Research Journal of Recent Sciences ________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res.J.Recent Sci. International Science Congress Association 1 Framework for the Comparison of Classifiers for Medical Image Segmentation with Transform and Moment based featuresMaria Hameed, Muhammad Sharif, Mudassar Raza, Syed Waqas Haider, Muhammad Iqbal Department of Computer Sciences, COMSATS Institute of Information Technology, Wah Cantt., 47040, PAKISTAN Available online at: www.isca.in Received 10th October 2012, revised 14th November 2012, accepted 7th February 2013 AbstractThe paper depicts and elaborates a new framework for the comparison of classifiers for medical image segmentation with transform and moment based features. Medical images modalities such as Ultrasound (US) bladder, Ultrasound (US) phantom, Computerized Tomography (CT) and Magnetic Resonance (MR) images are segmented using different algorithms namely, k-Nearest Neighbor (kNN), Grow and learn (GAL) and Incremental Supervised Neural Networks (ISNN). Segmentation is performed by applying feature extraction methods such as 2D Continuous Wavelet Transform (2D-CWT), Moments of gray level histogram (MGH) and a combined version of both 2D-CWT and MGH, called Hybrid features. With different iterations, the analysis results indicate that kNN performs better than GAL, and the performance of GAL is better than that of the ISNN for image segmentation. During analysis a comparison has been drawn between the performance of kNN, GAL and ISNN on the above three feature extraction schemes and also provides the qualitative and quantitative analysis of three classifiers. Results indicate that the performance of 2D-CWT and that of Hybrid features is consistently better than MGH features for all image modalities. The demonstrated frame work or the system is capable to meet the demand for selecting best approach in order to meet the given time constraints and accuracy standards in medical image segmentation. Keywords: kNN, GAL, ISNN, 2D-CWT, MGH Introduction Automatic tissue segmentation of images is helpful for radiologists, as it is used to facilitate doctors during diagnosis. Segmentation of medical images means to classify and identify the structure of interest in medical images. The overall objective is the computer-aided identification of the area of interest to help the doctors and radiologist during diagnosis and treatment of specific disease. Feature extraction is used for extracting sufficient and desired information from the image resulting by different variations from its features. Peculiar features having relevant information are chosen failing which culminates the segmentation process not to be executed correctly/properly1-5. For extracting right features, there is need of efficient feature extraction methods. In this paper three transform and moment based segmentation techniques namely, 2D-CWT, MGH and hybrid are analyzed with three different classifiers. In the literature, there are several approaches for image segmentation to be used for different applications, such as edge detection based segmentation, region growing based segmentation method, threshold based segmentation, level set method based segmentation, neural network based segmentation techniques10, Watershed algorithm based segmentation3,5, graph theory based segmentation1,11, clustering based segmentation12, active counter model based segmentation10,13, Marcove random field model based segmentation14, deformable model based segmentation15 and improved mean shift based segmentation16. In the literature there are different transform and texture features extraction based segmentation approaches are found17-19. Similarly 2D continuous, discrete wavelet transforms and 2D discrete cosine transform based feature extraction methods for segmentation are represented by Wang et al. and Ghazali et al.20-21. The main problem with some of the above methods is that they need too much computational resources and time for segmentation process. Some of them require too many parameters for proper performance yet these fail to meet the desired performance