A Computer Aided Pulmonary Nodule Detection System Using Multiple Massive Training SVMs

A computer aided pulmonary nodule detection system for chest radiography is proposed. The system consists of three models, viz., lung segmentation, lung nodule candidates detection and false positive reduction. Several innovations are offered in this system. The first one is that the detection of potential lung nodule candidates is conceived as a filtering process that searches for any region which has a spherical structure (where a potential nodule may happen to occur) in a chest radiograph by analysis eignvalues of Hessian matrix of an image. The second one is that a novel two stage classifier based method is developed to address the problem of false positive reduction in lung nodule detection, in which a rules based classifier is followed by a filter termed as multiple massive training supported vector machine (MTSVM), where the rules based classifier is employed to quickly remove obvious FP (outliers) so that their influence on the training of MTSVM was eliminated, the MTSVM is developed to further separate nodules from nonnodule candidates. Experimental results suggest that the proposed CAD scheme was superior to others in FPs reduction of lung nodule detection in chest radiography.