Automated detection and segmentation of pulmonary nodules on lung computedtomography (CT) scans can facilitate early lung cancer diagnosis. Existingsupervised approaches for automated nodule segmentation on CT scans requirevoxel-based annotations for training, which are labor- and time-consuming toobtain. In this work, we propose a weakly-supervised method that generatesaccurate voxel-level nodule segmentation trained with image-level labels only.By adapting a convolutional neural network (CNN) trained for imageclassification, our proposed method learns discriminative regions from theactivation maps of convolution units at different scales, and identifies thetrue nodule location with a novel candidate-screening framework. Experimentalresults on the public LIDC-IDRI dataset demonstrate that, our weakly-supervisednodule segmentation framework achieves competitive performance compared to afully-supervised CNN-based segmentation method.
展开▼
