Sie sind auf Seite 1von 5

IJSRD - International Journal for Scientific Research & Development| Vol.

4, Issue 05, 2016 | ISSN (online): 2321-0613

Identification of Intra-Cardiac Masses based on Advanced ACM

Segmentation using Multiclass SVM Classifier
M. Archana1 Dr.G.Harinatha Reddy2
M.Tech. Student 2Head of Dept.
Department of Electronics and Communication Engineering
NBKRIST, Vidyanagar, India
Abstract one of important task in cardiac disease diagnosis an image. Subsequently, an Adaptive color LSMA mass
is to identify intra-cardiac masses in echocardiogram. Based segmentation to get desirable characteristics of good image
on Multi-class SVM classifier to identify the difference segmentation should exhibit with reference to gray-level
between the intra-cardiac masses in echo-cardiograph to get images. The Multiclass SVM classifier is described an initial
good accuracy. First, a region of interest is cropped to define contour of the mass and also involving the motion, the
the mass area. Then apply global de-noising by NLM to boundary and the texture features.
eliminate the noise without disturbing the original structure.
Subsequently, the contour of the mass and it connected mass II. SVM CLASSIFIER
atrial wall are identified with adaptive color segmentation Support vector machine (SVM) originally separates the
model. Finally the motion, the boundary as well as the binary classes (k = 2) with a maximum margin criterion and
texture features are processed by SVM classifier. Our it is an effectively extend for multi-class classification. The
proposed method gives good accuracy and sensitivity. Intra- main aim is to determining the location of decision
cardiac masses are the condition of increasing clinical boundaries based on a part of the training data. It is a
significance not only because of its potential complications supervised learning theory. A special property of SVM is to
but also lack of clinical evidence to guide clinicians in minimize the empirical classification error and maximizes
selecting optimal therapies. the geometric margin. Similarly, an unsupervised learning
Key words: Automatic Classification, Intra-Cardiac Masses, method is applied when data is unlabeled. It is required to
Echocardiography, Multiclass SVM Classifier find clustering of the data into groups and vice versa [12].
The SVM algorithm is constructing a separating hyper-plane
I. INTRODUCTION with the maximal margin. SVM can be more accurate on
The intra-cardiac masses are occurred in middle ages. These moderately unbalanced data. Hyper-plane is defined as
diseases are very hazardous in cardiovascular disease [1]. distinguish between the largest distance and the nearest
Still there is no better treatment for cardiac masses training-data point of any class [13].
especially in malignant forms. They are available only in The figure (1) consists of Video decomposition,
single case reports [2]. Intra-cardiac masses are mainly Preprocessing, Adaptive color LSMA mass segmentation,
classified as two types: a) Tumor and b) Thrombus. 75% of Feature extraction, Multiclass SVM classifier, Classification
Primary tumors are rare in adults are benign. About 75% of results.
cardiac myxomas are located in the left ventricular filling
and 25% are in the right ventricular filling [3]. Mostly,
Patients are present with asymptomatic subjects depending
on tumor site and infiltration. In these cases, based on
echocardiogram we can observe an abnormal image. The
important characteristics of primitive cardiac tumors are
firmness or friability, degree of invasiveness growth to
determining symptoms, the tumor's size and its location [4,
5]. Intra-cardiac thrombi can appear elsewhere in the body
and be identified in the heart while in motion or may
develop. It is related with diverse therapy options [6]. A
fuzzy rule-DSS is presented for the diagnosis of coronary
artery disease. Here the identification of similar
echocardiography appearances of two masses is difficult [7].
It is based on storing and updating the gray level histogram
of the picture elements in the window and it requires more Fig. 1: Flow chart of our proposed method.
time [8]. SRAD is used for images corrupted with speckled
noise. In this method Lee and Frost filters are utilized the III. PREPROCESSING
coefficient of variation in adaptive filtering [9]. A CAD
system is used for the selection of lung nodules in CT A. Video Decomposition
(Computed Tomography) based on the region growing RG Echocardiography is one of the accepted tools to identifying
algorithms and ACM (Active Contour Model) [10]. The K- the condition of cardiac patients. Two-dimensional
SVD is one of the methods in sparse representation to echocardiography images are used to estimate cardiac
eliminate the global de-speckling noise [11]. The proposed chamber size, wall thickness and motion, valve motion and
method may present with several useful innovations. Using the pericardium [14]. Video means multiple frames. Video
a globally de-noising with NLM to remove speckle noise in is converted into frames using matlab codes and it act as

