Deeply Supervised Multi-Scale Fully Convolutional Networks for Segmentation of White Matter Hyperintensities

We present a method to address the challenging problem of segmentation of White Matter Hyperintensities (WMH) from multimodality MR images (T1 and FLAIR). Our method is based on deeply supervised multi-scale fully convolutional networks (FCNs), that are executed in two sequential stages and can directly map a whole volumetric data to its volume-wise labels. In order to alleviate the potential gradient vanishing problem during training, we designed multi-scale deep supervision. Validated on the 60 training datasets of the MICCAI 2017 Grand Challenge on WMH segmentation, our method achieved an average Dice Similarity Coefficient of 74.6%, an average precision of 69.0% and an average recall of 82.5%, when we randomly chose 45 datasets for training and tested our method on the remaining 15 datasets.