DIAGNOSIS OF TUBERCULOSIS USING MATLAB BASED ARTIFICIAL NEURAL NETWORK

Transcription

1 IJIPA: 3(1), 2012, pp DIAGNOSIS OF TUBERCULOSIS USING MATLAB BASED ARTIFICIAL NEURAL NETWORK Chandrika V. *, Parvathi C.S., and P. Bhaskar Department of Instrumentation Technology, Gulbarga University P.G Centre, Yeragera , Raichur, Karnataka, India Abstract: This paper deals with the diagnosis of Pulmonary Tuberculosis using MatLab based artificial neural network. The architecture of artificial neural network is three layered (15-9-2) Back Progression. The artificial neural network model with shape features and symptoms are used in diagnosing pulmonary tuberculosis i.e. here we are using two features to train the neural network i.e. Shape and symptoms. The model is designed in such a way that the presence of Tuberculosis (TB) is detected and result is displayed. The whole system is designed on the MatLab 7.0 version platform. The Dicom formatted X-ray images are read by converting them into MATRIX format. Then from these images shape features are extracted. The extracted features are fed to the neural network which is trained before these features are fed. Depending on the features and symptoms of the read image, the neural network detects the presence of TB. The model was applied to the validating sample, with accuracy, sensitivity and specificity at 71.25%, 73.68%, and 69.05% respectively. A Graphical User Interface (GUI) has been developed to read the image, to select the region of interest, to process the image, to plot histogram and finally to display the result whether the patient is having TB or Not. These are the fine features of our model. Keywords: Pulmonary Tuberculosis, Neural Network, Back Propagation, TB Symptoms, X-ray, Artificial Neural Network. 1. INTRODUCTION Tuberculosis (TB) is one of the most important public health problems worldwide. There are 9 million new TB cases and nearly 2 million TB deaths each year [1]. Case-finding and the management of pulmonary tuberculosis is an essential target of tuberculosis control programs. However, pulmonary tuberculosis (PT) is becoming more and more of a serious problem, particularly in countries affected by epidemics of human immunodeficiency virus (HIV)-TB co-infection [2]. The diagnosis of PT using prompt and accurate methods is a crucial step in the control of the occurrence and prevalence of TB. However, the diagnosis of PT is quite complex, so there is no unified standard at present. Frequently, there is over diagnosis and missed diagnosis and it is a thorny question in the field of TB control. Some of the methods used earlier are based on distance or pair wise distance measurement, and their performance is around 60% to 65% [3]. * crgolds@rediffmail.com Artificial neural network (ANN) is theoretical mathematical model acting like human brain which is one kind of information management system based on the imitation of cerebrum neural network architecture and the function [4]. ANN has the functions of self-learning, the associative memory, and highly parallel, fault-tolerant and formidable non-linearity handling ability [5] and can make rational judgment to complex questions according to obtained knowledge and the experience of handling problems. ANNs have been applied in the fields of signal processing, pattern recognition, quality synthetic evaluation, forecast analysis, etc. [6] This study seeks to develop a diagnostic model of TB that is based on ANN to explore the feasibility of it in TB diagnoses. A GUI program is a graphical based approach to execute the program in a more user friendly way. It contains components such as push buttons, text boxes, radio buttons, pop-up menus, slider etc. with proper labels for easy understanding to a less experienced user. These components help the user to easily understand how to execute or what to do