All rights reserved by 387

Identification of Intra-Cardiac Masses based on Advanced ACM Segmentation using Multiclass SVM Classifier
(IJSRD/Vol. 4/Issue 05/2016/096)

video to frame converter [15]. When compared with motion edge operator is used to get an initial contour. Whenever the
heart in two successive frames then these texts and labels sub windows are in same intensity then iterations are ends to
are static. After subtracting two successive frames the static get final contour. Here the sigma value of both Gaussian
data is removed. Finally, the original is image is cropped by filter and gradient threshold are 0.15 [26]. The closest initial
scanned region [16]. contour is easily identified with the help of SVM classifier.
In an iterative process to identify the features of an image.
B. Automatic ROI Selection
A special case of ROI is the change in the shape of the ROI V. FEATURE EXTRACTION
similar to the change in the location of the edges based on
the edge tracking algorithm. In order to select mass area a Feature extraction is defined as locating those pixels have
ROI containing the mass and its surrounding tissues are some distinctive characteristics in an image. Classification
defined [17]. The cardiac chamber has its own unique system is an important roles in feature extraction.
features with lower intensities and uniformed distributions Depending on the two rules such as the motion and the
[18]. In coarse-to-fine iteration method, the clustering boundary feature the cardiologist can identifying the intra-
windows are searched to get a fine position with half size of cardiac masses in an echocardiogram [27]. In poor image
sub windows as shown in fig 6(b). Whenever the sub quality the two intra-cardiac masses have different texture
windows are in same intensity then iterations are ends [19]. features. Here three clauses of features are the motion, the
boundary and the proposed texture features. In these clauses
C. Globally De-Speckling by NLM motion feature is an important factor in mass recognition to
The Image 6(c) de-noising algorithm is used to extract the get high degree of mobility.
noise from during acquisition and transmission to get a good The base length of thrombus is much longer than
image quality in medical image processing, A video de- that of a tumor. One of the methods for texture feature is
nosing frame that integrates optical flow into a non-local GLCM [28]. In this method to classifying homogeneous and
means framework with noise level estimation as shown in heterogeneous areas. Furthermore, we can introduce
fig 6(d) [20].The approximate K-nearest neighborhood can Multiclass SVM classifier for the better classification of the
significantly reduce the complexity of classical NLM intra-cardiac masses in an echocardiography.
methods. The main aim is a high degree of redundancy in an
image [21]. In NLM, similar patches are aggregated together VI. MULTICLASS SVM CLASSIFIER
with weights based on patch similarities. Initially, segment The main aim of Multiclass SVM classifier is to assign
depth is to create a binary image with any broken boundary labels with in time by using support vector machines. A
appendages are removed as shown in fig 6(e) but it is an finite set of elements can be drawn the labels. Here to
insufficient method to find similar patches in a whole image reduce the single multiclass problem into multiple binary
[22]. Another way is to compress the search windows in classification problems. Mainly, it approaches two methods:
larger size to get local optimization and filtering the One-Against-All and One-Against-One.
additional speckles to get clear border as shown in fig 6(f)
[23]. Mathematically, we can consider {a1, a2, ,al} is an A. 1A1Method
input noisy sequence with l frames and q = (u, v, t) To construct K (K- 1)/2 classifiers are applied on each pair
represents space-time volume, and P(q) denote a patch at of classes. Where each one is trained on data from two
location q. For each pixel q, the set of approximate K- classes. So it performance can be reduced due to imbalanced
nearest neighbors training data sets [29]. According to the maximal output
(AKNN) N (q) = {P (qi )} K i=1 before applying the sgn function. In this method training
To remove the opposite characteristics of two data from a and b classes can be given as
patches the same Y_ Sum, the average intensity Z_ mean R Min ( ab)' ab + =1 ab
z'= Z similar Sign ((ab)' ab)) denote x is in a class, is added
Where z' is reshaped from column vector. by one. Otherwise, b is increment by one.
Segmentation is based on low-level features for color and
texture with partitioning an image into disjoint and
homogeneous regions. It can be achieved by finding the
boundaries between the regions [24].In segmentation,
adjacent regions have different values with respect to the
uniform characteristics and these boundaries of each
segment should be simple and must be spatially accurate.
The two important low level parameters such as color and
texture are used for content based re-trivial of information in
images [25]. The segmentation of an images are composed
of more and less textures .These texture properties are not
well defined because of all textures have individual regions
bounded by the marked edges in the binary image to create a
segmented image as shown in fig 6(h).The binary image is
dilated and eroded with a structure of square or triangle. The Fig. 2: One-Against-One approach

All rights reserved by 388

Identification of Intra-Cardiac Masses based on Advanced ACM Segmentation using Multiclass SVM Classifier
(IJSRD/Vol. 4/Issue 05/2016/096)