level. The main work in this paper is to find out the best combination of classifiers with feature extraction schemes to achieve efficient segmentation for medical images. Recently, grow and learn (GAL) and incremental supervised neural network (ISNN) are compared under two feature extraction methods (moment of grey level histogram (MGH) and two dimension continuous wavelet transform (2D-CWT)). Neural network and SVM based classifiers22 are compared to check which classifier has better performance. Similarly, different classifiers23-24 are compared for checking performance results. In this paper KNN, GAL and ISNN under MGH, 2D-CWT and hybrid are comparatively analyzed to find out best combination of classifier and feature extraction scheme. MethodologyIn the proposed work kNN, GAL and ISNN are compared with each other as classifiers under MGH, 2D-CWT and hybrid Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 2 feature extraction method. According to recent work ISNN performs better than GAL but according to the proposed work GAL results are better than ISNN by comparing their no of nodes, computational time and performance. It can also be seen from results that kNN is better classifier than GAL and GAL is better than ISNN by comparing their computational load and performance. The performance evaluation is given on the basis of four modalities which are: US bladder image, US phantom image, CT image and MRI. For accurate performance, the results are taken on the basis of 11 images of MRI modality. The proposed work is expressed diagrammatically in figure 1. The step wise explanation of proposed work is as follows: Figure-1 Specific Processing Blocks Training Point Selection: In the segmentation phase first step is the selection of training points from the original image. Here 100 training points are selected then Select the points from each class in such a way e.g. if image has two classes (1 and 2) then select half points from class 1 and remaining half from class 2. Selection is the most important step in segmentation process. If points will not select correctly then segmentation cannot be performed accurately. Feature Extraction: After selecting points, the second step is extraction of features by using three feature extraction methods which are 2D-CWT, MGH, and hybrid (the combine version of both 2D-CWT and MGH) as competitors 20-21. Extract 9 feature vectors from the test data (original image) and also from the training data. Here statistical moments are use for feature extraction. The equation for the th order moments is as follows: (1) Where, m = Mean intensity. zi = Random variable intensity. P (zi) = Histogram of the intensity levels in a region. L = Possible intensity levels. Here, 9 statistical moments are used for feature extraction which are: Mean (2) Standard Deviation (3)Smoothness = 23 1 (4)Third Moment= .) (5) Uniformity= (6) Entropy=))).Log (7) (8) (9) (10)Feature vector in MGH is given below: XT = [Fm1, Fm2, Fm3, Fm4, m5, Fm6, Fm7, Fm8, m9], Fm1=�measures average intensity, m2=�measures average contrast, Fm3=�measures smoothness, m4=�measures skewness of histogram, Fm5=�measures uniformity in histogram, Fm6=�measures randomness, Fm7, Fm8, m9=�having least information and for completeness of feature vector dimension. 2D-continuous wavelet transform CWT splits a continuous time function in to wavelets. It has the ability to create a time-frequency representation of an image for getting more information. CWT evaluation is macro based (not pixel based). Here scale parameter is used for transformation. Scaling function is responsible for improving the coverage of the wavelet spectrum. At high scale value, image components having low frequency are fitted with rich and opposite is the case at low scale value. 