Abstract

The purpose of this study is to develop a technique for reducing the number of false positives affecting lung nodule computer–aided detection in computed tomography (CT) images. Contiguous 2D regions of interest found on segmented lung areas from sections of a CT scan are merged to form volumes of interest (VOIs). Feature vectors are then computed by submitting each VOI to the 3D ranklet transform, i.e., a non–parametric, orientation–selective and multi–resolution transform developed and evaluated herein. Finally, a support vector machine classifier is used to discriminate VOIs containing nodules from those containing normal tissue. The proposed approach is evaluated on data consisting of 25 nodules marked by experienced thoracic radiologists and 1048 non–nodules randomly selected within the segmented lung volume of healthy patients. By achieving 96% of sensitivity at 1% of false positive fraction, leave–one–out performances seem to be promising.