Comparative performance analysis of segmentation techniques

436 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
436
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
15
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Comparative performance analysis of segmentation techniques

  1. 1. InternationalINTERNATIONAL Communication Engineering & Technology (IJECET), ISSN 0976 – Journal of Electronics and JOURNAL OF ELECTRONICS AND6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEME COMMUNICATION ENGINEERING & TECHNOLOGY (IJECET)ISSN 0976 – 6464(Print)ISSN 0976 – 6472(Online)Volume 3, Issue 2, July- September (2012), pp. 238-247 IJECET© IAEME: www.iaeme.com/ijecet.htmlJournal Impact Factor (2012): 3.5930 (Calculated by GISI) ©IAEMEwww.jifactor.com COMPARATIVE PERFORMANCE ANALYSIS OF SEGMENTATION TECHNIQUES Amandeep Singh A.P Gursimran singh sandhu ECE, Lovely professional university, near ECE, Punjab Technical University Phagwara Punjab, India sandhugursimran@gmail.com deepsview@yahoo.comABSTRACTThe study presented in this article focuses on comparative analysis of Segmentation techniques.Segmentation techniques are applied to extract Region of Interest (ROI) from medical images obtainedfrom different medical scanners such as Ultrasound, CT or MRI. Global thresholding, AdaptiveThresholding, Region grow and Active contour using level set techniques has been used in the proposedsegmentation analysis. The approach consists of two steps: Apply segmentation technique to extract mostdiscriminative regions from image and calculate the parameters from the resulting image obtained by theapplied techniques. Parameters are precision, accuracy sensitivity, specificity. Segmentation techniqueshave been tested on medical and synthetic data sets and results are compared with each other. Testsindicate that using level set contour significantly improves the ability of extracting region of interest withunbroken boundaries and Adaptive thresholding acquires most of the details present in the image. Manualglobal thresholding have the highest success rate of extracting the region of interest.KeywordsGlobal threshold; Adaptive threshold; Region grow; Level set contour; Binary classification; Hybrid segmentation I. INTRODUCTIONThe research presented in this article is part of an on-going Mtech thesis aimed at developing anautomated hybrid imaging system for segmentation of tumor present in medical images obtained byComputed Tomography (CT) scans. Farzaneh Keyvanfard et al [1] Segmenting of human organs in CTscans using gray level information is particularly challenging due to the changing shape of organs inmedical images and the gray level intensity overlap in soft tissues. Medical image segmentation requiresextracting specific features from an image by distinguishing objects from the background. Medical imagesegmentation aims to separate known anatomical structures from the background for research, cancerdiagnosis, quantification of tissue volumes, radiotherapy treatment planning and study of anatomicalstructures. 238
  2. 2. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEMECancer diagnose can be manually performed by a human expert who simply examines an image,determines borders between regions, and classifies each region this process is called segmentation interms of image processing. This is perhaps the most reliable and accurate method of image segmentationbecause the human visual system is immensely complex and well suited to the task. But the limitationstarts in volumetric images due to the quantity of clinical data. Implementation of image processingincrease the rate of similar CT interpretation between different analysers, now its just 20% and to relieffor the analyzers from routine CT analysis.Nader H. Abdel-massieh et al [2] [3], thresholding iscommonly used image segmentation technique, In this method, pixels that are alike in grayscale (or someother feature) are grouped together. Often a image histogram is used to determine the best setting for thethreshold. After thresholding image is converted into logical image the pixels range above thresholdbecome 1 or white pixels and pixel range below threshold become 0 or black pixels. Bio medical imagesmay have multiple modes and multiple thresholds may be helpful. In general multilevel thresholding isless reliable than single level thresholding. Mostly because it is very difficult to determine thresholds thatadequately separate objects of interest. N. Otsu et al [4] Global Thresholding choose threshold T thatseparates object from background global thresholding is a single threshold method of thresholdingtechnique. When the pixel values of the components and that of background are fairly consistent in theirrespective values over the entire image, global thresholding could be used. In adaptive thresholding,different threshold values T1,T2,T3 etc for different local areas are used. This more sophisticated versionof thresholding can accommodate changing lighting