2 38 Chandrika V. Parvathi C.S. and P. Bhaskar to execute the program. When an user responds to a GUI's components by pressing a pushbutton or clicking a check box or radio button or by entering some text using text box, the program reads the necessary information for that particular event, hence GUI programs are also known as event driven programs. MATLAB provides a tool called GUIDE (GUI Development Environment) for developing GUI programs. GUI approach is employed in various fields. In some systems GUI is built to facilitate users to apply the developed system and understand hierarchy. GUI that acts as an intermediate media creates a form of communication between users and the developed object detection system. 2. SUBJECTS AND METHODS 2.1 Methodology The block diagram of the proposed setup is as shown in Figure 1. Initially the neural network is trained with shape features and symptoms with more number of X-ray images (data base). In figure 1, we observe that the data i.e. the images are read and are converted into matrix format. Then from these images shape features are collected. This stage is called as feature collection. Then these features are fed to the neural network for testing. In this stage the features of testing images are compared with the extracted features of the neural network during training process. After once the symptoms are fed the neural network fetches the result as TB or NON-TB on the window of the PC. (There is one more stage here i.e. database to neural network, in this stage training of the neural network takes place with symptoms & shape features). A GUI has been developed to perform all the processes i.e. reading the image, training the neural network, for drawing the histogram and to display the result. The GUI model is shown in figure 2. We compared 27 cases with a final diagnosis of TB to 28 non-tb cases, for a total of 55 patients in the modeling sample. 38 TB cases and 42 non-tb cases were used for the validation sample. In the modeling sample were randomly assigned into two groups, one for training and one for testing in a 4:1 ratio, respectively. Using the training sample adopt advisor study system, we obtained the distinction function through the training. We then distinguished the unknown sample category using these distinction functions, while the testing sample was used to examine the reliability of the recognition function, which was obtained from the training sample. This is how the ANN network architecture and the judgment training end points were determined. The 25 cases in the validation sample were used to evaluate the generalizability of the network. We evaluated network diagnosis performance using the sensitivity and the specificity. Figure 1: Block Diagram of the Proposed Setup Figure 2: Graphical User Interface (GUI) 2.2 ANN MODEL DESIGN Figure 3 shows the ANN model design. Figure 3: ANN Training Network

3 Diagnosis of Tuberculosis Using Matlab Based Artificial Neural Network The Network Type and the Layer The Back-Propagation Network is a multi-layered forward feed network for the weight training of nonlinear differentiable functions. The BP network mainly is used for approximation of functions, pattern recognition, classification, the data compression. In the practical application of ANN, 80%-90% of the ANN model adopted the BP network or its variations. The BP network is also central to the forwarding network and constitutes the most vital element of the ANN. ANN with one hidden layer can be used for approximation for any closed interval, continuous function. Therefore, a three-layered (including input layer) BP network may complete the random n dimension to m dimension mapping. Therefore this analysis uses a three-layered BP network with one hidden layer Input and Output Variable Choice Training samples were analyzed using single factor Logistic regression, screening significant parameters for TB diagnosis as input variable. Parameters identified in this analysis included the shape variables and symptoms. The network output has two kinds: the first kind is the TB group, for which the expected export value is 1; the second kind is the non-tb group, for which the expected export value is Number of Hidden Level Neurons Determining the number of hidden layer neurons is a very complex issue. Because of the lack of a strong analytical formula for calculating this value, in the past, this was often determined simply according to designer s experience and repeated trials. To address this in my research, we designed a BP network with a hidden layer with variable neuron in order to determine best number of hidden layer neurons through comparisons of errors Activation Function Activation function is central to both the neuron and the network. The capacity and efficiency of a network to solve questions depend on the activation function which used in the network, to a great extent beside related to the network architecture. The Sigmoid activation function has the function of nonlinearity magnification to coefficient; it can transform the signal from an input of 8 to 8, to an output of 1 to 1. Because the magnification coefficient is smaller for larger input values and bigger for smaller input values. As such, we chose to use the Sigmoid activation function The Pretreatment of Clinic Data Different parameters used in diagnoses had different expression methods and dimensions, and there was a significant difference between their ranges. If raw data were directly input into the neural network, the network would adjust weight primarily in accordance with data whose numerical values are greater. So the frequency of error did not reflect the data whose numerical values were smaller. So raw data had to be changed into those fit for neural network by means of pretreatment to improve the learning ability and astringency function of the neural network. It was also important to normalization, pretreated input data for the network, which used the sigmoid excitation function and error back-propagation learn algorithm for raising their learning ability and generalization performance. The input data of network should be in the interval (0, 1)as our neural network understands values between 0-1, so 1 and 0 were used to indicate YES and NO for the binary variable data. So we normalize the raw data whose values vary from 0-255in to binary form whose values vary from 0-1. This is what is called as normalization. Normalization treatment is widely used for selection of quantitative data as follows: Where, x i = raw data y i xi x i max x x i min imin x imin = minimum pixel value of the raw data (1) x imax = maximum pixel value of the raw data. The data collected were used for the raw data matrix of the ANN diagnostic after they were quantitative and normalized according to this principle. 3. IMPLEMENTATION METHOD ANN was implemented with self-edited program using the Neural-Network-Toolbox in MATLAB Detecting the Lung Area To find the lung area and hence the shape of the lung, several stages of processing is applied to the image. By using a high pass FFT filter and suitable cutoff frequency, the soft tissue within the lungs can be isolated. The resulting high frequency areas are

