Abnormality Detection in Endoscopic Images using

GETS Int’l Trans. Computer Science and Engr., Vol.9, No.1

Abnormality Detection in Endoscopic Images using HSI Segmentation and Curvature Computation

1

B.V.Dhandra1, Ravindra Hegadi1 Department of P.G. Studies and Research in Computer Science Gulbarga University, Gulbarga-585106 Karnataka, INDIA [email protected], [email protected]

Abstract. In this paper, a method for detecting possible presence of abnormality in the endoscopic images of lower esophagus is presented. The pre-processed endoscopic color images are segmented in the HSI color space. The zero-crossing method of edge detection is applied on the binary image corresponding to the segmented image. For the large contours, the Gaussian smoothing is performed for eliminating the noise in the curve. The curvature for each point of the curve is computed considering the support region of each point. The possible presence of abnormality is identified, when curvature of the contour segment between two zero crossing has the opposite curvature signs to those of such neighboring contour segments on the same edge contours. The experimental results for the proposed method show successful abnormality detection in the test images. Keywords. Endoscopy, abnormality, HSI color space, curvature, support region.

1 Introduction The technique of endoscopy has expanded the understanding of numerous gastrointestinal diseases since from its wide spread use in the late 1960s. As the video endoscope containing the intensity light source, suction equipment, guided camera, etc, passes under direct vision, through the esophagus and the stomach into a portion of duodenum, it transmits the video clipping of tissues for the display, the storage and the analysis [1]. Endoscopy of lower gastrointestinal system provides real time image information and is being used increasingly to identify abnormalities and disorders of the colon [2]. Colonic polypoid lesions are the most common pathology found during endoscopy. The abnormality of polyps and tumors are mainly detected when the surface of the lipoma is eroded or irregular in contrast to a smooth surface [3]. Normally, the creases of colon haustra, which are seen as contours in the endoscopic image, are smooth and are of arc shapes. However, the presence of polyps or tumors will lead to the shape of these contours being seen as distorted. Such distorted shape is reflected by the change of curvature sign along a normal smooth contour of the same curvature sign. Thus the possible presence of abnormality can be detected, if the contour’s curvature is analyzed. This approach is used by Krishnan et.al.[9] for the 2

Corresponding author.

159

160

Abnormality Detection in Endoscopic Images using HSI Segmentation


intestinal abnormality detection from Endoscopic images based on the Canny’s method for edge detection followed by curvature analysis. Another method proposed by Hiremath et.al [12] to detect the possible presence of abnormality using color segmentation of the image based on 3σ-interval method [5] for edge detection followed by curvature analysis for the image contours is a significant work in the abnormality detection. In this paper a new method is proposed for segmentation of image in HSI color space for finding the abnormal region and detection of edges on this region by adapting zero crossings method, which reduces the time required for the image pre processing with the significant improvement in the results as compared to the results obtained by the methods proposed by Krishnan et.al.[9] and Hiremath et.al [12].

2 Methods The edge contours in the color endoscopic images are extracted using hue, saturation and intensity (HSI) color space segmentation and edge detection by zero crossings method. The contours are then smoothed along its length to make them suitable for curvature computation, since curvature is represented by second order differential making it very noise sensitive. Abnormality in the image is detected through curvature analysis along the contours. 2.1 Segmentation in HSI color space When human views a color object, he describes it by its hue, saturation and intensity or brightness. The hue is a color attribute that describes a pure color, where as saturation gives a measure of the degree to which a pure color is diluted by white light and brightness is a subjective descriptor [4]. The HSI color model decouples the intensity component from the color carrying components, hue and saturation, in a color image. As a result the HSI model is an ideal tool for developing algorithms for image processing based on color descriptions that are natural and intuitive to humans. The proposed method converts the input RGB color image into HSI color space. This image will be having three components namely hue, saturation and intensity. In this method the saturation component is thresholded with a threshold value st to obtain a binary image. Then, every pixel of this binary image is multiplied to the corresponding pixel in the hue component of the HSI image to obtain the product image. On the other hand a compliment transform is applied on the intensity component of the HSI image resulting in bright pixels into dark and visa versa, which can be performed by subtracting each pixel by 255. Such obtained image is converted in to binary image by thresholding with the threshold value it. The product image and above thresholded image are combined or merged to obtain the final image on which edge detection is performed as explained in section 2.2. The segmentation in HSI color space is described with the block diagram given in Fig. 1 below.

