Three-Dimensional CT Image Segmentation by Combining 2D Fully Convolutional Network with 3D Majority Voting

We propose a novel approach for automatic segmentation of anatomical structures on 3D CT images by voting from a fully convolutional network (FCN), which accomplishes an end-to-end, voxel-wise multiple-class classification to map each voxel in a CT image directly to an anatomical label. The proposed method simplifies the segmentation of the anatomical structures (including multiple organs) in a CT image (generally in 3D) to majority voting for the semantic segmentation of multiple 2D slices drawn from different viewpoints with redundancy. An FCN consisting of “convolution” and “de-convolution” parts is trained and re-used for the 2D semantic image segmentation of different slices of CT scans. All of the procedures are integrated into a simple and compact all-in-one network, which can segment complicated structures on differently sized CT images that cover arbitrary CT scan regions without any adjustment. We applied the proposed method to segment a wide range of anatomical structures that consisted of 19 types of targets in the human torso, including all the major organs. A database consisting of 240 3D CT scans and a humanly annotated ground truth was used for training and testing. The results showed that the target regions for the entire set of CT test scans were segmented with acceptable accuracies (89 % of total voxels were labeled correctly) against the human annotations. The experimental results showed better efficiency, generality, and flexibility of this end-to-end learning approach on CT image segmentations comparing to conventional methods guided by human expertise.