2D-CWT is applied (by Gaussian wavelet) for eight different scale values to the original image such as 1.0, 1.6, 2.6, 3.9, 4.0, 5.0, 5.4 and 7. Time and frequency domain equations for 2D-CWT 20 are given in equation 11 and 12 below respectively. dxdycwt - - (11) Input Image Training point selection (Expert) Feature Extraction (2DCWT, Hybrid, MGH) Supervised Classification (KNN, GAL, ISNN) Expert Satisfaction NO Segmented Image YES Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 3 )2,1)2,1)2,1swswcwt (12) In above equations 11 and 12, ‘a’ and ‘b’ are translation parameters and ‘s’ is a scale parameter for wavelet . Whereas, ‘x’ and ‘y’ are spatial domain coordinates and w1, w2 are frequency domain coordinates. Hybrid features are formed by combining both 2D-CWT and MGH features. Nine dimensional hybrid features vector is formed by combining first five features from 2D-CWT and remaining four features from MGH features. First five features of 2D-CWT correspond to original image plus features generated using four different scale parameters as first features carry much information. The scale parameter values are 1, 1.6, 2.6 and 3.9 which generate four filtered transformed images. The remaining four features are taken from MGH. These features are selected based on their high information content. The important features that carry much information compared to others in MGH are mean, standard deviation, uniformity and entropy. In figure 2, MGH, 2D-CWT and hybrid features are shown: (a) MGH Features (b) 2D-CWT Features for eight different scale values (c) Hybrid Features Figure-2 Features extracted by three feature extraction methods (a) MGH , (b) 2D-CWT and (c) Hybrid features Classification: After feature extraction, the classification process takes place. GAL, ISNN and kNN are used as classifiers. Classification process has two phases namely the training phase and testing phase. The data is trained and weights are assigned to that data in the training process, then labels are assigned to the whole original image in the testing process. Hence, this is the reason to use supervised classifiers in which expert chooses training points from each class of input image which is to be segmented. Here kNN, GAL and ISNN are competitor classifiers to be used for classification. The performance of kNN is much superior to that of the other two classifiers used in this work. It is a non-parametric classifier which is defined in equation 13 as: p() = k/LV (13) Where L are available training samples, is the test object (input feature vector) and p() is the probability around x. Construct the region around and then count the number of samples in this region. k is the no of samples (or no of neighbors) and V is the volume in that region. In kNN, fix the count k and determine V. An object is classified by a more votes of its neighbors. K is always a positive integer and usually small. Here we use k=1. ISNN is a two layer neural network which is proposed by Enhui et al. 22. In ISNN, classification takes place by assigning weights to the training data during training and then labels are assigned to the test data (original image) by finding minimum distance between the weights of the test data and training data. Grow and learn GAL is same as ISNN. The only difference is that when class of winner node is equal to the class of the input vector there is no increment in weights. Enhui et al. 22 claimed that ISNN is better than GAL by comparing the segmentation time. But when we calculated the results of classification which is shown in the next section, GAL perform better than ISNN. GAL performance percentage is better than ISNN as less no of nodes generate during training and use less training and segmentation time as compared to ISNN. Results and Discussion In this section, segmentation of four modalities takes place. The modalities are segmented using supervised classifiers as kNN, ISNN and GAL under the transform and moment based feature extraction namely MGH, 2D-CWT and hybrid feature extraction method. The simulation is performed on IBM Compatible, Intel Pentium IV PC by using Matlab 7.10 on windows 2007.As there is supervised classification usage therefore 100 training points are selected from the original data of each modality by an expert. Select the points equaly from each class of all modalities (except phantom image where 50 points are selected from liquid medium to reduce noise). There are nine dimensional feature vectors therefore the size of training data is 100*9 and size of training labels is 100*1. The size of test data is (no. of rows of image) *(no. of columns of image) *9. US bladder is segmented in to two classes (inner (1) and outer side (2) of bladder). Therefore 50 training points are selected from class 1 and 50 from 2. US phantom is segmented in to three classes. In phantom image, 25 points are selected from class 1 and 2 whereas 50 points are selected from class 3. In both modalities training points are select equally from each class. Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 4 (a) (b) (c) (d) Figure-3 (a) US bladder image, (b) US phantom image, (c) CT head image, (d) MR head image Segmented results are shown in figure 4, 5, 6 and figure 7. Note that three classifiers are analyzed with three different feature extraction schemes. (a) (b) (c) (d) (e) (f) ™®(g) (h) (i) Figure-4 Segmented US bladder image by (a) KNN, (b) GAL, (c) ISNN using hybrid; (d) KNN, (e) GAL, (f) ISNN using 2D-CWT; (g) KNN, (h) GAL, (1) ISNN using MGH (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure-5 Detected US phantom image by (a) KNN, (b) GAL, (c) ISNN using hybrid; (d) KNN, (e) GAL, (f) ISNN using 2D-CWT; (g) KNN, (h) GAL, (1) ISNN using MGH (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure-6 Segmented CT head image by (a) KNN, (b) GAL, (c) ISNN using hybrid; (d) KNN, (e) GAL, (f) ISNN using 2D-CWT; (g) KNN, (h) GAL, (1) ISNN using MGH Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 5 (a) (b) (c) (d) (e) (f) (g) (H) (i) Figure-7 Segmented MR head image by (a) KNN, (b) GAL, (c) ISNN using hybrid; (d) KNN, (e) GAL, (f) ISNN using 2D-CWT; (g) KNN, (h) GAL, (1) ISNN using MGH Comparison of GAL and ISNN under MGH, 2D-CWT and Hybrid at different iterations: The test images are segmented using GAL and ISNN at different iterations. In the existing work, iteration no of GAL and ISNN were selected as 2000, and 10,000 respectively 22. Therefore GAL and ISNN were not performing efficiently as too much computational time were consuming. Therefore after testing the results at different iterations, the suggested and tested iteration would be 20, 50 and 100 for both GAL and ISNN because at these iterations, GAL and ISNN are efficiently perform and use less training and segmentation time. Table 1 , 2 and 3 shows the comparison of GAL and ISNN at 20 , 50 and 100 iterations. From the above tables it can be observed that the iteration no. 20, 50 and 100, results are more accurate. It is also observed that, the GAL is better than ISNN using all selected iterations as GAL generates less number of nodes than ISNN. Therefore GAL has good performance and low computational time than ISNN under different feature extraction schemes. Performance analysis of KNN, GAL, ISNN under MGH, 2D-CWT, hybrid: The Performance results of segmentation for all four modalities are evaluated on the basis of one image. In the next section for identify more real performance of MR image we evaluate performance on the base of 11 MR images. Performance is the percentage comes by comparing training labels with testing labels. Here we also compare each classifier and feature extraction methods by comparing performance, computational time (both training time and segmentation time). Table-1 Comparison of GAL and ISNN at 20 Image ANN Features TT(sec) ST(sec) Total time(s) (ST+TT) NON Performance (%) usbladder GAL Mgh 0.061746 2.495491 2.557237 6 100 ISNN Mgh 0.073545 2.502339 2.575884 5 100 GAL 2d - cwt 0.061578 2.448884 2.510462 3 100 ISNN 2d-cwt 0.073083 2.467418 2.540501 3 100 GAL Hybrid 0.061922 2.585885 2.647807 4 100 ISNN Hybrid 0.073207 2.629772 2.702979 5 100 usphantom GAL Mgh 0.062149 2.210711 2.272860 eq 5 Uneq 8 98 ISNN Mgh 0.072470 2.136215 2.208685 6 8 96 GAL 2d - cwt 0.062934 2.238110 2.301044 5 6 99 ISNN 2d-cwt 0.073083 2.293207 2.366290 6 12 99 GAL Hybrid 0.062992 2.184066 2.247058 5 9 99 ISNN Hybrid 0.073174 2.177684 2.257058 6 11 99 CT head GAL Mgh 0.068176 2.132186 2.200362 19 93 ISNN Mgh 0.092177 2.177600 2.269777 21 93 GAL 2d-cwt 0.064881 2.041324 2.106205 18 99 ISNN 2d-cwt 0.078795 2.062013 2.140808 19 