CLASSIFICATION OF BREAST CANCER BASED ON IMAGE mammogram TEXTURE FEATURE
ABSTRACT: Breast cancer is the number one killer of women in the world. By the early detection of breast cancer the life expectancy for the patient can be improved about 95%. In the United States the early detection of breast cancer patients were able to save the lives approximately 12 to 37 people per day. While in Indonesia, according to the health profile of the Ministry of Health of the Republic of Indonesia Year 2012 the highest cancer which is suffered by Indonesian women was breast cancer with the incidence rate of 2,2% per 1000 women. If this condition is not controlled, in 2030 there will be 26 million people suffering of breast cancer and 17 million people will die. Detection of breast cancer by radiologists all this time was based on direct observation by naked eye and experience of the radiologist. By using that method, there were weaknesses in analyzing the mammogram image because there were some condition that were difficult to distinguish using the direct observation, because of the noise and low contrast images and other human factors, such as fatigue, mood, etc. This condition causes the need for a method that is able to assist radiologist in diagnosing breast screening results more precisely. The higher the value of accuracy is found, it will be the smaller abnormality detection errors that occur thereby it can minimize the circumstances which could endanger the lives of patients. Based on the data obtained, there were approximately 5-10% misdiagnoses in analyzing the mammogram image. This study used the intact image of the breast without crop on the ROI and without segmentation, then did the pre-processing and texture-based feature extraction to obtain accurate classification results. Classifications of breast abnormality in this study were divided into three classes: normal, benign and malignant. Phase of the study consisted of a pre-processing stage, the removal of a label or artifacts that was found on the mammogram image, contrast enhancement by comparing several methods (CLAHE, NSCT, CS) to obtain the best method, image cropping, filtering and smoothing with a median filter to the image. Furthermore, the output of the pre-processing stage will be input to the feature extraction stage. The Feature extraction that was used in this study were texture-based namely GLCM, GLRLM and histogram. The results from feature extraction will be input to classification process. The classification process used the MLP with BEP learning. For classification, by testing techniques using k-fold cross validation, the optimal values was obtained at 10th fold. MLP was compared to other classifiers namely SVM and Na������¯ve Bayes. The maximum value of classification results was obtained using CLAHE contrast enhancement methods, GLCM-based feature extraction method and MLP classifier with accuracy of 98,33%, sensitivity of 100% and a specificity of 97,5% for 2 classes normal and abnormal. If we used hierarchical classification, then performed more abnormal classification (benign and malignant), we obtained the accuracy of 82,5%, sensitivity of 80% and specificity of 85%. While for the classification of three classes (normal, malignant and benign) directly without hierarchy obtained accuracy of 90%, sensitivity of 85% and specificity of 87,5%. In this study, to examine the influence of enhancement process (contrast improvement and filtering), we conducted a two-class classification of normal and abnormal images without enhancement. By that examination, we obtained that the results were worse if the image didn���¢�¯�¿�½�¯�¿�½t begin with the enhancement process namely accuracy of 83,333%, sensitivity of 92,5% and specificity of 80%.
