An Accurate Plaque Characterization using Feed Forward Neural Network

An Accurate Plaque Characterization using Feed Forward Neural Network

Mr. Z. Ahamed Yasar 1 Prof. D. Sangeetha2

Department of ECE Department of ECE

Sri Krishna College of Engineering & Technology Sri Krishna College of Engineering & Technology Coimbatore, India Coimbatore, India

Abstract- Computer Aided Diagnosis CAD is useful in detection of carotid atherosclerosis into symptomatic or asymptomatic in cardiac health. With the help of patents CAD system the symptomatic versus asymptomatic plaque is classified in ultra sound images. The system involves four steps preprocessing, feature extraction, feature selection and classification using feed forward neural network. Preprocessing is used to remove unwanted portions in the image, feature extraction is used to extract the required image portion, feature selection is used to select the required matrix values for classification of images into symptomatic or asymptomatic and blood clot will be detected using feed forward neural network. Feed forward neural network classifier classifies symptomatic or asymptomatic plaque and detects the condition of plaque and blood clot. The system was evaluated with 93 ultra sound carotid images and obtained classification accuracy is 94.50%.

Keywords- Computer Aided Diagnosis (CAD), Region Of Interest, Discrete Wavelet Transform (DWT), Feed Forward Neural Network(FFNW).

1. INTRODUCTION

   Stroke is the one of the leading cause of death in the world. Stroke will disrupt the blood flow which block the oxygen supply to the part of brain cells and the cells will die with lack of oxygen supply to the cells. This disturbance is due to blockage or leakage in blood vessels. Stroke and heart disease are related to high blood pressure and atherosclerosis. Presence of plaque in blood vessels will block the blood flow and cause stroke or heart disease [2],[3],[4].

   For atherosclerotic cardiovascular diseases diagnosis of plaque can make real difference in the treatment. To reduce the risk of stroke and heart disease plaques are surgically removed from affected area [6],[7]. The diagnosis system, detection and carotid surgery leads to the quality of the successful treatment. Ultrasound imaging is cost effective and it will produce good result for medical imaging. Even though the ultrasound imaging has significant advantages in detecting cerebrovascular disease it still has some limitations. The correlation between histological evaluation of carotid plaques and ultrasonographic features are poor. Due to the low spatial resolution and ultrasound artifacts these limitations appear. By improving the ultrasonographic image quality using extraction of good features and adequate image

   preprocessing techniques the diagnosis of plaque may improve [8],[9].

   To detect the carotid atherosclerosis into symptomatic or asymptomatic using computer-aided diagnosis (CAD) is useful in the analysis of cardiac health. With the help of the patented CAD system called Atheromatic the symptomatic and asymptomatic plaque classification is done in carotid ultra sound images. This operation involves four steps 1.pre-processing 2.feature extraction using discrete wavelet transform 3.feature selection with averaging algorithms 4.classification using feed forward neural network.The block diagram of the classification system is shown in the fig 1. The ultra sound image is pre-processed and given for feature extraction. The feature extraction is done using discrete wavelet transform. From the feature extracted image feature selection is obtained. The feature selected values are fed to the feed forward neural network for symptomatic and asymptomatic plaque classification.

2. MATERIALS AND METHODS

   Fig 1 shows the proposed system block diagram. The ultrasound images are preprocessed and fed to feature extraction using DWT. The extracted features are selected and classified using feed forward neural network.

   1. Ultrasound images and preprocessing

      In this paper we use 93 ultra sound carotid plaque images out of that 46 were asymptomatic and 47 were symptomatic. Ultrasound imaging is the second popularly used technique after x-ray to map the human body and its functions.

      Ultrasound imaging uses same principle used by bats and sonar. In medical imaging the operation of x-ray is a sound wave that is passed on human body it will strike the body organs and bounce back a echo which consist of information about the patient body as image. This image consists of the information about body organs to diagnose the symptoms about several diseases.

      Ultrasound images

      Pre- processi ng

      Feature extraction

      Feature selection

      Feed forward neural network

      Fig 2 . (c) blood clot

      Fig 1. Block diagram for classification

      The images obtained from ultrasound scan of the human body may contain some unwanted portions which may not be useful for diagnosis of plaques. To remove the unwanted portions in image the region of interest (ROI) is calculated for the images to be processed on the system. The ROI calculated image constitutes less than 25% of the original image, so the classification process will become more easily by calculating ROI and it will be useful to detect the symptomatic and asymptomatic plaques with this method. Figure 2(a), 2(b) and 2(c) shows the symptomatic plaque, asymptomatic plaque and blood clot. Figure 3(a), 3(b), 3(c) shows the ROI selected symptomatic plaque, asymptomatic plaque and blood clot.

