Nodule-Plus R-CNN and Deep Self-Paced Active Learning for 3D Instance Segmentation of Pulmonary Nodules

Abstract

Accurate and automatic segmentation of pulmonary nodules in 3D thoracic Computed Tomography (CT) images is of great significance for Computer-Aided medical Diagnosis (CAD) of lung cancer. Currently, this important task remains challenging for lack of the voxel-level annotation and training strategies that balance target/background voxels in thoracic CT images. In this paper, a new region-based network, called Nodule-plus Region-based CNN, is proposed to detect pulmonary nodules in 3D thoracic CT images effectively while synchronously generating an instance segmentation mask for every detected instance. Our new network is constructed with a stack of convolutional blocks in which lateral connections are used to alleviate the difficulty of vanishing gradients. In addition, in order to reduce annotation workload and make best use of unannotated samples, we proposed a new Deep Self-paced Active Learning (DSAL) strategy by combining Active Learning (AL) and Self-Paced Learning (SPL) strategies. For the purpose of evaluating the performance of our proposed Nodule-plus R-CNN, we conduct a series of experiments on the public LIDC-IDRI dataset, and our model achieves 0.66 Dice and 0.96 TP Dice, which are state-of-the-art best results of pulmonary nodule segmentation. When the amount of available annotated samples is limited, our model trained with the DSAL strategy performs much better than that trained with the standard strategy.

Keywords

• Pulmonary nodule segmentation
• 3D CT images
• R-CNN
• active learning
• self-paced learning