Segmentation of lungs nodules by iterative thresholding method and classification with Reduced Features

For lungs cancer detection, manual detection of lung nodules is critical for a radiologist if the size of the nodules are less. Also, due to excessive work pressure, some nodules may be missed during detection. To overcome these situations, an automatic decision making system is required. This paper presents an automatic CAD system based on CT-scan images. The proposed system uses iterative based thresholding method instead of the traditional histogram based thresholding methods. After extraction of the lungs region, nodules present within the region are detected by rule-based filtering method. From the extracted features of the nodules, feature selection process is applied to select the important features for better classification and to reduce the complexity. The proposed method is validated on 142 patients consisting of 500 cases, out of which 306 are nodule candidate. CT-scan images are taken from publically available large dataset LIDC-IDRI (Lung Image Database Consortium and Image Database Resource Initiative) and the images are in DICOM (Digital Imaging and Communications in Medicine) format. This system used 10 k-fold cross-validations to evaluate classifier performance. The proposed method gives an accuracy of 99.4% with polynomial kernel based support vector machine (SVM) classifier and gives 99.96% accuracy using k-nearest neighbourhood (KNN) classifier.