A novel CNN-based Computer-Aided Detection system for Lung Cancer
Since ConvNets can be trained end to end in a supervised fashion while learning highly discriminative features, removing the need for handcrafting nodule descriptors, they are well suited to be used in pulmonary nodule CAD system. We will use the large number of CT cases available in the NLST to train some novel ConvNets to (1) detect the candidate nodules in CT, (2) reduce the false positive, and (3) predict the probability of lung cancer according to the various features of detected nodules.
(1) Demonstrate that the suspected pulmonary nodules can be auto segmented with high precision without pre-specified feature sets, and the probability of suffering lung cancer can be predicted directly according to the various characteristics of candidate nodules.
(2) Use the mass datasets to train a novel 2D nodule detection system, which can predict the candidate bounding boxes of the suspected pulmonary nodules locating in some particular 2D slices of CT scans, so these bounding boxed can help doctors to fix lesion areas easily and efficiently and make precise clinical decisions.
(3) Exploit 2D ConvNets to nodule detection and false positive reduction, and determine the optical model parameters for more accurate segmentation of lesions.
(4) Use ConvNets to analysis the various characteristics of nodules from the large amount of available data in NLST, and build an end-to-end ConvNets, which can predict the probability of suffering lung cancer directly under a set of pulmonary nodule CT scans.
Fengyan wang, postgraduate, college of Information Science and Technology, University of Science and Technology of China, China