96 GAL Hybrid 0.064251 2.028647 2.092898 15 97 ISNN Hybrid 0.076433 2.086171 2.162604 20 97 M head GAL Mgh 0.065476 2.086582 2.152058 23 93 ISNN Mgh 0.079302 2.122091 2.201393 29 91 GAL 2d-cwt 0.063322 2.006598 2.069920 15 99 ISNN 2d - cwt 0.073378 1.993947 2.067325 15 97 GAL Hybrid 0.063661 2.024473 2.088134 9 98 ISNN Hybrid 0.075210 2.042749 2.117959 10 96 Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 6 Table-2 Comparison of GAL and ISNN at 50 Image ANN Features TT(sec) ST(sec) Total time(s) (ST+TT) NON Performance (%) usbladder GAL Mgh 0.152369 2.460263 2.612632 5 100 ISNN Mgh 0.169764 2.480009 2.649713 6 100 GAL 2d-cwt 0.149049 2.462202 2.611251 4 100 ISNN 2d-cwt 0.165828 2.464427 2.630255 6 100 GAL Hybrid 0.149489 2.460621 2.610110 4 100 ISNN Hybrid 0.167519 2.457495 2.625014 5 100 usphantom GAL Mgh 0.158182 2.165314 2.323496 eq 4 Uneq 7 98 ISNN Mgh 0.173404 2.160697 2.334101 6 9 96 GAL 2d - cwt 0.150029 2.132721 2.282750 4 5 99 ISNN 2d-cwt 0.166719 2.150088 2.316799 7 12 99 GAL Hybrid 0.156568 2.154378 2.310946 5 7 99 ISNN Hybrid 0.175836 2.029196 2.205032 6 12 99 CT head GAL Mgh 0.163830 2.064177 2.228007 13 93 ISNN Mgh 0.183346 2.085147 2.268493 22 91 GAL 2d-cwt 0.158120 2.083779 2.241899 14 99 ISNN 2d - cwt 0.172598 2.098529 2.271127 21 96 GAL Hybrid 0.161238 2.038472 2.19971 18 99 ISNN Hybrid 0.175836 2.029196 2.205032 20 96 MR head GAL Mgh 0.165803 2.025049 2.190852 24 90 ISNN Mgh 0.184973 2.098944 2.283917 34 90 GAL 2d - cwt 0.151466 1.984369 2.135835 8 99 ISNN 2d-cwt 0.170414 2.012171 2.182585 12 94 GAL Hybrid 0.152088 1.999095 2.151183 10 98 ISNN Hybrid 0.175413 2.077108 2.252521 10 96 Table-3 Comparison of GAL and ISNN at 100 Image ANN Features TT(s) ST(s) Total time(s) (ST+TT) NON Performance (%) Usbladder GAL Mgh 0.299269 2.504400 2.803669 5 100 ISNN Mgh 0.324763 2.511745 2.836508 6 100 GAL 2d - cwt 0.298291 2.488165 2.786456 4 100 ISNN 2d-cwt 0.329013 2.563743 2.892756 6 99 GAL Hybrid 0.300438 2.492378 2.792816 3 100 ISNN Hybrid 0.322719 2.450499 2.773218 3 99 Usphantom GAL Mgh 0.301321 2.183727 2.485048 eq 4 uneq 7 98 ISNN Mgh 0.320866 2.208548 2.529414 4 9 98 GAL 2d-cwt 0.301311 2.167776 2.469087 5 8 99 ISNN 2d - cwt 0.326056 2.161553 2.487609 10 8 98 GAL Hybrid 0.297357 2.157490 2.454847 5 8 99 ISNN Hybrid 0.321572 2.140950 2.462522 6 13 98 CT head GAL Mgh 0.315699 2.078036 2.393735 18 94 ISNN Mgh 0.349238 2.118836 2.468074 28 91 GAL 2d - cwt 0.309857 2.078821 2.388678 14 98 ISNN 2d-cwt 0.342323 2.072783 2.415106 21 92 GAL Hybrid 0.322229 2.114087 2.436316 18 97 ISNN Hybrid 0.361032 2.209156 2.570188 15 92 MR head GAL Mgh 0.320556 2.139310 2.459866 23 91 ISNN Mgh 0.345737 2.112361 2.570459 29 91 GAL 2d-cwt 0.300305 1.986979 2.287284 9 99 ISNN 2d - cwt 0.330680 2.024693 2.324998 14 96 GAL Hybrid 0.321592 2.130531 2.452123 15 99 ISNN Hybrid 0.350029 2.144014 2.494043 10 97 Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 7 Table-4 Performance Comparison of KNN, GAL, ISNN under MGH, 2D-CWT, hybrid for four different images Image ANN Features Total time(sec) Performance (%) usbladder kNN Mgh 0.340182 100 GAL Mgh 2.612632 100 ISNN Mgh 2.649713 100 kNN 2d-cwt 0.294831 100 GAL 2d - cwt 2.611251 100 ISNN 2d-cwt 2.630255 100 kNN Hybrid 0.287111 100 GAL Hybrid 2.610110 100 ISNN Hybrid 2.625014 100 usphantom kNN Mgh 0.266570 100 GAL Mgh 2.323496 98 ISNN Mgh 2.334101 96 kNN 2d-cwt 0.251623 100 GAL 2d - cwt 2.282750 99 ISNN 2d - cwt 2.316799 99 kNN Hybrid 0.282517 100 GAL Hybrid 2.310946 99 ISNN Hybrid 2.205032 99 CT head kNN Mgh 0.465149 100 GAL Mgh 2.228007 93 ISNN Mgh 2.268493 91 kNN 2d - cwt 0.401738 100 GAL 2d-cwt 2.241899 99 ISNN 2d - cwt 2.271127 96 kNN Hybrid 0.394566 100 GAL Hybrid 2.2.19971 99 ISNN Hybrid 2.205032 96 MR head kNN Mgh 0.421762 100 GAL Mgh 2.190852 90 ISNN Mgh 2.283917 90 kNN 2d - cwt 0.387594 100 GAL 2d - cwt 2.135835 99 ISNN 2d-cwt 2.182585 96 kNN Hybrid 0.400457 100 GAL Hybrid 2.151183 99 ISNN Hybrid 2.252521 99 The results come by finding and comparing performance and computation time of KNN, GAL and ISNN under MGH, 2DCWT and hybrid feature extraction method as shown in table 4, it is clearly observed that KNN has 