Fig. 1. Segmentation in HSI color space

2.2 Edge Detection and Curvature Computation The image after segmentation in HSI color space will be a binary image. The edge detection scheme is applied on this image based on zero crossings method. An edge detection technique based on the zero crossings of second derivatives [10] explores the fact that step edge corresponds to an abrupt change in the image function. Thus, for an image function f(x)

d 2 f ( x) =0 dx 2

(1)

is taken as the condition for the edge detection by zero crossing. The output of the process is a series of edge contours. After edge detection using zero crossing technique, only large contours are used for curvature analysis. Due to the discrete boundary representation and quantization errors, false local concavities and convexities along a contour are formed. This noisy nature of binary contours must be taken into account to obtain reliable estimates of contour curvature. Hence, a Gaussian filter is used to smooth the contour points to reduce the noise effect [6]. However, the width of Gaussian filter, w, that controls the degree of smoothing has to be chosen suitably. A large value of w will remove all small details of the contour curvature, while a small value will permit false concavities and convexities to remain in the contour, thus enforcing an appropriate choice of w. To overcome this problem a support region is employed which will dynamically determine the Gaussian parameter.

161

162

Abnormality Detection in Endoscopic Images using HSI Segmentation

2.3 Determination of Support Region


η(t, w) =

1 2πw2

 −t 2   2 e  2w 

163

(5)

where w is the width of the filter, which needs to be determined. The smoothed curve is denoted by set of points (X(t,w), Y(t,w)), where, x(t , w) = x(t ) ⊗ η (t , w) , y (t , w) = y (t ) ⊗ η (t , w) where ⊗ denotes the convolution.

Fig. 2. Representation of Support Region

The support region concept can be explained using the Figure 2. The support region for each point on the curve is the number of points obtained from the implementation of the Algorithm 1. Algorithm 1 (i) Determine the length of the chord joining the points Pi − k , Pi + k as l i ,k = Pi − k Pi + k

(2)

(b)

l i ,k ≥ l i ,k +1

d i,k l i, k

≥

d i , k +1 l i , k +1

(3)

Window Len =2xSupp. RegionD(Pi)+1, Width w =Support Region D(Pi) / 3

(7)

For the continuous curve C, expressed by {x(s), y(s)}, where s is the arc length of the edge point, the curvature can be expressed as: k (s) =

(x&

x&&y& − &x&y& 2

+ y& 2

)

32

(8)

where, x& = dx ds , &x& = d 2 x ds 2 , y& = dy ds , &y& = d 2 y ds 2 .

for d i ,k ≥ 0

Now, the region of support of Pi is the set of points satisfying either condition (a) or condition (b), that is,  Pi − k ,........., Pi −1 , Pi , Pi +1, ........, Pi + k  D( Pi ) =    | condition (a) or Condition (b) 

The measurement of the curvature of the point is based on the local properties within its region of support, and the length of Gaussian smooth filter is proportional to the region of support [7]. This implies that the neighboring points closer to the point of interest should have higher weights than those points further away. This method is less sensitive to noise. The Gaussian filter applied here will have the following window length and width [9]:

After finding the support region, the curvature for each point on the curve is calculated in the following way.

(ii) Let di,k be the perpendicular distance from Pi, to the line joining Pi − k Pi + k , start with k=1, compute li,k and di,k until one of the following conditions hold (a)

(6)

For digital implementation, the coordinate functions x(s) and y(s) of the curvature are represented by a set of equally spaced cartesian grid samples. The derivatives in the equation (8) are calculated by finite differences as: x& i = x i − x i −1 , y& i = y i − y i −1 , &x&i = xi −1 − 2 x i + x i +1 , &y&i = y i −1 − 2 y i + y i +1

(4)

2.4 Gaussian Smoothing

A planar curve can be defined in parametric form as (x(t), y(t)) ∈ R2, where t is the path length along the curve. Smoothing is performed by the convolution of x(t) and y(t) with the Gaussian filter. A one dimensional Gaussian filter is defined as

(9)

The algorithm for curvature computation is presented below. Algorithm 2 (i) Transform the input RGB color image into HSI color space. (ii) Obtain the binary image by thresholding the saturation component with threshold value st. (iii) Obtain the product image by multiplying the above image with the hue component.

