Fully automatic localization and segmentation of intervertebral disc from 3D multi-modality MR images by regression forest and CNN

Abstract

In this paper, we propose a fully automatic framework to localize and segment intervertebral discs (IVDs) from 3D Multi-modality MR Images. Random forest regression is employed to coarsely localize the IVD. Then IVDs are segmented sequentially by training the specific convolutional neural network (CNN) classifier for each IVD. We compared the performance using single- and multi-modality images. Evaluated on the MICCAI 2016 IVD on-site challenge datasets, our method achieved a mean localization distance of 0.64mm and a mean Dice overlap coefficient of 90.8%. The results show that our method is robust and comparable with state-of-the-art methods.