Multi-scale and Modality Dropout Learning for Intervertebral Disc Localization and Segmentation

Abstract

Automatic localization and segmentation of intervertebral discs (IVDs) from volumetric magnetic resonance (MR) images is important for spine disease diagnosis. It dramatically alleviates the workload of radiologists given that the traditional manual annotation is time-consuming and error-prone with limited reproducibility. Compared with single modality data, multi-modality MR images are able to provide complementary information. However, how to effectively integrate them to generate more accurate segmentation results still remains open for studies. In this paper, we introduce a multi-scale and modality dropout learning framework to segment IVDs from four-modality MR images. Specifically, we design a 3D fully convolutional network which takes multiple scales of images as input and merges these pathways at higher layers to jointly integrate multi-scale information. Furthermore, in order to harness the complementary information from different modalities, we propose a modality dropout strategy to alleviate the co-adaption issue during the training. We evaluated our method on the MICCAI 2016 Challenge on Automatic Intervertebral Disc Localization and Segmentation from 3D Multi-modality MR Images. Our method achieved the best overall performance with the mean segmentation Dice as 91.2% and localization error as 0.62 mm, which demonstrated the superiority of our proposed framework.