Development of an Automated Medical Diagnosis System for Classifying Thyroid Tumor Cells using Multiple Classifier Fusion

Transcription

1 Technology in Cancer Research and Treatment ISSN Volume 4 Number 5 October April 6. Epub ahead of print. Development of an Automated Medical Diagnosis System for Classifying Thyroid Tumor Cells using Multiple Classifier Fusion DOI: /tcrt An automated medical diagnosis system has been developed to discriminate benign and malignant thyroid nodules in multi-stained fine needle aspiration biopsy (FNAB) images using multiple classifier fusion and presented in this paper. First, thyroid cell regions are extracted from the auto-cropped sub-image by implementing mathematical morphology segmentation method. Subsequently, statistical features are extracted by two-level wavelet decomposition based on texture characteristics of the thyroid cells. After that, decision tree (DT), k-nearest neighbor (k-nn), Elman neural network (ENN) and support vector machine (SVM) classifiers are used separately to classify thyroid nodules into benign and malignant. The four individual classifier outputs are then fused together using maority voting rule and linear combination rules to improve the performance of the diagnostic system. The classification results of ENN and SVM classifiers show an overall diagnostic accuracy (DA) of 90%, sensitivity (Se) of 85% and 00%, specificity (Sp) of 90% and 90% respectively. However, the best diagnostic accuracy of 96.66% is obtained by multiple classifier fusion with maority voting rule and linear combination rules. The experimental results show that the proposed method is a useful tool for the diagnosis of thyroid cancer and can provide a second opinion for a physician s decision. B. Gopinath, Ph.D. * N. Shanth Ph.D. Department of Electronics and Communication Engineering, Info Institute of Engineering, Coimbatore 6407, India Department of Computer Science and Engineering, Nandha Engineering College, Erode 63805, India This co-author contributed equally to this work. Key words: Benign; Classifier fusion; Malignant; Medical diagnosis system; Thyroid cancer. Introduction The thyroid is a small gland that present in the lower part of the human neck. The function of the thyroid gland is to produce hormones. These thyroid hormones deliver energy to cells of the body. Many types of tumors can develop in the thyroid gland. Most of these tumors are benign and not cancerous. However, 5-0% of thyroid nodules are discovered to be thyroid malignant (). Fine needle aspiration biopsy examination in the investigation of thyroid nodules is relatively simple, minimally invasive, painless and inexpensive with high accuracy. It can provide a clear diagnosis in most of the situations when performed by well-trained clinicians. Traditionally, clinicians study the features of thyroid biopsy samples obtained from patients using FNAB technique, examine them under a microscope Abbreviations: ASM: Angular Second Moment; DA: Diagnostic Accuracy; DT: Decision Tree; DWT: Discrete Wavelet Transform; ENN: Elman Neural Network; FNAB: Fine Needle Aspiration Biopsy; GLCM: Gray-level Co-occurrence Matrix; H&E: Hematoxylin-Eosin; k-nn: k-nearest Neighbor; LC-: Linear Combination Type-; LC-: Linear Combination Type-; LC-3: Linear Combination Type-3; MGG: May-Grunwald-Giemsa; MV: Maority Voting; p-value: Probability Value; RGB: Red Green Blue; SD: Standard Deviation; Se: Sensitivity; Sp: Specificity; SVM: Support Vector Machines; W CR : Weight as a Function of Classification Rates of the Individual Classifiers. *Corresponding author: B. Gopinath, Ph.D. Phone: gopiphd@yahoo.com 653

2 654 Gopinath and Shanthi and make udgments based on their experience. However, this udgment is subective and often leads to considerable variation. To overcome these problems and improve the reliability of thyroid malignant diagnosis, it is important to develop computational tools for automated cancer diagnosis that operate on a set of FNAB thyroid cytological images taken from glass slides. Such automated cancer diagnosis facilitates a mathematical conclusion which can provide a second opinion for physician (3). In recent years, considerable efforts have been made in the development of the automated medical diagnosis system to improve the clinicians confidence in the analysis of medical images (3-7). Karakitsos et al. investigated the use of neural networks to classify benign and malignant thyroid lesions and achieved 97.8% of diagnosis accuracy using May-Grunwald-Giemsa (MGG) stained thyroid FNAB images (). Daskalakis et al. designed a multiclassifier system for discriminating benign from malignant thyroid nodules using Hematoxylin-Eosin (H&E) stained cytological images (3). Alexander et al. conducted an interesting analysis on testing 65 indeterminate thyroid nodules among which 85 were malignant using a gene-expression classifier. The gene-expression classifier correctly identified 78 of the 85 nodules as suspicious with 9% sensitivity (4). Zahir et al. investigated the utility of the combined classification and regression trees (CART) classifier to differentiate benign from malignant thyroid nodules in patients referred for thyroid surgery. The CART classifier produced a sensitivity and specificity of 80.0% and 94.