Attribute-Based Assessment of Lung Nodules in CT Using Support Vector Machine and Random Forest

An attempt to provide a tool for detection of certain attributes of pulmonary nodules is presented in this paper. The support vector machine and random forest are employed to determine the nodule calcification and likelihood of malignancy on the basis of ten intensity and geometric features. Training and validation relies on lung nodule cases from the public LIDC-IDRI database with over a thousand computed tomography studies. Lesion annotation provided by four radiologists in terms of delineation and quantitative assessment of selected attributes yields ca. 2500 nodules available for the analysis. In both classifications involving two classifiers the accuracy exceeding \(80\%\) was achieved.