conditions in the image. The fundamental drawbackof histogram-based region detection is that histograms provide no spatial information (only thedistribution of gray levels).Region-growing approaches exploit the important fact that pixels which are close together have similargray values. The first region-growing method was the seeded region growing method. This method takesa set of seeds as input along with the image. The seeds mark each of the objects to be segmented. Theregions are iteratively grown by comparing all unallocated neighboring pixels to the regions. Thedifference between a pixels intensity value and the regions mean is used as a measure of similarity. Thepixel with the smallest difference measured this way is allocated to the respective region. This processcontinues until all pixels are allocated to a region. Seeded region growing requires seeds as additionalinput. The segmentation results are dependent on the choice of seeds. Noise in the image can cause theseeds to be poorly placed. Unseeded region growing is a modified algorithm that doesnt require explicitseeds. Region Growing offers several advantages over conventional segmentation techniques. Unlikegradient and Laplacian methods, the borders of regions found by region growing are perfectly thin (sincewe only add pixels to the exterior of our Region) and connected. The algorithm is also very stable withrespect to noise. Region will never contain too much of the background, so long as the parameters aredefined correctly. Other techniques that produce connected edges, like boundary tracking, are veryunstable. Most importantly, membership in a region can be based on multiple criteria. We can takeadvantage of several image properties, such as low gradient or gray level intensity value, at once.An important class of segmentation methods is model based methods. Caselles, R. Kimmel [5] [6] ActiveContours, also known as Evolving Fronts . Active contour is an interface usually used to separatestructures and background on the image. There are two principal approaches to build an active contour:explicit or Lagrangian approach, and resulting interfaces called snakes, implicit or Eulerian approach, andresulting interfaces called level sets. These methods are used in the domain of image processing to locatethe contour of an object. Trying to locate an object contour purely by running a low level imageprocessing task such as canny edge detection is not particularly successful. Often the edge is notcontinues, i.e. there might be holes along the edge, and spurious edges can be present because of noise.The level set method makes it very easy to follow shapes that change topology, for example when a shapesplits in two, develops holes, or the reverse of these operations. An active contour tries to improve on thisby imposing desirable properties such as continuity and smoothness to the contour of the object. This 239
  3. 3. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEMEmeans that the active contour approach adds a certain degree of prior knowledge for dealing with theproblem of finding the object contour. An active contour is modeled as parametric curve, this curve aimsto minimize its internal energy by moving into a local minimum. In this paper we use level set contourmethod along with other methods for analysis.This paper introduces the use wide range of data set and presents a comprehensive analysis determiningthe optimal segmentation technique for the applied to CT scans. This paper is focusing on a robustimplementation technique for segmenting medical volumes and performing binary analysis methodologyon images obtained from a CT scanner. The rest of this paper is organized as follow: The followingsection illustrates the proposed medical image segmentation system analysis methodology have beenexplained in section 2. The results and analysis of the implemented of segmentation techniques formedical image segmentation is illustrated in section 3. Finally, section 4 includes the conclusions andfuture scope.II. METHODOLOGYBinary classification problems often result in a predicted probability surface, which is then translated intoa presence–absence classification map in more generalize way it is act of discriminating an item into oneof two groups based on specified measures or variables. This paper consists of four main binaryclassification methods sensitivity, specificity, accuracy and precision. Sensitivity and specificity arestatistical measures of the performance of a binary classification test, also known in statistics asclassification function. Sensitivity (also called recall rate in some fields) measures the proportion of actualpositives which are correctly identified as such (e.g. the percentage of sick people who are correctlyidentified as having the condition). Specificity measures the proportion of negatives which are correctlyidentified (e.g. the percentage of healthy people who are correctly identified as not having the condition).The accuracy