% respectively (5). Acharya et al. conducted a research work on the automated classification of benign and malignant thyroid lesion in 3D contrastenhanced ultrasound using the combination of wavelets and textures. They achieved a classification accuracy of 98.9%, a sensitivity of 98% and a specificity of 99.8% (6). Gopinath et al. presented an automated diagnosis system for classifying multi-stained thyroid FNAB cytological images using single classifiers. In this work, classification was performed based on the statistical feature vectors using Elman neural network (ENN), support vector machine (SVM) and k-nearest neighbors (k-nn) classifiers (7). Gopinath et al. tested maority voting based classification of papillary and medullary thyroid carcinomas using FNAB cytological images and SVM based approach in the diagnosis of thyroid malignancy using statistical texture features (8, 9, 3). From the literature survey, it is observed that the development of a generalized automated diagnosis system for thyroid cancer using multi-stained thyroid FNAB images is not addressed in the literature (, 3). Furthermore, research on thyroid lesion classification using statistical texture features derived from discrete wavelet transform (DWT) and fusion of classifiers is scanty in the literature. Statistical texture methods are motivated from Julesz s findings that human visual systems usually recognise textured obects based on the statistical distribution of their image gray-levels. Julesz et al. analysed the spatial distribution of gray-values by computing texture features at each point in the image and deriving a set of statistics from the distributions of the texture features (0). Moreover, the values of texture features are highly scale-dependent. Multi-scale methods use the concept of decompositions that result different values of texture features for different scales. Discrete wavelet transform can act as basic function for this multi-scale approach (). Thus, the novelty of the proposed computer-aided diagnosis system can be evaluated by incorporating statistical texture features derived from DWT and fusion of classifiers for classifying multi-stained thyroid FNAB images irrespective of staining protocol used. The study starts with segmentation of thyroid cells by the morphology segmentation method. Statistical textural features are subsequently extracted using two-level DWT decomposition and discrimination of benign and malignant thyroid nodules is performed by DT, k-nn, ENN and SVM classifiers. Finally, fusion of classifiers using maority voting rule and linear combination rules are used to improve the performance of the basic learning methods. Materials and Methods Fine Needle Aspiration Biopsy Fine needle aspiration biopsy is the study of cellular samples obtained through a fine needle under negative pressure. FNAB is recommended as the first and most decisive diagnostic step in the work-up of patients with nodular thyroid disease. Thyroid FNAB can be performed effectively by either a surgical registrar or consultant surgeon in the outpatient clinic. The sample material was collected from the patient using a 3-gauge needle connected to a 0 ml syringe mounted on a syringe holder. The sample was transferred onto glass slides. FNAB specimens on glass slides were analyzed by a clinician based on the cytological features. After examining all the slides, the clinician will make a cytological diagnosis report. The classifications are Benign, Malignant, Suspicious and Inadequate. Normally the accuracy of FNAB technique is around 95%, false negative and false positive results vary from 0 to 5% (). In the development of automated medical diagnosis systems, the cytological images are captured from the glass slides of the sample material. The classification process is then automated using various image processing techniques. The developed diagnosis system uses the thyroid FNAB images obtained from an on-line image atlas of the Papanicolaou Society of Cytopathology, which are reviewed and approved by the atlas committee (). The training image set consists of 80 images (40 benign and 40 malignant images) whereas the testing image set has 30 images (0 benign and 0 malignant images). Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