100% performance in the segmentation of all 4 modalities and it has less computational time as compared to GAL and ISNN. Similarly GAL performance is better than ISNN and use less computational time than ISNN. In the above comparison we clearly observe that KNN is better than GAL and ISNN. GAL is better classifier than ISNN under all feature extraction methods as discuss above. Hybrid and 2D-CWT both perform equally and better than MGH. Hybrid has slightly better performance than 2D-CWT. 2D-CWT is good in performance and use less computational time than MGH. Performance Evaluation: In the previous session we evaluate the performance of all images on the basis of 1 image but for more accurate results and to identify the original performance we evaluate the performance on the bases of 11 image comparison. Performance is evaluated using Leave-One-out Cross-Validation technique. For more accurate performance results, the 11 MRI images are analyzed in order to come up with solid conclusion. There are total 11 patients MR real images of size 128×128 taken by Siemens MRI system in Abrar CT and MRI centre, Peshawar road, Rawalpindi. Leave-one out cross validation procedure is defined as in it one image is kept for testing and remaining 10 images for training i.e. classifiers are trained using training data of other 10 images and 11th image is tested. In simple words in first step 1st image will be a test data and 2 to 11images will be a training data, then in second step 2nd image will be the test data and 1,3 to 11 images will be the training data and so on. In each step compare the resultant segmented image labels with its known training point’s labels and performance percentage is Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 8 calculated for each image. Then for each image we would have classifier’s performance. Quantitative and Qualitative performance analysis of classifiers for 11 MR images Using LOOCV: Table 5 shows the comparison of classifiers (kNN, GAL and ISNN) using Hybrid, 2DCWT and MGH features for 11 MR images. Performance percentages are calculated for each match with the visual results as it is mentioned before. Table 6 is constructed by averaging performance percentages. Computational time is also shown for each classifier which is equal to training time plus segmentation time. It is clearly visible that for Hybrid and 2D-CWT features kNN performs best in time compared to GAL and ISNN and also its percentages are higher than GAL and ISNN. In table 6 the overall results which numerically match with the visual results. Priority wise classifiers in taking less computational time for Hybrid and 2D-CWT features are kNN, GAL and ISNN. I give these results at 50 iterations. Since MGH features are not robust therefore kNN performance is not good. Some images shows less performance of kNN than as compared to GAL and ISNN but overall result i.e. kNN is better than GAL and GAL is better than ISNN. Overall results i.e. kNN is better than GAL and GAL is better than ISNN. Also it shows that percentages for hybrid and 2D-CWT features are higher than those for MGH features which show the effectiveness of Hybrid and 2D-CWT features. Table 8 shows the overall results. Hybrid and 2D-CWT Features are equal in performance. The difference lies between them in their computational time. 2D-CWT features take less time than Hybrid features because Hybrid has to wait for execution of MGH function also. According to their computational time, 2D-CWT features are executed faster than Hybrid and MGH. Priority wise features better in computational time are 2D-CWT, MGH and Hybrid. Priority wise features better in performance are Hybrid, 2D-CWT and MGH round of performance is better than GAL and ISNN. Computational time of kNN is also much healthier than both neural networks classifiers. Quantitative and Qualitative performance analysis of feature extraction methods for 11 MR images: Features are compared by fixing classifiers. Table 7shows the comparison of features (Hybrid, 2D-CWT and MGH) under kNN, GAL and ISNN. Performance percentages are calculated for each image using Leave One out