A new computationally efficient CAD system for pulmonary nodule detection in CT imagery.

摘要: Early detection of lung nodules is extremely important for the diagnosis and clinical management cancer. In this paper, a novel computer aided (CAD) system pulmonary in thoracic computed tomography (CT) imagery presented. The paper describes architecture CAD assesses its performance on publicly available database to serve as benchmark future research efforts. Training tuning all modules our done using separate independent dataset provided courtesy University Texas Medical Branch (UTMB). testing that created by Lung Image Database Consortium (LIDC). LIDC data used here comprised 84 CT scans containing 143 ranging from 3 30mm effective size are manually segmented at least one four radiologists. uses fully automated segmentation algorithm define boundaries regions. It combines intensity thresholding with morphological processing detect segment nodule candidates simultaneously. A set 245 features each candidate. sequential forward selection process determine optimum subset two distinct classifiers, Fisher Linear Discriminant (FLD) classifier quadratic classifier. comparison between classifiers presented, based this, FLD selected system. With an average 517.5 per case/scan (517.5+/-72.9), proposed front-end detector/segmentor able 92.8% LIDC/testing (based merged ground truth). mean overlap regions delineated three or more radiologists ones approximately 63%. Overall, specificity false positives (FPs) case/patient average, correctly identify 80.4% (115/143) 40 features. 7-fold cross-validation analysis only shows sensitivity 82.66% FPs scan/case.