3 Computer-aided Diagnosis of Thyroid Tumors 655 Image Preprocessing The thyroid FNAB images are stained by various types of stains and they are called as multi-stained cytological FNAB images. In automated diagnosis system, the diagnostic accuracy is affected by the unwanted background staining information and hence it is very important to carefully remove the unwanted background information in multi-stained FNAB images. In this study, the FNAB images are initially auto-cropped to obtain high density cell portion followed by thresholding operation. Finally, mathematical morphology based image segmentation method is applied to segment the required benign and malignant cell regions of thyroid nodules in multi-stained FNAB images. The preprocessing stage consists of converting RGB image into gray-scale image, cropping pixel sub-image from the input slide image by auto-cropping and thresholding. In automated segmentation and classification methods, high density cell regions in slide images are normally identified before segmentation for simple and fast computing (3). In auto-cropping technique, a cropping window is moved over the entire image and mean value is calculated for each window. Statistically, a window with high density cell region in the slide image will have least mean value. Thus, a window having least mean value is cropped and used for further processing which will improve the diagnostic accuracy. The original image with pixels and cropped image with pixels are shown in Figure A and B. Thresholding is the most common method of separating obects from its background in image that assigns a pixel to one class if its gray-level is greater than a specified threshold value and otherwise assigns it to the other class. using the histogram-based Otsu s method (4). The histogram h(r k ) of a digital image f(x,y) with L intensity level is defined as, h(r k ) 5 n k, where, k 5 0,,,..., L [] where, r k is the k th intensity level and n k is the number of pixels in the image. From probability theory, p(r k ) is an estimation of the probability of occurrence of intensity level r k. p(r k ) can be obtained as a normalized histogram by dividing all elements of h(r k ) by the total number of pixels n in the image and treated as a discrete probability density function p r (r k ), hr ( k ) pr( rk) = n The threshold value T is calculated so as to maximize the between-class variance, σ B, defined as, where, ω T σ = ω ( µ µ ) + ω ( µ µ ) B 0 0 t t L = p ( r ) and ω = p ( r ) 0 k k k k k = 0 k= T T µ = qp ( r )/ ω, µ = qp ( r )/ ω and µ = qp ( r ) 0 k k 0 k = 0 L L [] [3] k k t k k k= T k = 0 Each pixel in the image is compared to the threshold value. When a pixel s intensity is higher than the threshold value, the pixel is set to white; otherwise, it is set to black. Mathematically, the output image g(x,y) for the input image f(x,y), at global threshold value T, is given in Equation [4]. A global thresholding approach with a constant thresholding value is used. An appropriate threshold value is calculated g(x,y) { 0 if f(x,y) T otherwise [4] Figure : (A) FNAB cytological image of thyroid nodules. (B) Auto-cropped image. (C) Result of thresholding operation. Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

4 656 Gopinath and Shanthi The result of thresholding operation of an auto-cropped sample thyroid FNAB image is shown in Figure C. Cell Segmentation Segmentation of cells in microscopic images is a fundamental subect in quantitative image analysis. Because of the malignant characteristics of cancer cells are contained in the cell nucleus, the isolation of the cell nucleus is an important task of segmentation. In the past years, many methods for the segmentation of cell regions in microscopic biomedical images have been presented (5-7). The methods most commonly used for image segmentation can be categorized into two classes, namely region-based and boundary-based approaches (8). Mathematical Morphology Mathematical morphology, which is based on regionbased approach, provides a quantitative description of geometrical structures. It is used for extracting shape and size information from an image. In general, morphological operators transform the original image into another image through the interaction with the other image of a certain shape and size, which is known as the structuring element. Geometric features of the images that are similar in shape and size to the structuring element are preserved, while other features are suppressed (6, 8). The two basic morphological operators are dilation and erosion, from which many operations can be derived. The auto-cropped and thresholded FNAB images are treated with the morphological operators opening and closing along with structuring elements. If sets I and Se are referred to as the input image and structuring element, respectively, and s is an element of Se, then opening and closing are defined as, Where, Opening: (I Se) 5 (I Se) Se [5] Closing: (I Se) 5 (I Se) Se [6] Dilation :I Se = I [7] s s Se Erosion:I Se I [8] = S s Se The size parameter of a structuring element must be selected in accordance with the size of the structure to be extracted from the image. For the opening operation, all the elements smaller than the structuring element are removed. For the Figure : (A) Auto-cropped FNAB cytological image of thyroid nodules. (B) Result of morphological opening operator. (C) Result of morphological closing operator. (D) Original image with contour superimposed. closing operation, all the elements, present as holes, smaller than the structuring element are filled (9). Hence, the size of the structuring element must be larger than the size of the target thyroid cell obects. Thus, disk-shaped structuring elements with radii of three and eight pixels are selected for opening and closing respectively. Thus, small obects around the cell obects in the image are removed by performing the opening operation and holes inside the cells are filled by performing the closing operation. An image complement is performed to clearly distinguish foreground thyroid cell obects from the background image. Figure A shows automatically cropped FNAB cytological image of thyroid nodules. The results of morphological opening and closing operators are given in Figure B and C respectively. The contour superimposed image on the original image are presented in Figure D. Feature Extraction Feature extraction is a method of generating a description of an obect in an image in terms of measurable values. The extracted features represent the properties of the obect. These features can be used with a classifier to assign the class for the obect. Since most of the microscopic images exhibit textures and the cells in the microscopic images represent repeated irregular texture pattern, many microscopic image analysis systems deal with texture features and texture analysis methods for characterizing microscopic images. Two-level Wavelet Decomposition In computer vision, it is difficult to analyze the information content of an image directly from the gray-level intensity of the image pixels. For this purpose, the image information can be transformed into a set of details appearing at different resolutions. At different resolutions, the details of an image generally characterize different physical structures of the image. In this study, the wavelet transform and its associated scaling functions are used to decompose the thyroid FNAB images into different resolutions. In two-level wavelet decomposition, the low pass and high pass decomposition filters are applied in both horizontal and vertical directions on the thyroid FNAB images, followed by a sub-sampling of each output image. Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

5 Computer-aided Diagnosis of Thyroid Tumors 657 This will generate four wavelet coefficient images, i.e., LL, HL, LH and HH channels. The sub-bands labeled LH, HL and HH represent the finest scale wavelet coefficients, while the sub-band LL represents the coarse scale (i.e.) approximation. The process is repeated on the LL channel and the two-level wavelet decomposition is constructed (8, 9). The statistical features mean, standard deviation, entropy, variance, energy, homogeneity, contrast and correlation of the sub-bands of two-level decomposed images are calculated from thyroid FNAB images and stored in feature library as, Mean ( µ ) = f(, i )/ N [9] (( f i ) N) SD ( σ ) = (, ) µ / [0] Variance ( V) = σ [] Contrast = f(, i )( i ) [] Correlation = f( )( i µ )( µ ) / σσ i [3] f(, i ) Homogeneity = + i Energy or ASM = f( ) Entropy = f(, i )log f(, i ) [4] [5] [6] where, f() is the gray-level value for each pixel in the region of interest, N is the total number of pixels in the region of interest and µ i, µ, σ i, σ, σ are means and standard deviations of f (). Classification of Thyroid Nodules Image classification refers to the process of grouping unknown test samples into classes, where each resulting class contains similar samples according to some similarity criterion. In this study, a simple decision tree (, ), k-nearest neighbor (k-nn) algorithm (3), ENN (4, 5) and linear kernel-based SVM (3, ) are tested as the classification models. The tuning parameters of these four classifiers have been selected to obtain the optimized results as given in Table I. The developed medical diagnosis system consists of training and testing phases as explained in Figure 3. The FNAB training set consists of 80 images among which 40 images are benign and the remaining 40 images are malignant. These 80 images are used for training. In the training phase, 640 statistical texture features (80 8) are extracted from LL region of automatically cropped and segmented training set images and stored in the feature library. The testing set images are 30 images (0 benign and 0 malignant). In the testing