164 Abnormality Detection in Endoscopic Images using HSI Segmentation

(iv)

Find the compliment image of intensity component and threshold with a threshold value it. (v) Merge the product image with the binary image to obtain the HSI segmented image. (vi) Find the edge contours by edge detection using zero-cross method. (vii) Determine the support region for each contour point. (viii) Smooth the contour by a Gaussian filter with the width proportional to the support region. (ix) Compute the curvature for each point on the Gaussian smoothed curve using equation (8) and (9). In the process of curvature computation we come across with two special conditions for which the alternate solutions need to be given. They are: (i) (ii)

when the edge point is on a straight line, the curvature for that point is assigned to zero. when the support region for an edge point is 1, this point will not be smoothed. So, the smoothing on this point is performed using the following equation:

(x)i, y) i ) = 1 (x i , y i ) + 1 [(x i −1 , y i −1 ) + (x i +1 , y i +1 )] 2

4

GETS Int’l Trans. Computer Science and Engr., Vol.9, No.1 165 to the segmented image are found using the zero crossings method as shown in Fig. 3(d). The Fig. 3(e) shows a large edge contour for the image. This edge contour is smoothed using Gaussian filter as shown in Fig. 3(f). The width of the filter depends on the support region for each point on the edge contour. The curvature profile along an edge contour is obtained. This plot is having many rugged zero crossings. In order to analyze the signals effectively and automatically by the computer, the curvature is smoothed by a one-dimensional Gaussian filter, with standard deviation σ=4. The smoothed curve profile is shown in Fig. 3(g). Fig. 4 and Fig. 5 shows the curvature profile obtained for 5 abnormal images and two normal images respectively. Column (a) shows the endoscopic color image. Column (b) shows the smoothed large edge contour obtained after the edge detection by the proposed method. Column (c) is the smoothed curvature profile for the image using the proposed method. From the Fig. 4 and Fig. 5 It can be noticed that the proposed method has generated large number of continuous contours in the abnormal images and it has generated very less number of small contours in the normal images. The presence of large edge contours and the more number of zero crossings in the curvature profile of the abnormal images are the good indicators of the presence of the abnormality.

(10)

) ) where, (xi , y i ) is the smoothed point of (xi , y i ) .

2.5 Abnormality detection in endoscopic image

The detection of possible presence of abnormality is performed by analyzing the curvature change along each edge contour. The curvature of each edge point on the edge contour is computed using the Algorithm 2. Two thresholds, cth and nth, are used in the analysis. cth is the curvature threshold value, and nth is number of edge points in a segment. Along the edge contour, if the absolute curvature of the point is bigger than cth, the point counting starts until the absolute curvature value of the point is less than cth. If the point count is bigger than nth, an edge segment is formed. The possible presence of abnormality in the image is detected when the curvature of a segment has opposite sign to those of such neighboring segments on the same edge contour. Also such a segment is bounded by two significant zero crossings.

3 Experimental Results and Discussion The experiment is carried out on 20 normal and 15 abnormal images obtained from the medical expert. The images are collected from Olympus V70 Endoscopic equipment. The Matlab 6.1 software is used for implementation of the algorithm. The Fig. 3(a) shows an abnormal endoscopic image. The RGB color image is converted into HSI color space as shown in Fig. 3(b). Fig. 3[c] shows the region after segmenting the image using HSI color space segmentation. The edges corresponding

Fig. 3. Results for an abnormal image [a] Color image showing abnormality, [b] Image after conversion to HSI color space, [c] HSI segmented image, [d] Edge detected by zero crossing method, [e] A large edge contour corresponding to the image, [f] Edge after smoothing, [g] Smoothed curvature profile of the edge.

166 Abnormality Detection in Endoscopic Images using HSI Segmentation


167

4. Conclusion In this paper a new novel technique is proposed for detection of possible presence of abnormality in an endoscopic image based on the segmentation of image in HSI color space. The model is well suited for describing colors in terms that are practical for human interpretation [4]. The miss detection of bright spots in the image due to reflection is avoided, thus reducing the time required for further image processing. However, this is not so in the method adopted by Krishnan et. al. [2], since, it is based on the edge detection of intensity image. The present method does not require much time for pre-processing as compared to the method proposed by Hiremath et.al. [12]. It generates large continuous edge contours, as compared to the method proposed by Krishnan et. al. [2]. Large continuous edge contours are more advantageous than shorter ones in so far as the curvature computation and determination of number of zero-crossings is concerned. The present method located more number of zerocrossings in the edge contour curvatures and it does not generate large edges in the normal image and hence is a good indicator of the presence of abnormalities in the color image than the method proposed by Krishnan et. al. [2].

Acknowledgements Authors are grateful to Dr. M. K. Ramakrishna, MS., Sri Lakshminarayana Nursing Home, Raichur, for providing endoscopic images and rendering manual segmentation for the present study. The authors are also indebted to Dr. P. Nagabushan, Dr. G. Hemanthkumar and Dr. D.S. Guru, Dept. of Studies in Computer Science, University of Mysore, for their helpful discussions and encouragement during this work. Fig. 4. Curvature profile of five abnormal images based on proposed method (a) Original Color image, (b) Large smoothed edge and (c) Smoothed curvature for the corresponding edge.

