LUNG NODULE CLASSIFICATION UTILIZING SUPPORT VECTOR MACHINES

Wail A.H Mousa and Mohammad A. U Khan *t

Department of Electrical Engineering King Fahd University of Petroleum and Minerals

Dhahran 31261, Saudi Arabia wail,maukhan@kfupm.edu.sa

ABSTRACT Lung cancer is one of the deadly and most common diseases in the world. Radiologists fail to diagnose small pulmonary nodules in as many as 30% of positive cases. Many methods have been proposed in the literature such as neural network algorithms. Recently, support vector machines (SVM)’s had received an increasing attention for pattem recognition. The advantage of SVM lies in better modeling the recognition process. The objective of this paper is to apply support vector machines (SVM)’s for classification of lung nodules. The SVM classifier is trained with features extracted from 30 nodule images and 20 non-nodule images, and is tested with features out of 16 noduleinon-nodule images.The sen- sitivity of SVM classifier is found to be 87.5%. We intend to automate the pre-processing detection process to further enhance the overall classification.

I. INTRODUCTION

The detection of lung nodules in early stages of growth can result in saving lives. Radiologists fail to diagnose small pulmonary nodules in as many as 30% of positive cases. To assist radiologists in their job, S . Lo and others reported in 1995 [I] a development of a double-matching method and an artificial visual neural network algorithm for lung nodule detection. The fundamental operation carried out in their neural network system is the local two-dimensional convo- lution. Out of 207 suspected nodules areas (SNA’s), 52 real nodules were found with 155 false positive (FP) regions. The success rate of their algorithm was in terms of sensitiv- ity of 86%.

Later on, in 1996, J. Lin et. al. [Z] developed a neural- digital computer-aided diagnosis system based on a param- eterized two-level convolution neural network. It performs an automatic suspect localization, feature extraction, and di- agnosis of a particular pattern-class aimed at a high degree

‘Authors would like to acknowledge the suppon o f King Fahd University ofPetroleum and Minerals, Dhahran 31261, Saudi Arabia.

tAuthors would like to acknowledge the Image, Speech and Intelligent Systrms Group, University OfSouthampton for using their SVM MATLAB toolbox.

’

of true positive (TP) fraction detection and a low FP frac- tion detection. The system was successful in detecting and classifying 66 nodules out of 110 SNA’s, while 46 were con- sidered to be FP’s. However, their sensitivity was 80% ob- tained with less computational complexitiy compared to the previous algorthim reported in [ I ] .

In 1998 [3], a computer-aided diagnosis system based on a two-level artificial neural network was established. The first neural network (NN) was dedicated for detecting SNA’s. Features from the SNA’s detected by the first NN were passed onto a second NN. Such features included curvature peaks computed for all pixels in each SNA. The output ofthis sys- tem is thresholded at a chosen level of significance to give a positive detection. They used 90 real nodules and 288 sim- ulated nodules for the training. Then the system was tested to give 89% - 96% sensitivity.

Support Vector Machine (SVM) is a technique for leam- ing from examples that is well-founded in statistical leam- ing theory. It was used for isolated handwritten digit recog- nition, object recognition, speaker identification, face de- tection in images, text recognition, etc [4]. It has received an increasing attention for pattem recognition. This use of SVM for classification problems has several advantages not found in methods such as those based on NN algorithms. For example, traditional NN approaches have faced difficul- ties with finding a general model to fit any data. This dif- ficulty is removed when the problem is solved using SVM

The objective of this paper is to utilize SVM classifier for classifying lung nodules. In section 11, materials and methods used are presented. Section Ill, provides system performance evaluation. Finally, conclusion and discussion is provided in section IV.

[SI.

11. MATERIALS AND METHODS

Our lung classification scheme consists of preprocessing step for signal enhancement and SVM-based classifier.