phase, 40 statistical texture features (30 8) are extracted from LL region of automatically cropped and segmented testing set images and these features are compared with the features available in the feature library for the classification of benign and malignant thyroid nodules with a help of classifiers. Multiple Classifier Fusion Combining various classifiers in the prediction of normal and abnormal thyroid cells in the FNAB images can be used for combining different opinions in decision making. The advantage is that a multiple classifier is better than a single Table I Tuning parameters for the four classifiers. Classifier Possible tuning parameters Tuning parameters used in the current study Decision tree Possible criterions for choosing a split are Gini s diversity index, twoing rule and maximum deviance reduction. Gini s diversity index k-nn k The number of nearest neighbors used in the classification. k = Possible distance metrics are Euclidean, Cityblock, Cosine, Correlation, Euclidean distance metric and Hamming. ENN Number of neurons in the hidden layer. Number of neurons in the hidden layer 0 Possible transfer functions are Log-sigmoid, Tan-sigmoid and Linear Transfer functions. The transfer function for hidden layers is Tan-sigmoid and for the output layer is Linear transfer function. SVM Possible kernels are Linear kernel, Quadratic kernel, Polynomial kernel, Gaussian Radial Basis Function kernel, and Multilayer Perceptron kernel. Linear kernel Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

6 658 Gopinath and Shanthi Figure 3: Training and testing steps in medical diagnosis system. all the others. This searching rule has been used in association with the outputs of odd number of classifiers, i.e., k-nn, ENN and SVM classifiers to avoid a tie situation as in the case of even number of classifiers. The final classification decision C out is given by the maority of benign results or malignant results of individual classifiers C i as given in Equation [7]. C out = maority (C i ) where, i 5,, 3 [7] where, C decision of k-nn classifier C decision of ENN classifier C 3 decision of SVM classifier In linear combination multiple classifier fusion method, an even number of decisions can be fused to obtain a final decision with a concept of weighted voting. The final decision is the linear combination of set of weights and the classification results. Choosing the weights to improve the performance of the combined classifier is a challenging task. In this study, the classification rates of the individual classifiers are used as their weights (W CR ). Mathematically, given n set of basic classifiers with their individual classification decisions C i and a set of weights W CRi, the final classification decision C out of multiple classifier fusion is defined by the Equation [8]. Figure 4: Multiple classifier fusion. classifier if the single classifiers are correctly weighted and combined (7, 8). In this study, the output opinions of individual classifiers are fused using maority voting rule and linear combination as shown in Figure 4. The maority voting rule is normally used to find a particular member in any given sequence which has more votes than n Cout = WCRiCi where, i =,, 3, [8] i= To achieve optimized performance of multiple classifier fusion in terms of diagnostic accuracy, sensitivity and specificity, three types of linear combination methods have been tested for combining different classifiers as, (a) Linear combination type- (LC-) (b) Linear combination type- (LC-) and (c) Linear combination type-3 (LC-3) Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

7 Computer-aided Diagnosis of Thyroid Tumors 659 Linear combination type- is arrived by the combination of DT, k-nn, SVM, ENN classifiers and a set of weights W CRi as, C = W C where, i =,, 3, 4 out 4 i= CRi i [9] Linear combination type- is the combination of k-nn, SVM, ENN classifiers and a set of weights W CRi as, Figure 5: (A) Uncropped thyroid FNAB image. (B) Foreground thyroid cells. (C) Background stain information. C = W C where, i =,, 3 out 3 i= CRi i [0] Linear combination type-3 is the combination of SVM, ENN classifiers and a set of weights W CRi as, Results C = W C where, i =, out i= CRi i [] The useful feature vectors can be extracted for performing the classification procedure, if the segmentation of thyroid cells is effectively done. To enhance the performance of the segmentation algorithm, the concept of auto-cropping was implemented so that the amount of useful foreground thyroid cell information becomes more than the unwanted background stain information. The result of morphology segmentation algorithm for uncropped FNAB slide image and auto-cropped FNAB sub-image are shown in Figures 5 and 6 respectively. Figure 5B and 5C shows foreground thyroid cells and background stain images for uncropped