of a measurement system is the degree of closeness of measurements of a quantity to thatquantitys actual(true) value. IMAGE NOISE ADDITION SEGMENTATION TECHNIQUES COMPARISION BINARY CHART CLASSIFICATIONFigure 1. Process DiagramThe precision of a measurement system, also called reproducibility or repeatability, is the degree to whichrepeated measurements under unchanged conditions show the same results. Image is first converted in togray scale because resulting images of biomedical scans are gray in nature .Salt and paper noise is used innoise addition process, different segmentation techniques are applied on the resulting image of noiseaddition process. Every applied segmentation technique produces a segmented image which is in logicalscale. Binary classification methodology is used to statistically represent the performance of each segment 240
  4. 4. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEMEtechnique and comparison is done on basis values obtained by binary classification methods. Thesemethods of classification simply work on existence of desired and undesired objects, edges in the image.Binary classifier use terms to indicate presence of a edges and objects. Some of the images have theobjects, clear edges and our test says they are positive. They are called true positives (TP). Some have theobjects, but the test claims they dont. They are called false negatives (FN). Some dont have the objects,and the test says they dont - true negatives (TN). Finally, we might have image with no clear edges thathave a positive test result - false positives (FP). Thus, the number of true positives, false negatives, truenegatives, and false positives add up to 100% of the set. Specificity (TNR) is the proportion of objectsthat tested negative (TN) of all the objects that actually are negative (TN+FP). As with sensitivity, it canbe looked at as the probability that the test result is negative given that the image is not having object.With higher specificity, fewer will be the undesired objects which r not present in original image.Sensitivity (TPR) is the proportion of objects that tested positive (TP) of all the objects that actually arepositive (TP+FN). It can be seen as the probability that the test is positive given that the presence ofobject in image.The accuracy is the proportion of true results (both true positives and true negatives) in the population. Precision orpositive predictive value is defined as the proportion of the true positives against all the positive results (both truepositives and false positives). An accuracy of 100% means that the measured values are exactly the same as thegiven values. In the medical domain, the most important performance measures are both specificity and sensitivity.Optimally one would want both high specificity and high sensitivity measures. However, theoretically these twomeasures should have a negative correlation. Since accuracy reflects both the sensitivity and specificity in relation toeach other, this descriptor was selected to determine the overall correctness of the classifier. To evaluate theperformance of each classifier; specificity, sensitivity, precision, accuracy rates are then calculated from each of themisclassification matrices.Table1-Description table of TP, TN, FP, FN True positive (TP) Correctly identified True negative (TN) Correctly rejected False positive (FP) Incorrectly identified False negative(FN) Incorrectly rejectedTable 2-Defination of binary classification Sensitivity True Positive / Total Positive Specificity True Negative / Total Negatives Accuracy (True Positives + True Negatives) / (True Positive + True Negative + False Positive + False Negative) Precision True Positive / (True Positive + False Positives) 241
  5. 5. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEME III. RESULTS The segmentation algorithms are applied to various formats of images and results are given below. The binary analysis Parameters has value range from 0 to 1. (a) (b) (c) (d) (e) (f)Figure 2. Segmentation results for Eye image .(a) Original eye image effected by salt paper noise (b) Global threshold applied on original eye image (c) Resultof Adaptive threshold segmentation applied on eye image.(d) Eye image result of region grow thin segmentation (e) Eye image result of region grow thinsegmentation (c) Result of Level set segmentation applied on eye image. (a) (b) (c) (d) (e) (f)Figure 3. Segmentation results for Bone image .(a) Original bone image effected by salt paper noise (b) Global threshold applied on original bone image (c)Result of Adaptive threshold segmentation applied on bone image.(d) Bone image result of region grow thin segmentation (e) Bone image result of regiongrow thin segmentation (c) Result of Level set segmentation applied on Bone image. (a) (b) (c) (d) (e) (f)Figure. 4. Segmentation results for Lena image .(a) Original lena image effected by salt paper noise (b) Global threshold applied on original lena image (c)Result of Adaptive threshold segmentation applied on lena image.(d) Lena image result of region grow thin segmentation (e) Lena image result of regiongrow thin segmentation (f) Result of Level set segmentation applied on lena image. 242
