Massive Training Radial Basis Function Neural Network for Distinguishing Between Nodule and Non-Nodule

Abstract

Early detection of lung nodules is extremely important for the diagnosis and clinical management of lung cancer. In this paper, a novel computer aided detection (CAD) system for the detection of pulmonary nodules in conventional chest radiograph is presented. The proposed approach is based on radial basis function neural network. The massive training radial basis functions (MTRBFNN) is presented for classification between nodule and non-nodule. The MTRBFNN is trained by large number of overlapped sub-regions which are extracted from regional of interest (ROI). The efficiency of the MTRBFNN was assessed by ROC curves. The ROC curve shows the total sensitivity as a function of the number of non-nodules (false positives) at a certain point on the curve per image. When the MTRBFNN was applied, FPs decreased so that at some special operating points on the ROC curve, the reduction was up to 18% (99/550). The MTRBFNN is able to reduce false-positive rate in this paper from 3.93 (550/140) to 0.71 (99/140) false positives per image, and in total, gained a sensitivity of 92% (129/140).