   2. Feature extraction

   For feature extraction 2-D DWT and its averaging algorithms are used in this paper. In image processing feature extraction is the form of dimensionality reduction, the dimensionality reduction is achieved when the input to an algorithm is too large to be processed in it. Feature extraction is the process of transforming the input data into set of features that can be applied to the classifier. The extracted features are properly chosen that the features should have relevant information from the input data in order to get reduced representation of input data instead of full size input.

   Fig 3 . (a) ROI selected symptomatic carotid image

   Fig 3(b) ROI selected asymptomatic carotid image

   Fig 3(c) ROI selected blood clot

   The detail and approximation coefficients are calculated using The equations (1) and (2).

   Ho=

   =

   [2 ]

   (1)

   Go=

   =

   2 (2)

   Fig 2. (a) symptomatic carotid image

   Fig 2. (b) asymptomatic carotid image

   In this paper we have evaluated the image with several wavelet functions. Out of those functions biorthogonal 3.1 (bior3.1) produced better result hence bior3.1 family is used in this work [5] .

   In 2-D DWT the input image rows are filtered with one high-pass filter and one low-pass filter combination. The high-pass filter output is known as detailed coefficient and low-pass filter output is known as approximation coefficient. These high pass filter and low-pass filter output are down-sampled to get the redundant sampled value which has high information content. With this operation the bandwidth gets reduced and the frequency resolution gets

   doubled. The down sampled output of high-pass and low- pass filter is filtered with sets of high and low-pass filters for columns. The first level 2D-DWT produces four matrices values namely Dp, Dv1, Dd1 and A1, the second level 2D-DWT produces four matrices in each of the obtained matrices Dp, Dv1, Dd1 and A1.The figure 4 and 5 shows DWT decomposition and image decomposition. The elements of thes matrices are intensity values and its number of elements is too high which cannot be used for classification directly. Therefore we used two averaging methods and energy of the intensity values to select the features used for classification in feed forward neural network. Two averaging methods and the energy of the intensity values are calculated using the equation (3),(4) and (5). The features obtained using these equations for symptomatic and asymptomatic plaques are shown in the table I.

   Avg Dp= (1/L×M)

   TABLE I .Symptomatic versus asymptomatic features obtained using DWT BIOR 3.1 wavelet function.

   Features	Symptomatic	Asymptomatic
   Avg Dp	0.1411	-0.4069
   Avg Dv1	0.0727	0.1427
   Energy(E)	0.0051	0.0045

   C .Feed Forward Neural Network

   Artificial neural networks are very popular image processing and data mining tool. The function of artificial neural networks is attempts to model the algorithm like human thought which can be efficiently run on a computer. The human brain is formed with neurons which send activation signals in the brain to control the human function,

   = | 1 , | =() |1 , |

   Avg Dv1= (1/L×M)

   = | 1 , | =() |1 , |

   Energy (E)=(1/L²×M

   (3)

   (4)

   likewise human brain the artificial neural network also contains artificial neurons which sends activation functions between one another and calculate the outputs according to their inputs. Therefore neural networks can be useful for several applications such as clustering, classification, function approximation, time series prediction, and descriptive modeling. In this paper the system is evaluated with feed forward neural network.

   = (1 , )² =()(1 , )²

   (5)

   Dd1

   Dp

   Dv1

   A1

   Image

   Fig 4. DWT decomposition

   A2 Dp	Dv2 Dd2

   Fig 5. image decomposition

   Fig 6 . feed forward neural network

   Informations in feed forward neural network will moves only in forward direction, from the input neurons to the output neurons through hidden neurons. To activate the inputs in network the input activation function is activated at input neurons and to get the output the output activation function is activated at the output neurons. The equation (6) and (7) shows the input and output activation functions. The classification operations are performed accurately in hidden neurons and output is transferred to output neurons. The neurons in each layers sends activation to the other neurons in next layers called weights. The weights consist of source and destination addresses which delivers the data to the proper destination the weights are given by Wi,j where i is the source and j is the destination.

   a(X)= b+i wixi (6) h(X)= g(a(X)) (7)

   In the classification of plaque and blood clot using feed forward neural network the symptomatic plaque and blood

   clot are given a value of 1 and the asymptomatic plaque and blood clot are given a value of -1. The feed forward neural network provides better classification accuracy by separating the 1 and -1 by a hyperplane.