  6. 6. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print),ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEME (a) (b) (c) (d) (e) (f) Figure 5. Segmentation results for Hex shapes image .(a) Original hex shapes image effected by salt paper noise (b) Global threshold applied on original hex shapes image (c) Result of Adaptive threshold segmentation applied on hex shapes image.(d) Hex shapes image result of region grow thin segmentation (e) Hex shapes image result of region grow thin segmentation (f) Result of Level set segmentation applied on hex shapes image. (a) (b) (c) (d) (e) (f) Figure 6. Segmentation results for Frog image .(a) Original frog image effected by salt paper noise (b) Global threshold applied on original frog image (c) Result of Adaptive threshold segmentation applied on frog image.(d) Frog image result of region grow thin segmentation (e) Frog image result of region grow thin segmentation (f) Result of Level set segmentation applied on frog image. (a) (b) (c) (d) (e) (f) Fig. 7. Segmentation results for Frog image.(a) Original frog image effected by salt paper noise (b) Global threshold applied on original frog image (c) Result of Adaptive threshold segmentation applied on frog image.(d) Frog image result of region grow thin segmentation (e) Frog image result of region grow thin segmentation (f) Result of Level set segmentation applied on frog image. Table 3. Results of Sensitivity applied on segmentation methods Images Eye Bone Lena Frog Hex shapes Two cell Techniques Sensitivity Sensitivity Sensitivity Sensitivity Sensitivity Sensitivity Result of segmentation 0.75 1 1 1 0.25 1 after Global thresholding Result of segmentation 0.75 0.66 0.25 0.8 0.25 1 after Adaptive thresholding Result of segmentation after 0.75 0 1 0.4 0.5 0 region grow-thick Result of segmentation after 0.75 0 1 0.4 0.5 0 region grow-thin Result of segmentation after 0 1 0 1 1 0 level set 243
  7. 7. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print),ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEME Table 4. Results of Specificity applied on segmentation methods. Images EYE Bone Lena Frog Hex Two cell Techniques Specificity Specificity Specificity Specificity Specificity Specificity Result of segmentation 0.5 1 0.5 1 0.66 0.5 after Global thresholding Result of segmentation after Adaptive 0.5 0 0.25 0.66 0 0.5 thresholding Result of segmentation 0.5 1 0.5 1 1 1 after region grow-thick Result of segmentation 0.5 1 0.5 1 1 1 after region grow-thin Result of segmentation 0.5 0.66 0 0.33 1 0 after level set Table 5. Results of Accuracy applied on segmentation methods. Images Eye Bone Lena Frog Hex shapes Two cell Techniques Accuracy Accuracy Accuracy Accuracy Accuracy Accuracy Result of segmentation 0.66 1 0.75 1 0.42 0.75 after Global thresholding Result of segmentation after Adaptive 0.66 0.33 0.25 0.75 0.14 0.25 thresholding Result of segmentation 0.66 0.5 0.75 0.62 0.71 0.5 after region grow-thick Result of segmentation 0.66 0.5 0.75 0.62 0.71 0.5 after region grow-thin Result of segmentation 0.16 0.83 0 0.75 1 0 after level set Table 6. Results of Precision applied on segmentation methods. Images Eye Bone Lena Frog Hex Two cell Techniques Precision Precision Precision Precision Precision Precision Result of segmentation 0.75 1 0.66 1 0.25 0.66 after Global thresholding Result of segmentation after Adaptive 0.75 0.4 0.25 0.8 0.25 0.66 thresholding Result of segmentation 0.75 0 0.66 1 0.5 0 after region grow-thick Result of segmentation 0.75 0 0.66 1 0.5 0 after region grow-thin Result of segmentation 0 0.75 0 0.71 1 0 after level set 244
  8. 8. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 – 6464(Print),ISSN 0976 – 6472(Online) Volume 3, Issue 2, July July-September (2012), © IAEME Table 7. Overall performance results of different segmentation methods. Average Average Average Average Average Overall Techniques Sensitivity Specificity Accuracy Precision Performance Result of segmentation 0.83 0.69 0.76 0.72 0.75 after Global thresholding Result of segmentation 0.61 0.31 0.39 0.51 0.455 after Adaptive thresholding Result of segmentation afte 0.44 0.83 0.62 0.48 0.5925 region grow-thick Result of segmentation afte 0.44 0.83 0.62 0.48 0.5925 region grow-thin Result of segmentation after 0.5 0.41 0.45 0.41 0.4425 level set 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Average Average Average Accuracy Average Precision Overall sensitivity Specificity Performance Results of segmentation after Global thresholding Results of segmentation after Adaptive thresholding Results of segmentation after region grow thick Results of segmentation after region grow thin Results of segmentation after Level set Figure 8. Graphical representation of performance of segmentation