4 40 Chandrika V. Parvathi C.S. and P. Bhaskar non-uniform, so the filtered image is first dilated to a suitable degree and then eroded to separate these uniform regions. A binary image is then generated by thresholding the image. Thresholding can introduce further non-uniformity to these regions but can be rectified by an additional step of dilation and erosion. An approximation of the lung area after performing these steps is shown by Flowchart 1. Many of the errors present in the processed image (especially the upper left and right hand corners) can be attributed to the high frequency nature of the patient text on the sample radiograph. 3.2 Rib Supression The density of the ribs affects the image by changing the luminance values of the underlying textures. This can affect the detection of nodules. A method for suppressing the contrast of the ribs and chest clavicles may be implemented using an algorithm such as the one suggested by K. Suzuki, H. Abe, H. MacMahon, K. Doi [7]. The previously suggested method obtains a representation of the bone structure in a radiograph by using a dual-energy subtraction technique. This involves a multi-kv (multi-kilovolt) radiographic analysis using two separate radiographs, each captured at a different kv rating. The generated bone structure is then used to train a classifier and suppress the ribs in a lung radiograph. 3.3 Shape Description We discriminated true TB using shape. Matching nearest-neighbor connected pixels were grouped; to account for varying bacillary orientations and magnification, we bypassed size calibration by employing two shape descriptors that were invariant to rotation, translation, skew transformations and scale: 1) axis ratio (1 for circles, higher for line segments) and 2) eccentricity, a ratio of distance between elliptical foci to major axis length (1 for line segments, 0 for circles). The typical axis ratio of for TB cavity was significantly different from approximately one for non-tb objects; similarly, TB eccentricity was and centered at zero for non-tb. To maximize rod-shaped object recognition, we empirically chose conservative threshold cut-offs (axis-ratio > 1.25 and eccentricity >0.65) as indicating TB. Objects below the thresholds were labeled red as non-tb objects. Calculating the mean TB size µ and standard deviation from a broth image, we labeled all size outliers µ ± 1.5 in blue as possible and within µ ± 1.5 in green as definite TB objects. Flowchart 1: Flowchart showing the classification steps for automatic identification and labeling of Tuberculosis. 3.4 Training and Testing Flowchart 2 and 3 explains the method of training and testing. During training the TB and NON-TB images are read then in the next stage noise removal takes place in the preprocessing stage then shape features are calculated. These vector values are fed to the neural network. This is repeated for all the images. This is how the neural network is trained. Flowchart 2: Flowchart Showing the Steps for Training of Database for ANN.