   In this paper three sets of images are used, two sets are used for training and one set is used for testing. The input image is compared with testing set of images, if the input image value is -1 or near to -1 the feed forward neural network will classify it as asymptomatic and if the input image value is 1 the feed forward neural network will classify it as symptomatic.

3. RESULTS

   classification of symptomatic and asymptomatic plaques. The performance calculation is shown in the table II.

   TABLE II

   FFNW input	93
   True negative(TN)	42
   False negative (FN)	3
   True positive (TP)	44
   False positive (FP)	2
   Accuracy (%)	94.5
   Specificity (%)	95
   Sensitivity (%)	93.6
   PPV (%)	95.6
   NPV (%)	93

   Performance calculation of symptomatic and asymptomatic plaques

   1. Significant features

      The feature selected values such as horizontal coefficient, vertical coefficient and energy are fed to the feed forward neural network for detection of unknown class. The features are obtained using DWT with bior3.1 wavelet.

   2. Classification results

   In this paper we compared the performance of several wavelet functions. The biorthogonal bior3.1 function performed better compared to other wavelet functions. In this paper three sets of images are used, two sets are used for training and one set is used for testing. In this method 93 images are used for testing among those 46 images were asymptomatic and 47 images were symptomatic. These images are classified using feed forward neural network to find the unknown class.

   The feed forward neural network performs better result than the previous classification methods. In this method its performance is calculated by number of true positive (TP), true negative (TN), false positive (FP) and false negative (FN). True positive (TP) is the number of symptomatic samples identified as symptomatic. True negative (TN) is the number of asymptomatic plaques classified as asymptomatic. False negative (FN), is the number of symptomatic samples classified as asymptomatic, and false positive (FP) is the number of asymptomatic samples classified as symptomatic. The probability that the technique will classify the symptomatic cases which are calculated by the formula TP/ (TP + FN) called as sensitivity. The probability that the technique will identify the asymptomatic cases which are calculated by the formula TN/ (TN +FP) called as specificity. The ratio of the number of correctly classified samples to the total number of samples which are calculated by the formula (TP + FP)/ (TP

   + FP + TN + FN) called as accuracy. PPV is the consideration of symptomatic subjects among those who were labeled symptomatic by the technique. NPV is the consideration of asymptomatic subjects among those who were labeled asymptomatic by the technique. Our result shows that the feed forward neural network offers better result than the previous methods with accuracy of 94.5%, sensitivity of 93.61% and specificity of 95% for the

4. CONCLUSION

   It is very difficult to diagnose the blood clot symptomatic and asymptomatic plaques from ultrasound images using image processing. Only experienced surgeons can able to detect the difference between blood clot, symptomatic and asymptomatic plaque in ultrasound scanned images. Our present system diagnoses some of the difficulty in classification of plaques and blood clot. It can be used as a diagnostic tool in modern clinical practice in detection of carotid plaques. Our system uses DWT for feature extraction and diagnoses the classes with accuracy, specificity and sensitivity. With this method we obtained highest classification accuracy of 94.5%. This accuracy is higher than the methods implemented in similar studies in the literature. Therefore we believe the proposed method can be considered as an effective technique which serves as a effective tool for the vascular surgeons in patients for treatment of risky stenosis. To improve the accuracy furthermore, our feature work includes in studying more feature extraction methods to improve accuracy.

5. REFERENCES

/ol>

1. The Internet Stroke Center. [Online]. Available: http://www.strokecenter. org/patients/stats.htm

2. N. R. Sims and H. Muyderman, Mitochondria, oxidative metabolism and cell death in stroke, Biochim. Biophys. Acta., vol. 1802, no. 1, pp. 8091, Jan. 2010.

3. D. R. Labarthe, Epidemiology and Prevention of Cardiovascular Diseases: A Global Challenge. Gaithersburg, MD: Aspen, 1998.

4. Y. Nagai, K. Kitagawa, M. Sakaguchi, Y. Shimizu, H. Hashimoto, H. Yamagami, M. Narita, T. Ohtsuki, M. Hori, and M. Matsumoto, Significance of earlier carotid atherosclerosis for stroke subtypes, Stroke,

   vol. 32, no. 8, pp. 17801785, Aug. 2001.

