3D Fully Convolutional Networks for Intervertebral Disc Localization and Segmentation

Abstract

Accurate localization and segmentation of intervertebral discs (IVDs) from volumetric data is a pre-requisite for clinical diagnosis and treatment planning. With the advance of deep learning, 2D fully convolutional networks (FCN) have achieved state-of-the-art performance on 2D image segmentation related tasks. However, how to segment objects such as IVDs from volumetric data hasn’t been well addressed so far. In order to resolve above problem, we extend the 2D FCN into a 3D variant with end-to-end learning and inference, where voxel-wise predictions are generated. In order to compare the performance of 2D and 3D deep learning methods on volumetric segmentation, two different frameworks are studied: one is a 2D FCN with deep feature representations by making use of adjacent slices, the other one is a 3D FCN with flexible 3D convolutional kernels. We evaluated our methods on the 3D MRI data of MICCAI 2015 Challenge on Automatic Intervertebral Disc Localization and Segmentation. Extensive experimental results corroborated that 3D FCN can achieve a higher localization and segmentation accuracy than 2D FCN, which demonstrates the significance of volumetric information when confronting 3D localization and segmentation tasks.