image of pixels. Figure 6B and 6C shows foreground thyroid cells and background stain images for auto-cropped image of pixels. The percentage of useful foreground thyroid cell information and unwanted stain information for both uncropped and auto-cropped images are computed and tabulated in Table II. From Table II, it is naturally concluded that the auto-cropped FNAB sub-image has 50.35% of useful thyroid cell information compared with a 0.4% of thyroid cell information in uncropped image. Thus, the auto-cropped sub-images with Figure 6: (A) Auto-cropped thyroid FNAB image. (B) Foreground thyroid cells. (C) Background stain information. high density thyroid cell regions have greater useful information than the uncropped images. If the feature extraction methods are operated on these sub-images, effective feature vectors can be derived for solving the classification problems. The FNAB training set images are divided into two groups as benign set and malignant set with 40 images in each group. These 80 images are trained; eight statistical texture features are extracted and stored in feature library. Thus, the total number of features extracted from the training set images becomes high. Hence, best feature vectors with high significant levels are selected by calculating percentage variability in median values between benign and malignant image sets and conducting statistical t-test. The probability values (p-values) obtained by t-test and percentage variability in median values between benign and malignant image sets for the eight statistical features mean, standard deviation, variance, contrast, homogeneity, energy, entropy and correlation are tabulated in Table III. As given in Tables III, the p-values of the features mean, standard deviation, variance and energy are lesser than 0.05 indicating that there is a significant difference between benign and malignant feature sets of auto-cropped thyroid FNAB images segmented by the morphology segmentation Table II Comparison of foreground and background information. Sample image Image size Total number of pixels Total number of foreground thyroid cell pixels Total number of background stain pixels % of useful information Uncropped % Cropped % Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

8 660 Gopinath and Shanthi Table III Median and p-values of statistical features derived from auto-cropped FNAB images. Features Benign group Median value Malignant group % Variability in median value p-value Rank of features based on p-value Standard deviation Mean Energy Variance Homogeneity Entropy Correlation Contrast method. For the remaining features contrast, homogeneity, entropy and correlation, the p-values are greater than 0.05 meaning that these features exhibit no significant difference between benign and malignant feature sets. It is also observed that the percentage of variability in median values of the features mean, standard deviation, variance and energy are significantly high compared with the remaining features contrast, homogeneity, entropy and correlation. Furthermore, the features are ranked based on p-values and classification is performed by adding the features as given in Tables III and IV. The highly significant feature standard deviation gets first rank while the least significant feature contrast gets last rank. Discussion Table IV presents the performance of decision tree, k-nn, ENN, SVM single classifiers and multi-classifiers with maority voting rule and linear combination rules (LC-, LC- and LC-3) using the feature combinations SD, (SD MEAN), (SD MEAN ENERGY), (SD MEAN ENERGY VARIANCE) and 30 testing set images. It is clearly observed that the ENN and SVM classifiers performed better than the decision tree and k-nn classifiers. The SVM and ENN classifiers achieved 90% of diagnosis accuracy with different values of sensitivity and specificity. However, the highest diagnostic accuracy of 96.66% was reported by the multiple classifier fusion using maority voting rule and linear combination rules. Though the texture features standard deviation, mean, energy and variance have occupied top four ranks based on p-values obtained from t-test; the highest diagnostic accuracy of 96.66% has been achieved for maority voting and linear combination rules with single feature (standard deviation) along with other feature combinations. In the present study, the proposed computer aided diagnosis system has been designed using an optimal combination of Table IV Performance analysis of single and multi-classifiers. Performance of the classifiers (%) SD features (SD MEAN) features (SD MEAN ENERGY) features (SD MEAN ENERGY VARIANCE) features Classifiers DA Se Sp DA Se Sp DA Se Sp DA Se Sp DT k-nn ENN SVM Multi-classifier with maority voting rule Multi-classifier with LC- rule Multi-classifier with LC- rule Multi-classifier with LC-3 rule Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