5. U. Rajendra Acharya, Oliver Faust, S. Vinitha Sree, Filippo Molinari, Luca Saba, Andrew Nicolaides, and Jasjit S. Suri, Senior Member, IEEE An Accurate and Generalized Approach to Plaque Characterization in 346 Carotid Ultrasound Scans IEEE transactions on instrumentation and measurement, vol. 61, no. 4, april 2012.

6. European Carotid Surgery Trialists Collaborative Group, Randomised trial of endarterectomy for recently symptomatic carotid stenosis: Final results of the MRC European Carotid Surgery Trial (ECST), Lancet, vol. 351, no. 9113, pp. 137913871998, May.

7. Beneficial effect of carotid endarterectomy in symptomatic patients with high-grade carotid stenosis, New England J. Med., vol. 325, no. 7, pp. 445453, Aug. 1991.

8. T. S. Hatsukami, M. S. Ferguson, K. W. Beach, D. Gordon, P. Detmer, D. Burns, C. Alpers, and D. E. Strandness, Carotid plaque morphology and clinical events, Stroke, vol. 28, no. 1, pp. 95100, Jan. 1997.

9. D. W. Droste, M. Karl, R. M. Bohle, and M. Kaps, Comparison of ultrasonic and histopathological features of carotid artery stenosis, Neurol. Res., vol. 19, no. 4, pp. 380384, Aug. 1997.

10. M. B. Griffin, E. Kyriacou, C. Pattichis, D. Bond, S. K. Kakkos, M. Sabetai, G. Geroulakos, N. Georgiou, C. J. Doré, and A. Nicolaides, Juxtaluminal hypoechoic area in ultrasonic images of carotid plaques and hemispheric symptoms, J.Vascular Surg., vol. 52, no. 1, pp. 69 76, Jul. 2010.

11. Anita Jindal, Renu Dhir, Rajneesh Rani Diagonal Features and SVM Classifier for Handwritten Gurumukhi Character Recognition International Journal of Advanced Research in Computer Science and Software Engineering, Volume 2, Issue 5, May 2012, ISSN: 2277 128X.

12. C. I. Christodoulou, C. S. Pattichis, M. Pantziaris, and A. Nicolaides, Texture based classifcation of atherosclerotic carotid plaques, IEEE Trans. Med. Imag., vol. 22, no. 7, pp. 902912, Jul. 2003.

13. G. Simone, F.C. Morabito, R. Polikar, P. Ramuhalli, L. Udpa and S. Udpa Feature extraction techniques for ultrasonic signal classification International Journal of Applied Electromagnetics and Mechanics 15 (2001/2002) 291294 IOS Press.

14. J. Stoitsis, S. Golemati, K. S. Nikita, and A. N. Nicolaides, Characterization of carotid atherosclerosis based on motion and texture features and clustering using fuzzy c-means, in Proc. 26th Annu. Int. Conf. IEEE EMBS, San Francisco, CA, 2004, pp. 1407 1410.

15. J. Stoitsis, N. Tsiaparas, S. Golemati, and K. S. Nikita, Characterization of carotid atherosclerotic plaques using frequency- based texture analysis and bootstrap, in Proc. 28th Annu. Int. Conf. IEEE EMBS, New York, 2006, pp. 23922395.

16. K. R. Muller, S. Mika, G. Ratsch, K. Tsuda, and B. Scholkopf, An introduction to kernel based learning algorithms, IEEE Trans. Neural Netw., vol. 12, no. 2, pp. 181201, Mar. 2001.

17. Loizou C, Pattichis C, Istepanian R, Pantziaris M, Nicolaides A (2004) Atherosclerotic Carotid Plaque Segmentation. In: Proceedings of the 26th annual international IEEE EMBS conference, San Francisco, USA.

18. Pablo A. Estévez, Senior Member, IEEE, Michel Tesmer, Claudio A. Perez, Senior Member, IEEE, and Jacek M. Zurada, Fellow, IEEE Normalized Mutual Information Feature Selection ieee transactions on neural networks, vol. 20, no. 2, february 2009,pp. 189-201.

19. Pabitra Mitra, student member, IEEE, C.A Murthy, and Sankar k.Paul, Fellow IEEE Un supervised feature selection using feature similarity IEEE transaction on pattern analysis and machine intelligence, vol 24, no 3, mar 2002, pp. 301-312.

20. Wilhjelm JE, Gronholdt LM,Wiebe B, Jespersen SK, Hansen LK, Sillesen H (1998) Quantitative analysis of ultrasound B-mode images of carotid atherosclerotic plaque: correlation with visual classification and histological examination. IEEE Trans Med Imaging 17(6):910 922.