3D Tiled Convolution for Effective Segmentation of Volumetric Medical Images

Abstract

Convolutional Neural Networks (CNNs) have achieved remarkable performance in many 2D computer vision and medical image analysis tasks. In clinical practice, however, a large part of the medical imaging data available are in 3D. This has motivated the development of 3D CNNs in order to benefit from more spatial context. Although weight sharing in CNNs significantly reduces the number of parameters that have to be learned, state-of-the-art 3D methods still depend on patch processing due to GPU memory restrictions caused by moving to fully 3D. The size of the input patch is usually small if no specialized hardware with large GPU memory is used, limiting the incorporation of larger context information for a better performance. In this paper, we propose a 3D Tiled Convolution (3D-TC) which learn a number of separate kernels within the same layer. 3D-TC has the advantage of significantly reducing the required GPU memory for 3D medical image processing task but with improved performance. Results obtained from comprehensive experiments conducted on both hip T1 MR images and pancreas CT images demonstrate the efficacy of the present method. Our implementation can be found at: https://github.com/guoyanzheng/LPNet.