Fig. 5 Curvature profile of the two normal images based on proposed method (a) Original Color image, (b) Large smoothed edge and (c) Smoothed curvature for the corresponding edge.

References [1] C. Le. Guillou, J. M. Cauvin, B. Solaiman, I. M. Robaszkiewicz, “Knowledge Representation and case Indexing in Upper Digestive Endoscopy”, Proceedings of 22nd Annual EMBS International Conference, Chicago IL, 2000. [2] S. M. Krishnan, K. V. Asari, C. J. Yap, P. M. Y. Goh, “A neural Network based approach for Classification of Colon Abnormality”, Intl. Symposium on Intelligent Robotic Systems, 1998. [3] F. Silverstein, G. Tytget, Atlas of Gastrointestinal Endoscopy, Gower Medical Publication. [4] Refael Gonzalez, Richard E. Woods, Digital Image Processing, Pearson Edition Asia, 2nd Edition, 2002. [5] P. S. Hiremath, B.V. Dhandra, Iranna Humnabad, Ravindra Hegadi, G.G. Rajput, “Detection of esophageal Cancer (Necrosis) in the Endoscopic images using color image segmentation”, Proceedings of second National Conference on Document Analysis and Recognition (NCDAR-2003), Mandya, India, 2003. [6] V. Torre, T.A. Poggio, “On Edge Detection”, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 8, pp. 147-163, 1986. [7] N. Ansari, K.W. Huang, “Non-parametric Dominant Point Detection”, Pattern Recognition, Vol. 24, pp. 849-862, 1991. [8] C. Tech, R.T. Chin, “On the Detection of Dominant points on Digital Curves”, IEEE Trans. on Pattern Analysis and Machine Int., Vol. 11, pp. 859-872, 1989.

168 Abnormality Detection in Endoscopic Images using HSI Segmentation [9] S. M. Krishnan, X. Yang, K. L. Chan, S. Kumar, P. M. Y. Goh, “Intestinal Abnormality Detection from Endoscopic Images”, The 20th Annual International Conference of IEEE EMBS 98, Hongkong 1998. [10] D. Marr, E. Hildreth, “Theory of Edge detection”, in R. Kasturi and R.C. Jain Editors, Computer Vision, IEEE Los Alamitos, CA, pp. 77-107, 1991. [11] Milan Sonka, Vaclav, Hlavac, Roger Boyle, “Image Processing and Machine Vision”, PWS Publishing Company, 2001. [12] P.S.Hiremath, B.V.Dhandra, Ravindra Hegadi, G.G.Rajput, “Abnormality detection in endoscopic images using color segmentation and curvature computation”, 11th International Conference on Neural Information Processing, ICONIP-2004, ISI, Calcutta, India, LNCS, ISBN-3-540-23931-6, Springer-Verlag, 2004.

Biography Name: B. V. Dhandra Address: Professor and Chairman, P. G. Department of Studies and Research in Computer Science,Gulbarga University, Gulbarga, Karnataka, INDIA Education and work experience: M.A.(Statistics), M.Phil.,Ph.D., 21 years of Post Graduate teaching experience. Tel.:+91-8472-249682 E-mail: [email protected] Other information: Dr. B. V. Dhandra born in Gulbarga, India on January 1st 1955. He received MA degree in Statistics in 1979 and M.Phil in Statistics in 1986 from Karnataka University Dharwad, India. He obtained his Ph.D. degree from Shivaji University, Kolhapur, India in the year 1993. He served as lecturer during 1979 to 1993, as a reader during 1993 to 2001 and since 2001 he has been appointed as professor in Computer Science and presently heading the PG Department of studies and research in Computer Science, Gulbarga University, Gulbarga, India. His research interests are Pattern Recognition, Image Processing and Operations Research. Name: Ravindra Hegadi Address: Research Scholar, P. G. Department of Studies and Research in Computer Science, Gulbarga University, Gulbarga, Karnataka, INDIA Education and work experience: M.C.A., M.Phil., 9 years teaching experience. Tel.:+91-94480-23871 E-mail: [email protected]

GETS Int’l Trans. Computer Science and Engr., Vol.9, No.1 169 Other information: Ravindra Hegadi was born in Gulbarga, India, on February 28th 1970. He received the Master’s degree in Computer applications in 1995, and Master of philosophy in 2004 from Gulbarga University, Gulbarga, India. From February 1997 he is serving as lecturer in Government College, Raichur, India. From March 2004 he is persuing his Ph. D. research under the guidance of Prof. B. V. Dhandra, in the P.G. Department of studies and Research in Computer Science. His research interests are Medical image processing, Color image processing and Pattern Recognition.