B. 1AA Method
During the testing, data sets are classified by finding margin
from the linear separating hyper-plane. In this class, the final
output have largest margin that corresponds to the SVM. If
two or more classes outputs are very close to each other
than that points are labeled as unclassified [30]. For K class
data set is divided into K- two class data set. For n
training data can be given as
(p1; q1), , (pl ; ql),
Where p(i) Rn, i = 1,. , l and q(i) (1,.,k )
A new sample x represents the class that
corresponds to the farthest hyper-plane.
For the class of p(i) , the mth SVM is given by
Min (l)' l + C =1
Where C =1 represents to reduce the number
of training errors

Fig. 4(a): Classification results for proposed method


Fig. 3: One-Against-All approach

C. Advantages
1) It resistance can be merges the visual information in
the analyze image and greatly reduce the noise.
2) Fast evaluation of the learned data set function.
3) This classifier is an ability to detect the texture
boundaries. Fig. 4(b): Classification results for existing method
4) In the sense that the number of binary classifiers to (Thrombi)
construct the equal number of classes.
5) Multiclass SVMs have been demonstrated superior
classification accuracies to neural networks.
6) Sparseness of solution when dealing with large data
D. Applications
Multiclass SVM has been used successfully in many real-
world problems.
1) Multiclass SVMs are helpful in text and hypertext
categorization, handwritten digit and object
recognition, speaker identification, face detection in
2) It experimental result shows that to achieve
significantly higher accuracy than conventional
refinement schemes after an iterative is ends.
3) It also helpful to identify protein compounds in
medical science.
4) Hand-written characters can be recognized using
Multiclass SVM.
Fig. 4(c): Classification results for proposed method

All rights reserved by 389

Identification of Intra-Cardiac Masses based on Advanced ACM Segmentation using Multiclass SVM Classifier
(IJSRD/Vol. 4/Issue 05/2016/096)

Fig. 4(d): Classification results for existing method (Tumor)

Fig. 4(g): Simulation method for proposed method (Tumor):

Fig. 4(e): Simulation method for proposed method


Fig. 4(h): Simulation method for existing method (Tumor)

A proposed method (MSVM) is an identification of the
intra-cardiac masses in an echocardiogram. The ROI is
automatically defined by a coarse-to-fine iteration method.
After that to eliminate the noise with de-speckling by NLM.
It yields good noise attenuation and texture features without
disturbing an original image. Consequently, using an
adaptive color segmentation to get better results when
iterations are end. Finally, it is able to identify the intra-
cardiac masses with an accuracy of 99% and sensitivity of
100% and remaining features such as Specificity, NPV and
PPV values. For providing higher performance for further
research efforts.

[1] Columbus M: Libri 15 De re anatomica. Paris 1562.
Fig. 4(f): Simulation method for existing method (Thrombi)

All rights reserved by 390

Identification of Intra-Cardiac Masses based on Advanced ACM Segmentation using Multiclass SVM Classifier
(IJSRD/Vol. 4/Issue 05/2016/096)

