A Multi-Scale Fully Convolutional Networks Model for Brain MRI Segmentation

Abstract

1School of math and statistics, Nanjing University of Information Science & Technology, Nanjing 210044, China (Received October 01 2017, accepted January 15 2018) Abstract \u3002 Accurate \u00a0segmentation \u00a0for \u00a0brain \u00a0magnetic \u00a0resonance \u00a0(MR) \u00a0images \u00a0is \u00a0of \u00a0great \u00a0significance \u00a0to quantitative analysis of brain image. However, traditional segmentation methods suffer from the problems existing in brain images such as noise, \u00a0weak edges and intensity inhomogeneity (also named as \u00a0bias field). Convolutional neural networks based methods have been used to segment images; however, it is still hard to find accurate results for \u00a0brain \u00a0MR \u00a0images. \u00a0In \u00a0order \u00a0to \u00a0obtain \u00a0accurate \u00a0segmentation \u00a0results, \u00a0a \u00a0multi-scale \u00a0fully \u00a0convolution \u00a0networks model \u00a0(MSFCN) \u00a0is \u00a0proposed \u00a0in \u00a0this \u00a0paper. \u00a0First, \u00a0we \u00a0use \u00a0padding \u00a0convolutions \u00a0in \u00a0conv-layer \u00a0to \u00a0preserve \u00a0the resolution of feature maps. So we can obtain segmentation results with the same resolution as inputs. Then, different sized filters are utilized in the same conv-layer, after that, the outputs of these filters are concatenated together and fed \u00a0to \u00a0the \u00a0next \u00a0layer, \u00a0which \u00a0makes \u00a0the \u00a0model \u00a0learn \u00a0features \u00a0from \u00a0different \u00a0scales. \u00a0Both \u00a0experimental \u00a0results \u00a0and statistic results show that the proposed model can obtain more accurate results.