methods. IV. CONCLUSION The image segmentation is a relevant technique in image processing. Numerous and varied methods exist for many applications. Now that we have described the algorithms, we can compare the outputs and check which type of segmentation technique is better for a particular format. It is believed that the binary classification is best analysis method for biomedical image key factors which allow for the use of a segmentation algorithm in a Many object detection system:. Accuracy, Sensitivity, Precision, Specifici Specificity. On an average parameter set of the edge detection techniques, Global thresholding technique performed better than all other techniques for all the formats of images, sample of which can be found in results. Histogram based methods are found to be very efficient in terms of computation complexity when compared to other image segmentation methods. If significant peaks and valleys are identified properly and proper thresholding is fixed, this technique yields good result. Region grow technique operates well Region-grow wel 245
  9. 9. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEMEover all formats of images provided proper seed point is selected and range of threshold is properlydefined. This method performs well even when noise is present and it is reflected with a reasonable valuesof parameters in table. Adaptive thresholding yield fine results over all but in some type of image it yieldbetter results than all other segmentation methods, we can observe it by resulting image of segmentationmethods . The level set algorithm is guaranteed to converge but it may not return optimal solution fordetails of images. Region grow can enhance salt noise if seeds are not selected properly. As the adaptivethresholding perform nearly close to global thresholding and adaptive thresholding is an automaticprocedure of segmentation it is a better choice for hybrid segmentation. Region grows and Adaptivethresholding will be used for creating a hybrid segmentation method to improve the Detection andsegmentation of liver cancer from CT scan.ACKNOWLEDGMENTTHE AUTHOR IS THANKFUL TO ALL THE STAFF MEMBERS OF THE SCHOOL OF ECE, LOVELY PROFESSIONALUNIVERSITY FOR THEIR VALUABLE SUPPORT.REFERENCES[1] Farzaneh Keyvanfard” Feature selection and classification of breast MRI image “Artificial Intelligence and Signal Processing AISP 2011 International Symposium on (2011) pp. 54 – 58[2] Nader H. Abdel-massieh “Fully Automatic Liver Tumor Segmentation from Abdominal CT Scans” .[3] Amandeep singh, Jaspinder sidhu “Performance Analysis of Segmentation Techniques” International Journal of Computer Applications (0975 – 8887) Volume 45– No.23, May 2012[4] N. Otsu, “A Threshold Selection Method from Gray-Level Histograms”, IEEE Trans.Syst., Man, Cybern., vol. SMC-9 (1), pp. 62-66, Jan. 1979.[5] C.M. Li, C.Y. Xu, C.F. Gui, M.D. Fox, Level set evolution without re-initialization: a new variational formulation, in: IEEE Conference on Computer Vision and Pattern Recognition, San Diego, 2005, pp. 430–436.[6] Caselles, R. Kimmel, G. Sapiro, Geodesic active contours, in: Processing of IEEE International Conference on Computer Vision’95, Boston, MA, 1995, pp. 694–699.[7] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces, Springer-Verlag, New York, 2002.[8] P. K. Sahoo, S. Soltani and A. K. C. Wong, “A Survey of Thresholding Techniques”, Computer Vision, Graphics, and Image Processing, vol. 41, 133-260 (1988).[9] J. S. Weszka, R. N. Nagel, and A. Rosenfeld, “A threshold selection technique”, IEEE Trans. Comput., vol. C-23, pp. 1322-1326, 1974.[10] N. R. Pal and S. K. Pal, “A Review on Image Segmentation Techniques”, PatternRecognition, vol. 26, No. 9, pp. 1277-1294, 1993.[11] N. Lee et al., “Fatty and fibroglandular tissue volumes in the breastsof women 20-83 years old: Comparison of X-ray mammography andcomputer-assisted MR imaging,” Amer. J. Roentgenol., vol. 168, pp.501–506, 1997. 246
  10. 10. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –6464(Print), ISSN 0976 – 6472(Online) Volume 3, Issue 2, July-September (2012), © IAEME[12] L. Ludemann, P. Wust, and J. Gellermann, “Perfusion measurement using DCE-MRI: Implications for hyperthermia,” Int. J. Hyperthermia,vol. 24, no. 1, pp. 91–96, 2008.[13] N. Senthilkumaran et al,” Edge Detection Techniques for Image segmentation – A Survey of Soft Computing Approaches” nternational Journal of Recent Trends in Engineering, Vol. 1, No. 2, May 2009.[14] A. Korpel, " Acousto-Optics," in Applied Solid State Science, R. Wolfe, ed.,vol.3, Academic, New York (1972).[15] Shudong Wu, Feng Cheng and Francis T.S.YU, “Pattern recognition by OTF method”, J.Optics (paris), vol.20, 5, pp 201-204, 1989.[16] Joseph Rosen, “Three-dimensional optical Fourier transform and correlation”, Vol.22, No. 13, Optics Letters, 964-966, July 1, 1997[17] Ting-Chung Poon and Taegeum Kum, “Optical image recognition of three dimensional objects”, Vol.38, No.2, Applied Optics, 370-381, 10 Jan 1999. 247

×