Cross Validation technique. Fig. 8 is constructed using tables 5 and 6. Percentages in figure 8 are calculated by taking mean of 11 percentages (11 images) in each table. Table-5 Performance comparison of classifiers using hybrid, 2D-CWT and MGH features Performance % using Hybrid Performance % using 2D-CWT Performance % using MGH Image kNN GAL ISNN kNN GAL ISNN kNN GAL ISNN 1 95 90 88 95 81 79 82 84 83 2 92 88 84 96 90 90 84 82 82 3 97 90 88 95 86 88 86 90 86 4 97 96 90 100 92 89 90 89 86 5 96 88 91 96 92 84 89 90 81 6 96 96 90 99 91 89 84 88 90 7 97 88 87 95 95 88 80 83 80 8 96 92 92 95 91 91 89 90 84 9 99 93 94 100 95 93 95 86 81 10 96 94 90 97 93 90 88 81 81 11 92 91 90 95 92 87 89 82 81 Avg 96 91 89 97 91 89 87 86 83 Table-6 Concluded Performance Comparison of Classifiers Image Feature Classifier Training+ Segmentation Time (seconds) Performance (%) MR head (128×128) Hybrid kNN 0.366181 96 Hybrid GAL 2.218106 91 Hybrid ISNN 2.259258 89 2D-CWT kNN 0.361703 97 2D-CWT GAL 2.171699 91 2D-CWT ISNN 2.166921 89 MGH kNN 0.373317 87 MGH GAL 2.249207 86 MGH ISNN 2.269919 83 Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 9 Figure-8 Comparison of classifiers Table-7 Performance comparison of features under Knn, GAL and ISNN Performance % under KNN Performance % under GAL Performance % under ISNN Image Hybrid 2D-CWT MGH Hybrid 2D-CWT MGH Hybrid 2D-CWT MGH 1 95 95 82 90 81 84 88 80 83 2 92 96 84 88 90 82 84 90 82 3 97 95 86 90 86 90 88 89 86 4 97 100 90 96 92 89 90 89 86 5 96 96 89 88 92 90 91 85 81 6 96 99 84 96 91 88 90 89 90 7 97 95 80 88 95 83 87 89 80 8 96 95 89 92 91 90 92 91 84 9 99 100 95 93 95 86 94 93 81 10 96 97 88 94 93 81 90 92 81 11 92 95 89 91 92 82 90 88 81 Avg 96 97 87 91 91 86 89 89 83 Table-8 Concluded performance Comparison of features under Knn, GAL and ISNN Image Feature Classifier Feature Extraction Time (seconds) Performance (%) MR head (128×128) Hybrid kNN 3.7420 96 2D-CWT kNN 0.741782 97 MGH kNN 3.000182 87 Hybrid GAL 3.7420 91 2D-CWT GAL 0.741782 91 MGH GAL 3.000182 86 Hybrid ISNN 3.7420 89 2D-CWT ISNN 0.741782 89 MGH ISNN 3.000182 83 Reason for superiority of 2D-CWT features is that it allows multi-resolution texture analysis due to which noise is reduced in segmentation. Reason for superiority of Hybrid features is that it is a combined version of both 2D-CWT and MGH important features.Figure 9 shows the overall result of feature extraction methods as MGH, 2D-CWT and Hybrid and classifiers as Knn,GAL ans ISNN performance comparison. Also figure 9 is the graphical explanation of table 8. Conclusion In this study medical image segmentation of four modalities as us bladder, us phantom, CT head and MR Head images are evaluated using kNN, GAL and ISNN as classifiers under MGH, 2D-CWT and hybrid feature extraction methods. In the literature work it is claimed that ISNN is better than GAL but on the basis of this work we can conclude that GAL is better than ISNN because GAL uses less training time and segmentation time and generates less no of nodes during training than ISNN. The said observation is validated by testing GAL and ISNN at different iterations, so the preferable iterations would be 20, 50 and 100. The segmentation performance result of classifiers are evaluated on the bases of 11 MRI images and the conclusion can be drawn from the observed results that kNN is superior and outperformed both GAL and ISNN. For segmentation, GAL is better than ISNN under MGH, 2D-CWT and hybrid feature extraction methods. If there is comparison between features extraction methods, than hybrid and 2D-CWT gives much better performance than MGH. Research Journal of Recent Sciences ______________________________________________________________ ISSN 2277-2502Vol. 2(6), 1-10, June (2013) Res. J. Recent Sci. International Science Congress Association 10 Figure-9 Comparison of features References 1.Zhy C.M., Gu Guo-chang, Liu Hai-bo, Shen Jing and Yu Hualong, Segmentation of Ultrasound Image Based on Texture Feature and Graph Cut, International Conference on Computer Science and Software Engineering, 795-798, 12-14 (2008)2.Chi-Man Pun and Hong-Min Zhu, Image