9 Computer-aided Diagnosis of Thyroid Tumors 66 Table V Performance comparison of the proposed multi-stain diagnosis system and other existing studies. Studies Nature of study Staining protocol Classifiers Diagnostic accuracy (%) Karakitsos et al. () Daskalakis et al. (3) Gopinath et al. (8, 9) Gopinath et al. (7) Proposed medical diagnosis system Discrimination of benign from malignant thyroid lesions Discrimination of benign from malignant thyroid nodules Classification of Papillary and Medullary carcinoma in thyroid nodules Discrimination of benign from malignant thyroid nodules Discrimination of benign from malignant thyroid nodules May-Grunwald-Giemsa stained FNAB images LVQ-NN 97.8 H&E stained images PNN 89.6 k-nn/pnn/bayesian 95.7 Manually cropped multi-stained k-nn 95.8 FNAB images k-nn with maority voting rule 97.5 Auto-cropped multi-stained FNAB images Auto-cropped multi-stained FNAB images k-nn 70 SVM 90 ENN 90 Multi-classifier with maority voting rule Multi-classifier with LC- rule Multi-classifier with LC- rule Multi-classifier with LC-3 rule segmentation, feature extraction and classifier stages which are optimized in terms of parameter selection, feature size and testing sample size. Also, an optimized set of feature vectors is selected based on statistical t-test and size of testing images used. A direct comparison with previous studies is not feasible due to the differences in the image sets used for training and testing, feature size and the different staining protocols (3). However, a suitable comparison is made by keeping these limitations as given in Table V. Karakitsos et al. investigated the use of neural networks to discriminate benign and malignant thyroid lesions and achieved 97.8% of diagnostic accuracy using thyroid FNAB images stained by MGG staining protocol (). Daskalakis et al. designed a multi-classifier system for discriminating benign from malignant thyroid nodules using H&E stained FNAB cytological images and diagnostic accuracies of 89.6% and 95.7% were reported for PNN single classifier and combined classifiers respectively (3). These studies were not dealing with multi-stained FNAB images. In our previous studies (8, 9), we conducted experiments for discriminating thyroid malignancies with manually cropped thyroid FNAB images whereas the present study deals with automatically cropped thyroid FNAB images. Similarly, our previous study (7) with multi-stained and auto-cropped thyroid FNAB images demonstrated a reasonable diagnostic accuracy of 90% using ENN and SVM classification models. However, the present study achieved a promising diagnostic accuracy of 96.66% with multi-stained and auto-cropped thyroid FNAB images using maority voting and linear combination rules based multiple classifier fusion. Although the diagnostic accuracy of the proposed multi-stain diagnosis system is slightly lesser than the existing systems (, 8), it reaches 96.66% which may be considered as the most encouraging and working well irrespective of staining protocol used. Conclusion An automated medical diagnosis system to segment thyroid cell regions and classify them as benign or malignant was presented. Initially, thyroid cell regions were segmented from auto-cropped thyroid FNAB images by applying mathematical morphology segmentation method. In order to classify the segmented images, decision tree, k-nn, ENN and SVM classifiers were used among which the ENN and SVM classifiers exhibited superior diagnostic accuracy of 90% to recognize benign and malignant thyroid nodules using the statistical textural features derived by two-level wavelet decomposition. The improved diagnostic accuracy of 96.66% was achieved if multiple classifiers were fused using maority voting of k-nn, ENN and SVM classifiers and linear combination of single classifiers with LC-, LC- and LC-3 rules. Thus, the proposed medical diagnosis system for identifying thyroid cancer can be used as a second opinion tool to support physician s decision in daily clinical practice, when a definite diagnosis is difficult to be obtained irrespective of staining protocol used. Conflict of Interest All authors certify that this manuscript has not been published or considered for publication elsewhere. The authors have no conflicts of interest to declare. Acknowledgments The authors gratefully acknowledge Dr. V. Sindhu, M.D. (Pathology), for reviewing, verifying and validating the proposed automated diagnosis system for thyroid cancer. Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print