0-7803-7622-6/02/$17.00 02002 IEEE 111 - 153 IEEE ICIP 2002

I I -A. Data Base and Image Enhancement

The data base is composed of 38 nodule images and 28 non-nodule images. Several algorithms were applied on the images to enhance them. It happens that a combination of special domain high frequency emphasis filtering (SPD- HFEF), histogram equalization followed by homomorphic filtering was giving the best enhancement among other well known techniques. In SPD-HFEF, the image was first passed through a high pass filter to enhance the edges in the image. After that, some of the low pass image was added to the high passed image to restore some the energy back to the image. Based on the SPD-HFEF resultant image, global histogram equalization was required to improve the contrast. Conse- quently, noise was also emphasized. Therefore, homomor- phic filtering was applied and hence a very well enhanced image was ultimately obtained. The size of the nodulehon- nodule images is 100 x 100 pixels. Fig. 1 shows the 30 nodules used for training the classifier while Fig. 2 shows 20 non-nodules also used for the training process. Fig. 3(a) and Fig. 3(b) indicate the rest of the nodulehon-nodule im- ages, respectively. These 16 images were used for testing the classifier.

Fig. 1. The 30 nodulcs uscd for training the classifier. figure

Fig. 2. The 20 non-nodules used for training the classifier. figure

Fig. 3. The testing (a) 8 nodule images and (b) 8 non-nodule images, figure

":! .> b, ;. ;, d. .I d. ;, I-

Fig. 4. The feature space of the training data for both nod- ules and non-nodules. figure

11-B. Feature Extraction from Nodule/"-nodule Im- ages

Feature extraction can be done in several ways such as finding the circularity ofthe nodules themselves. Other fea- tures can he obtained by segmenting the nodules and per- forming edge detection then finding the mean and variances of the resultant edges related to the nodules. The means and variances of a thresholded nodulelnon-nodule areas were taken to be features used for classification after normalizing their values. This is because normalization of the features used to train SVM for classification is necessary for some kernels. Fig. 4 describes the feature space based on normal- ized mean and variance of the gray levels of the training set data.

11-C. Classification of Nodules from Non-Nodules based on SVM

SVM implements the idea of mapping input vectors X into a high dimensional feature space through non-linear mapping. The mapping is performed via a chosen a priori [6] to construct an optimal separating hyper-plane as Fig. 5 presents. According to Fig. 4, linear hyper-planes for separable and non-separable data cannot be implemented. However, using kernel mappings such as radial basis func- tion, a non-linear byper-plane will be constructed such that

Table I. polynomial results with d = 10 where Sen.: for Sensitivity, Acc.: stands for Accuracy, and Spe.: for Specificity. table

.' 0 e .

Input Space

0 . . . 0

* * . 0 . * Output Spnce

stands stands

* e: Fenma Spxe

Fig. 5. Mapping the input space into a high dimensional feature space. figure

it can easily distinguish between the class of nodules from the non-nodule class. Therefore, the optimization problem, in its dual form, will he presented as follows:

Given the trainingsamples{(Xi,y;)}~=,, findthe La- grange multipliers {X;}zl that maximizes:

subject to.

i= l

O ~ X i ~ C (3)

where X; is input pattern for the ith erample, y; is the cor- responding desired response, K(X i , X j) an inner product kernel that must satisfy Mercer k theorem 141, and C is a positive penalty component used to contml the amount of overlap that 1s allowed between classes. The idea of the kemel function is to enable operations to be performed in

the input space rather than the potentially high dimensional feature space. This will help us to reduce the dimensionality and, therefore, reduce the computational complexity. How- ever, the computation is still critically dependent upon the number of Xi. Ultimately, a good data distribution for high dimensional problem will generally require a large training set [SI. The common choice for K(Xi,Xj) could he as follows: 1. Gaussian Radial Basis Function (GRBF):

(4)

where U is defined to be the global basis function width. 2. Polynomial:

K(X,Y) = ((X'Y) + l ) d (5)

where d = 1,2, . . . is the degree of the polynomial

111. SYSTEM PERFORMANCE EVALUATION In order to select the hest classification performance, ex-

periments were made on the polynomial kernel function with different d's and different C's. The evaluation was based on the following criterion used in [3],[l],and [2]: 1. True Positive (TP): the percentage of nodules classified as nodules. 2. False Positive (FP): the percentage of non-nodules clas- sified as nodules. 3. False Negative (FN): the percentage ofnodules classified as non-nodules. 4. True Negative (Th'): the percentage of non-nodules clas- sified as non-nodules. In addition, some authors [3], evaluate their classifiers in terms of Sensitivjtj Accuracy and Specificity. These crite- rion are defined to he:

T P T P + F N

Sensitivity =

T P i T N T P + T N + F P + FN Accuracy =

(6) T N

T N i F P Specificity =

Also, a common method of characterizing classification per- formance called recejver operator characteristic curve or the ROC curve [7] will be presented. Basically, ROC curve demonstrates the classification rate (normalized'TP in our case) versus the false alarm rate (normalized FP in our case). The evaluation is done according to the tested data with dif- ferent trained classifiers as will he described in the coming subsections.

111-A. Polynomials The test using polynomials was used ford = 2, d = 3,

d = 4, and d = 10, respectively, for different C's. The best

I11 - 155

Fig. 6. Space ofthe SVM polynomial of degree IO with the training data. figure

/ /’

Fig. 7. Space of the SVM polynomial of degree I O with the training data and the tested data. figure

results were obtained by the polynomial of degree 10 and C = 0.483 and with Sensitivity= 87.5%, Accuracy= 75%. and Specificity= 62.5%. Tabel.1 demonstrates the exper- iment performed using polynomial kemel of degree IO. Fig.6 shows the space of the SVM polynomial of degree I O with the training data while Fig.7 shows the space of the SVM polynomial ofdegree IO with the training data in ad- dition to the tested data. Note that the pluses sign stands for nodules, the siars are used for non-nodules, triangles are the test nodules, squares for representing the test non- nodules, and circles indicates support vectors. Fig.8 is the ROC curve for the SVM polynomial of degree 10 for the tested data.

N. CONCLUSION AND DISCUSSION

In view of the results obtained by utilizing SVM for clas- sification of lung nodules from non-nodules, we can make the following evaluation. The sensitivity result of using the SVM polynomial of degree IO is 87.5%. This result is ac- ceptable based on the previous sensitivity values reported in [3],[l],and [Z]. However, comparison with previously re- ported techniques will require a standard image data base. To have a fair comparison, the SVM techniques and any

Fig. 8. ROC curve for the SVM polynomial of degree IO. figure

other technique should he applied to the same data and, con- sequently, the same features.

It might be a good idea to add a phase that can automate the process of detecting the suspected nodule areas (SN4’s) prior to the classification step by using well know tech- niques such as Template Marching and Correlaiion Meih- ods [7]. In addition, a better feature extraction can be made by computing the circulariiy of the nodules, using Fourier Descriptors, etc. The use of these features will probably increase the chance of obtianing a better performance.

Vr REFERENCES S. B Lo, S. A. Lou, J. Lin, M. Freedman, M. V. Chien, and S. K. Mun, “Artificial convolution neural network techniques and applications for lung nodule detection,” IEEE Tran. on MedicalImaging, vol. 14, no. 4, pp. 71 1- 718, December 1995. S. Lin, S. Bo, A. Hasegawa, M. Freedmans, and S. K. Mun, “Reduction of false positive in lung nodule de- tection using a two-level neural classifier,” IEEE Tran. on Medical Imaging, vol. 15, no. 2 , pp. 20&2217, April 1996. M. G. Penedo, M. J. Carreira, A. Mosquera, and D. Ca- hello, “Computer-adided diagnosis: A neura-network- based approach to lung nodule detection,” IEEE Tran. on Medical Imaging, vol. 17, no. 6, pp. 872-880, De- cember 1998. V. Vapnik, The Nature of Siaiisiical Learning Theory, Springer Verlag, 1995. C. J. C. Burges, “A tutorial on support vector machines for pattern recognition,” Tech. Rep., Bell Laboratories, Lucent Technologies, 2000. S. Raudys, Siaiisiical and Neural Network Classifiers, An Iniegraied Appmach io Design, Springer, 200 I . M. J. T. Smith and A. Docef, A Sfudy Guide for Digiial Image Pmcessing, Scientific Publishers, Inc, 1999.

111 - 156