Learning 3D Features with 2D CNNs via Surface Projection for CT Volume Segmentation

Abstract

3D features are desired in nature for segmenting CT volumes. It is, however, computationally expensive to employ a 3D convolutional neural network (CNN) to learn 3D features. Existing methods hence learn 3D features by still relying on 2D CNNs while attempting to consider more 2D slices, but up until now it is difficulty for them to consider the whole volumetric data, resulting in information loss and performance degradation. In this paper, we propose a simple and effective technique that allows a 2D CNN to learn 3D features for segmenting CT volumes. Our key insight is that all boundary voxels of a 3D object form a surface that can be represented by using a 2D matrix, and therefore they can be perfectly recognized by a 2D CNN in theory. We hence learn 3D features for recognizing these boundary voxels by learning the projection distance between a set of prescribed spherical surfaces and the object’s surface, which can be readily performed by a 2D CNN. By doing so, we can consider the whole volumetric data when spherical surfaces are sampled sufficiently dense, without any information loss. We assessed the proposed method on a publicly available dataset. The experimental evidence shows that the proposed method is effective, outperforming existing methods.