10 66 Gopinath and Shanthi References. Mahar, S. A., Husain, A., Islam, N. Fine needle aspiration cytology of thyroid nodule: diagnostic accuracy and pitfalls. Journal of Ayub Medical College 8, 6-9 (006).. Karakitsos, P., Cochand-Priollet, B., Guillausseau, P. J., Pouliakis, A. Potential of the back propagation neural network in the morphologic examination of thyroid lesions. Anal Quant Cytol Histol 8, (996). 3. Daskalakis, A., Kostopoulos, S., Spyridonos, P., Glotsos, D., Ravazoula, P., Kardar M., Kalatzis, I., Cavouras, D., Nikiforidis, G. Design of a multi-classifier system for discriminating benign from malignant thyroid nodules using routinely H&E-stained cytological images. Comput Biol Med 38, (008). DOI: 0.06/.compbiomed Alexander, E. K., Kennedy, G. C., Baloch, Z. W., Cibas, E. S., Chudova, D., Diggans, J., Friedman, L., Kloos, R. T., LiVols V. A., Mandel, S. J., Raab, S. S., Rosa J., Steward, D. L., Walsh, P. S., Wilde, J. I., Zeiger, M. A., Lanman, R. B., Haugen, B. R. Preoperative diagnosis of benign thyroid nodules with indeterminate cytology. N Engl J Med 367, (0). DOI: 0.056/NEJMoa Zahir, S. T., Binesh, F., Mirouliae M., Khaeh, E., Noshad, S. Malignancy risk assessment in patients with thyroid nodules using classification and regression trees. Journal of Thyroid Research, -9 (03). DOI: 0.55/03/ Acharya, U. R., Faust, O., Sree, S. V., Molinar F., Garberoglio, R., Sur J. S. Cost-effective and non-invasive automated benign and malignant thyroid lesion classification in 3D contrast-enhanced Ultrasound using combination of wavelets and textures: A class of ThyroScan Algorithms. Technol Cancer Res Treat 0, (0). 7. Gopinath, B., Shanth N. Computer-aided diagnosis system for classifying benign and malignant thyroid nodules in multi-stained FNAB cytological images. Australas Phys Eng Sci Med 36, 9-30 (03). DOI: 0.007/s Gopinath, B., Gupta, B. R. Maority voting based classification of thyroid carcinoma. Procedia Computer Science, 65-7 (00). DOI: 0.06/.procs Gopinath, B., Gupta, B. R. Classification of thyroid carcinoma in FNAB cytological microscopic images. Int J Healthc Inf Syst Inform 5, 60-7 (00). DOI: 0.408/hisi Julesz, B., Gilbert, E. N., Shepp, L. A., Frisch, H. L. Inability of humans to discriminate between visual textures that agree in second-order statistics revisited. Perception, (973). DOI: 0.068/p0039. Daubechies, I. Orthonormal bases of compactly supported wavelets. Comm Pure Appl Math 4, (988). DOI: 0.00/ cpa date accessed: 07/05/ Gopinath, B., Shanth N. Support vector machine based diagnostic system for thyroid cancer using statistical texture features. Asian Pac J Cancer Prev 4, 97-0 (03). DOI: 0.734/APJCP Otsu, N. A threshold selection method from gray-level histograms. IEEE Trans Syst Man Cybern 9, 6-66 (979). 5. Angulo, J., Flandrin, G. Automated detection of working area of peripheral blood smears by using mathematical morphology. Anal Cell Pathol 5, (003). 6. Hrebien, M., Stec, P., Nieczkowsk T., Obuchowicz, A. Segmentation of breast cancer fine needle biopsy cytological images. Int J Appl Math Comput Sci 8, (008). DOI: 0.478/ v x 7. Lee, J., Chung, Y. Integrating edge detection and thresholding approaches to segmenting femora and patellae from magnetic resonance images. Biomed Eng Appl Basis Comm 7, - (005). DOI: 0.405/S Gopinath, B., Shanth N. Automated segmentation of ELA cancer cells in microscopic images for evaluating the cytotoxic effect of selected medicinal plants. J Med Biol Eng 3, (0). DOI: /mbe Kimor Y. Mathematical morphology-based approach to the enhancement of morphological features in medical images. J Clin Bioinformatics, 33 (0). DOI: 0.86/ Haralick, R. M., Shanmugam, K., Dinstein, I. Textural Features for image classification. IEEE Trans Syst Man Cybern 3, 60-6 (973).. Perner, P., Perner, H., Müller, B. Mining knowledge for Hep- Cell image classification. Artif Intell Med 6, 6-73 (00). DOI: 0.06/S (0)00057-X. Chang, C., Huang, H., Chen, S. Automatic thyroid nodule segmentation and component analysis in ultrasound images. Biomed Eng Appl Basis Commun, 8-89 (00). DOI: 0.405/S Arif, M., Malagore, I. A., Afsar, F. A. Detection and localization of myocardial infarction using k-nearest neighbor classifier. J Med Syst 36, (0). DOI: 0.007/s Elman, J. L. Finding structure in time. Cogn Sci 4, 79- (990). 5. Song, Q. On the weight convergence of Elman networks. IEEE Trans Neural Network, (00). DOI: 0.09/ TNN Pham, D. T., Liu, X. Training of Elman networks and dynamic system modelling. Int J Syst Sci 7, -6 (996). DOI: 0.080/ Mahmoud Z., Rahat S., Ghasem M. M., Asadpour, V., Tayaran H., Raat M. Classification of voice disorder in children with cochlear implantation and hearing aid using multiple classifier fusion. BioMed Eng Online 0, 3 (0). DOI: 0.86/475-95X Lan, H., Carson, R., Provart, N. J., Bonner, A. J. Combining classifiers to predict gene function in Arabidopsis thaliana using large-scale gene expression measurements. BMC Bioinformatics 8, 358 (007). DOI: 0.86/ Received: November 5, 03; Revised: February 0, 04; Accepted: February 7, 04 Technology in Cancer Research & Treatment 04 April 6. Epub ahead of print