[2] McAllister H, Fenoglio J: Tumors of the cardiovascular Proc. ACM Multimedia 95, San Fransisco, CA, pp.
system. In:Hartmannn W, Cowan W (eds.): Atlas of 189-200, November 1993.
tumor pathology. Washington, DC: Armed Forces [20] J. S. Boreczky and L. A. Rowe, Comparison of Video
Institute of Pathology 1978: 120. Shot Boundary Detection Techniques, Storage and
[3] S.Maraj, G.S.Pressman and V.M.Figueredo, Primary Retrieval for Image and Video Databases IV, Proc. Of
cardiactumors, Int.J.Cardiol., Vol.133, no.2.pp.152- IS&T/SPIE 1996 Intl Symp. on Elec. Imaging: Science
156, Apr 2009. and Technology, San Jose, CA., February 1996.
[4] M.ORLON, E. V. B.: Benign polypoid tumors of the [21] Olshausen, B.A., Field, D.J.: Emergence of simple-cell
left auricle. Edinburgh M. J. 41: 227, 1954. receptive field properties by learninga sparse code for
[5] DAVIS, T. W. AND ANDRUS, E. C.: Mitral stenosis natural images. Nature 381 (1996) 607609.
in facsimile. New England J. Med. 251: 297, 1954 [22] Yang, J., Wright, J., Huang, T., Ma, Y.: Image super-
[6] Obeid AI, al Mudamgha A, Smulyan H. Diagnosis of resolution as sparse representation of raw image
right atrial mass lesions by transesophageal and patches. In: IEEE Conference on Computer Vision and
transthoracic echocardiography .Chest1993; 103:1447 Pattern Recognition (CVPR). (2008).
51. [23] Buades, A., Coll, B., Morel, J.M.: A non-local
[7] M. G. Tsipouras, T. P. Exarchos, D. I. Fotiadis, A. P. algorithm for image denoising. In: IEEE Conference on
Kotsia, K. V. Vakalis, K. K. Naka, and L. K. Computer Vision and Pattern Recognition (CVPR).
Michalis,Automated diagnosis of coronary artery (2005).
disease based on data mining and fuzzy modeling, [24] Y.Guo, Y. Wang and T. Hou,Speckle filtering of
IEEE Trans. Inf. Technol. Biomed. vol.12, no.4, ultrasonic images using a modified non local-based
pp.447-458 Jul.2008. algorithm,Biomed. Signal Process., vol.6,no.2,pp.129-
[8] T. Huang, G. Yang, and G. Tang,A Fast Two- 138,Apr.2011.
Dimensional Median Filtering Algorithm , IEEE [25] R.M. Haralick, L.G. Shapiro, \Image Segmentation
Trans.Acoust.Speech signal process., vol.27, no.1, Techniques,"Computer Vision, Graphics, and Image
pp.13-18 Feb.1979. Process-ing, Vol. 29 , No.1, pp. 100-132, Jan. 1985.
[9] Y.Yongjian, Yu and Scott T. Acton,Speckle Reducing [26] M.J. Swain and D.H. Ballard, \Color Indexing," Int'l
Anisotropic Diffusion, IEEE Trans.Image process., Jour-nal of Computer Vision, Vol. 7, No. 1, Nov. 1991,
vol.11, no.11, pp.1260-1270,Nov.2002. pp. 11-32.
[10] R.Bellotti, F.Carlo A CAD system for nodule detection [27] M. Mirmehdi and M. Petrou, Segmentation of color
in low-dose lung CTs based on region growing and a textures, IEEE Trans. Pattern Anal. Machine Intell.,
new active contour model, Med.Phys., vol.34, no.12, vol. 22, pp.142159, Feb. 2000.
pp.4901-4910, Dec.2007. [28] S.Maraj, G.S.Pressman and V.M.Figueredo, Primary
[11] M.Aharon,M.Elad and A.Bruckstien,K-SVD: An cardiactumors, Int.J.Cardiol., Vol.133, no.2.pp.152-
algorithm for designing over complete dictionaries for 156, Apr 2009.
sparse representation, IEEE Trans. Signal process., [29] M.Strzelecki, A.Materka,J.Drozdz ,Classification and
vol.54, no.11,pp.4311-4322, Nov.2006. segmentation of intracardiac masses in cardiac tumor
[12] Cortes, C.; Vapnik, V. (1995). Support-vector echocardiograms, Comput.Med.Imaging Graph,vol.30,
networks. Machine Learning 20 (3): 273.doi:10.1007/ no.2,pp.95-107,Mar.2006.
BF00994018. [30] L.M.Po and W.CW.Ma.a novel for step such algorithm
[13] Press, William H.; Teukolsky, Saul A.; Vetterling, for fast block motion estimation,IEEE T trans.circuits
William T.; Flannery, B. P. (2007). Section techno., vol. 6, No.3,PP.313-317, jun.1996.
16.5.Support Vector Machines. Numerical Recipes: [31] Melgani,F., and Bruzzne, L .2004. Classification of
The Art of Scientific Computing (3rd ed.). NewYork: hyper spectral remote sensing images with SVM.IEEE
Cambridge University Press. ISBN 978-0-521-88068-8. Trans. on Geoscience and remote sensing, vol.42,
[14] K. R. Muller, S. Mika, G. Ratsch, K. Tsuda, and B. pp.1778-1790.
Scholkopf, An introduction to kernel-based learning [32] Platt, J., Cristianini, N., Shawe-Taylor, J.: Large margin
algorithms, IEEE Trans. Neural Networks, vol. 12, pp. dags for multiclass classification.
181201, Mar. 2001. [33] In: Advances in Neural Information Processing
[15] 1. H. Feigenbaum, Echocardiography, LEA & Systems, vol. 12, pp. 547553. MIT Press, 2000.
FEBIGER, 1993. [34] Vapnik, V.: Statistical Learning Theory. Wiley, New
[16] B. Shahraray, Scene Change Detection and Content- York (1998).
Based Sampling of Video Sequences, Proc. SPIE, vol.
2419,pp. 2-13, 1995.
[17] F. Arman, A. Hsu, and M. Y. Chiu, Image Processing
on Compressed Data for Large Video Databases,
Proc.ACM Multimedia 93, pp.267-272, August 1993.
[18] H. Zhang, A. Kankanhalli, and S. Smoliar, Automatic
Partitioning of Full-Motion Video, Multimedia
Systems, 1(1), pp. 10-28, 1993.
[19] R. Zabih, J. Miller, and K. Mai, A Feature-Based
Algorithm for Detecting and Classifying Scene Breaks,

All rights reserved by 391