In this paper, we propose a novel framework with 3D convolutional networks(ConvNets) for automated detection of pulmonary nodules from low-dose CT scans,which is a challenging yet crucial task for lung cancer early diagnosis andtreatment. Different from previous standard ConvNets, we try to tackle thesevere hard/easy sample imbalance problem in medical datasets and explore thebenefits of localized annotations to regularize the learning, and hence boostthe performance of ConvNets to achieve more accurate detections. Our proposedframework consists of two stages: 1) candidate screening, and 2) false positivereduction. In the first stage, we establish a 3D fully convolutional network,effectively trained with an online sample filtering scheme, to sensitively andrapidly screen the nodule candidates. In the second stage, we design ahybrid-loss residual network which harnesses the location and size informationas important cues to guide the nodule recognition procedure. Experimentalresults on the public large-scale LUNA16 dataset demonstrate superiorperformance of our proposed method compared with state-of-the-art approachesfor the pulmonary nodule detection task.
展开▼