5 Diagnosis of Tuberculosis Using Matlab Based Artificial Neural Network 41 During testing same procedure is followed like training only, i.e. the features of the knowledge base are compared with the present features of the image and then the result is classified and displayed. accordance with TB features. Because we found that diagnosis on TB was influenced by 15 variables (Table 1), the input number was 15. We also found that the approach effect to function of BP net was best when the node number of the hidden layer was 9 after many trials. Because output variables were in the form of TB and NO-TB, the BP net framework was The network was trained by traingdm with tansig activation function for the hidden layer, and purelin linearity function for output. In this process, the target error was 0.01 and the biggest training time was Figure 4: GUI Showing the Result of Testing Image Above figure 4 shows the result of the given X-ray image after once the symptoms are added. Table 1 Variable and their Significance Sl. No Variables Sig. Flowchart 3: Flowchart Showing the Steps for Testing of Input Image 4. RESULTS AND DISCUSSIONS 4.1 Selection Results of The BP Network Architecture We used a three-layer BP network, including an input layer, an output layer, and a hidden layer, in 1. Rectangularity Circularity Sphericity Convexity Convexity Perimeter Cough Fever Weight Loss HIV Breathlessness Family History Past History Alcoholic Chest Pain Smoker 0.01

6 42 Chandrika V. Parvathi C.S. and P. Bhaskar 4.2 The Results of Training and Fit of BP Network The objective error, and largest training number of training samples were 0.01 and 1000 separately (Figure 5). Figure 5: Semi-Logarithmic Line Graph of Training Performance According to Table 2 Accurate rate, Sensitivity, and Specificity of (15-9-2) BP network diagnosis were 71.25% (57/80), 73.68% (28/38), and 69.05% (29/42), respectively. Table 2 Diagnostic Result of Testing Samples Diagnostic Status of disease Total result TB Non-TB TB Non-TB Total CONCLUSION Due to the complexity of TB diagnosis, there continues to be no unified standard for it. Over diagnosis and missed diagnosis are formidable problems in the process for TB control. The cost of new diagnostic methods, such as nucleic acid amplification tests is very high and the effectiveness of these tests has not been confirmed in developing countries. To aim directly at uncertainty information and artifacts in clinical diagnosis, the limitation of regression modeling can be overcome by the use of ANNs. Reasonable judgment, satisfactory predictions and ideal forecasts can be achieved by ANN based on existing knowledge and experiences in solving problems. It was confirmed that the sensitivity, specificity of TB diagnosis were 73.68%, and 69.05%, respectively by the (15-9-2)-BP network. These results indicate that the validity of diagnosis was good and the (15-9-2)-BP network could be further extended to new patient data. The results indicate that this could be used as a new diagnosis method for this complex problem. References [1] R.P. Tripathi, N. Tewari, N. Dwivedi, et al. Fighting Tuberculosis: An Old Disease with New Challenges. Med Res Rev, 2005, 25(1), [2] R. Colebunders, WE. Bastian. A Review of Diagnosis and Treatment of Smear-negative Pulmonary Tuberculosis. Int. J. Tuberc Lung Dis, 2000, 4, [3] S.A. Patil, and V.R. Udupi, Textile and Engineering Institute, Ichalkaranji, India Chest X- ray Features Extraction for Lung Cancer Classification, JSIR, 69, April 2010, pp [4] Y.J. Wu, Y.M. Wu, L.B. Qu, et al. Application of Artificial Neural Network in the Diagnosis of Lung Cancer. Chin J. Microbiol Immuno, 2003, 23(8), [5] F.E. Ahmed. Artificial Neural Networks for Diagnosis and Survival Prediction in Colon Cancer. Molecular Cancer, 2005, 4, 29. [6] W. Deng, P.H. Jin. Artificial Neural Networks and Its Applications in Preventive Medicine. Chin Pub Health, 2002, 18(10), [7] K. Suzuki, H. Abe, H. MacMahon, K. Doi, Image- Processing Technique for Suppressing Ribs in Chest Radiographs by Means of Massive Training Artificial Neural Network (MTANN), IEEE Transactions on Medical Imaging, 25(4), pp , 2006.