Attention 3D Fully Convolutional Neural Network for False Positive Reduction of Lung Nodule Detection

Deep Learning based lung nodule detection is rapidly growing. It is one of the most challenging tasks to increase the true positive while decreasing the false positive. In this paper, we propose a novel attention 3D fully Convolutional Neural Network for lung nodule detection to tackle this problem. It performs automatic suspect localization by a new channel-spatial attention U-Network with Squeeze and Excitation Blocks (U-SENet) for candidate nodules segmentation, following by a Fully Convolutional C3D (FC-C3D) network to reduce the false positives. The weights of spatial units and channels for U-SENet can be adjusted to focus on the regions related to the lung nodules. These candidate nodules are input to FC-C3D network, where the convolutional layers are re-placed by the fully connected layers, so that the size of the input feature map is no longer limited. In addition, voting fusion and weighted average fusion are adopted to improve the efficiency of the network. The experiments we implement demonstrate our model outperforms the other methods in the effectiveness, with the sensitivity up to 93.3 $$\%$$ .