Segmentation using Discrete Cosine Texture Feature, International Journal of Computers, 4(1), 19-26 (2010)3.Chen Jia-xin and Liu Sen, A medical image segmentation method based on watershed transform, The Fifth International Conference on Computer and Information Technology, 2005, CIT 2005, 634- 638, (2005)4.Zhy Chang-ming, Gu Guo-chang, Liu Hai-bo, Shen Jing and Yu Hualong, Segmentation of Ultrasound Image Based on Texture Feature and Graph Cut, International Conference on Computer Science and Software Engineering, 1, 795-798 (2008)5.Grau V., Mewes A.U.J., Alcaniz M., Kikinis R. and Warfield S.K., Improved watershed transform for medical image segmentation using prior information, IEEE Transactions on Medical Imaging, 23(4), 447-458 (2004)6.Waqas Haider, Muhammad Sharif Malik, Mudassar Raza, Abdul Wahab, Izhar Ahmed Khan, Umar Zia, Jawad Tanveer, Hadia Bashir, A Hybrid Method for Edge Continuity based on Pixel Neighbor Pattern Analysis (PNPA) for Remote Sensing Satellite Images, International Journal of Communications, Networks and System Sciences,5( 9A), 624-630 (2012)7.Regina Pohle and Klaus D., Toennies, Segmentation of medical images using adaptive region growing, Proceedings of SPIE (Medical Imaging), San Diego; 4322 1337-46 (2001)8.Sahoo P.K., Soltani S. and Wong A.K.C., A survey of thresholding techniques, Comput. Vis. Graph. Im. Proc., 41, 233–260 (1988)9.LIANG Xiao-yun, LUO Li-min and ZENG Wei-ming, Level Set Method Based Medical Image Segmentation (2003)10.Middleton and R. Damper, Segmentation of magnetic resonance images using a combination of neural networks and active contour models, Med. Eng.Phys., 26, 71-86 (2004)11.Jianzhong Wang, Di Liu, Lili Dou, Baoxue Zhang, Jun Kong and Yinghua Lu, Graph theory based algorithm for magnetic resonance brain images segmentation, 8th IEEE International Conference on Bioinformatics and Bioengineering, BIBE 2008, 1-5, (2008)12.Jain A.K. and Dubes R.C., Algorithms for clustering data, Prentice Hall, (1988)13.Lefebvre F., Berger G. and Laugier P., Automatic detection of the boundary of the calcaneus from ultrasound parametric images using an active contour model-clinical assessment, IEEE Trans. Med. Imag., 17(1) 45–52 (1998)14.Li H.D., Kallergi M., Clarke L.P., Jain V.K. and Clark R.A., Markov random field for tumor detection in digital mammography, IEEE T. Med. Imag., 14, 565–576 (1995)15.McInerney T. and Terzopoulos D., Deformable models in medical image analysis: a survey, Med. Im. Anal.,, 91–108 (1996)16.Bai Peirui, Fu Caihong, Cao Maoyong and Han Yan, Improved Mean Shift Segmentation Scheme for Medical Ultrasound Images, International Conference on Bioinformatics and Biomedical Engineering (iCBBE), 1-4, (2010)17.Zafer Iscan, Ayhan Yuksel, Zumray Dokur, Mehmet Korurek and Tamer Olmez Medical Image Segmentation with transform and moment based features, Digital Signal Processing, 19(5), 890-901 (2009)18.Parkkinen J., Selkainaho K. and Oja E., Detecting texture periodicity from co-occurrence matrix, Pattern Recogn. Lett., 11, 43–50 (1990)19.Dasarathy B. and Holter E., Image characterization based on joint gray level run length distributions, Pattern Recogn. Lett., 12 497–502 (1991)20.Wang N. and Lu C., Two-dimensional continuous wavelet analysis and its application to meteorological data, Journal of Atmospheric and Oceanic Technology, 27, 652-666 (2010)21.Ghazali K.H., Mansor M.F., Mustafa M.M. and Hussain A., Feature Extraction Technique using Discrete Wavelet Transform for Image Classification, 5th Student Conference on Research and Development, 1-4, (2007) 22.Enhui Zheng, Ping Li and Zhihuan Song, Performance analysis and comparison of neural networks and support vector machines classifier, Intelligent Control and Automation, (2004)23.McGunnigle G. and Chantler M.J., Comparison of three rough surface classifiers, Vision, Image and Signal Processing, IEE Proceedings-, 149(5), 263-271 (2002)24.Rezaei S., Tavakolian K. and Naziripour K., Comparison of Five Different Classifiers for Classification of Mental Tasks, 27th Annual International Conference of the Engineering in Medicine and Biology Society, 6007-6010, 17-18 (2006)