Abstract:
Computer aided systems has a crucial importance on lung nodule studies, since lung cancer is the leading cause of cancer related death for both men and women worldwide. Accurate characterization of lung nodules as malignant or benign may be di cult. CAD can assist radiologists in improving the accuracy of classi cation. The computer-assisted characterization of lung nodules involves several steps including segmentation, feature extraction and classi cation. In this study, we aim to optimize each step in order to improve the overall accuracy through on classi cation accuracy. The main objective of this study is to improve the characterization of detected nodules on chest x-rays by performance comparison of algorithms and optimum selection of classi er parameters. In this study, 154 posteroanterior chest x-ray images included in JSRT Database were used as test materials. The database consists of 100 malignant and 54 benign nodules. Our system involves pre-processing, detection, segmentation, feature extraction and classi cation steps. The aim of the pre-processing was to improve the quality of the images by contrast enhancement and noise reduction. We have determined 14 features (morphological features, statistical features and textural features) from each segmented nodule to make the classi cation more e cient. Initially in this work we have used k-nearest neighbor classi er and fuzzy classi er to classify the nodules as malignant or benign. We have tested the algorithm for di erent parameter values. According to our initial results, the optimal accuracy for k-NN classi cation is 68.8% and for the fuzzy classi cation it is 61.3%. The initial results reveal that this methodology has the potential to assist radiologists as a second opinion tool in the classi cation of benign and malignant lung nodules.|Keywords : CAD,Chest X-ray, Lung Nodule Characterization, Fuzzy Classi cation.
