False-Positive Reduction of Pulmonary Nodule Detection Based on Deformable Convolutional Neural Networks
Haiying Yuan, Zhongwei Fan, Dong Ding, Zengyang Sun
Abstract
As a crucial component of a computer-aided diagnosis (CAD) system, the false-positive reduction plays an important role in the timely diagnosis of pulmonary nodules. Own to the similarity of the true and false-positive nodules in early morphology, it's a huge challenge to distinguish exactly between these two nodules. Hence, a novel convolutional neural network (CNN) framework based on the residual network is constructed to address this thorny issue. The deformable convolution component is performed on Computed Tomography (CT) images to adaptively reflect different spatial information, and the deformable feature images can reflect the complex structure appropriately. This efficient Deformable Convolutional Neural Networks (DCNN) model has been performed on the Lung Nodule Analysis 2016 dataset, which achieves an average competitive performance metric score of 0.835, and the excellent sensitivity of 0.941 and 0.958 occur to 4, 8 false-positive per scan.