for MRI feature extraction often doesn't need edge thinning methods. 6) We can use SUSAN edge detection for corner detection too and this is one of the most important features of this algorithm. Corner detection plays an important role in MRI feature extraction. 7) The performance of SUSAN edge detection in MRI feature extraction is more precise than Sobel edge detection.

Figure 1. Sobel and Susan Edge Detection algorithms implemented on an MRI.

6. Results
We have implemented two MRI sample images. Sobel and SUSAN edge detection operators have been implemented on them and we can see the results in the following figures. Then we have compared these two edge detection algorithms from the view point of MRI feature extraction and we have got the following results: 1) In the following figures, features of two images have been extracted using SUSAN Principle and Sobel Operator for comparison. It can be clearly seen that SUSAN provides much better edge localization and

0-7803-9521-2/06/$20.00 2006 IEEE.

1105

Figure 2. Sobel and Susan Edge Detection algorithms implemented on an MRI.

7. Conclusion
In this paper we have shown that between Sobel and SUSAN edge detection algorithms, response given by SUSAN edge detection was better than the best result of Sobel detector used in these MRI images feature extraction.

8. References
[1]. Scott E Umbaugh, "Computer vision and image processing", Prentice Hall, 1999. [2]. Davies E.R., "Machine vision: Theory, Algorithm, Practicalities", Academic Press, 1997. [3]. Pratt W.K., "Digital image processing", A Wiley Interscience Publication, 19991. [4]. Gonzalez R.C., Woods R.E., "Digital Image Processing", Addison Wesley, 2002. [5]. Rosenfeld A., Avinash C.K., "Digital picture processing", Academic Press, 1982. [6]. H. Soltanian-Zadeh, M. Kharrat, D. J. Peck, "Polynomial transformation for MRI feature extraction". [7]. Forstner, W., and Gulch, E., "A fast operator for detection and precise location of distinct points, corners and centers of circular features", Intercommission Workshop, Interlaken, June 1987, 149-155. [8]. Harris, C. G., and Stephens, M., "A combined corner and edge detector", 4th. Alvey Vision Conference, 1988, 147-151.

0-7803-9521-2/06/$20.00 2006 IEEE.

1106

0-7803-9521-2/06/$20.00 2006 IEEE.

1107