Residual and Dense Connection Combine Fully Convolutional Network for Infant Brain MRI Segmentation

Abstract

In \u00a0the \u00a0development \u00a0of \u00a0medical \u00a0image \u00a0segmentation, \u00a0the \u00a0application \u00a0of \u00a0convolutional \u00a0neural networks \u00a0has \u00a0begun \u00a0a \u00a0profound \u00a0revolution. \u00a0The \u00a0deep \u00a0learning \u00a0model \u00a0is \u00a0famous \u00a0for \u00a0excellent \u00a0flexibility, efficiency and accuracy. The U-Net model is the beginning of task in the segmentation of medical images, which includes the basic operations of convolution, maxpooling, deconvolution, and concatenation. However, the \u00a0U-Net model is disable \u00a0to perform well on many \u00a0types \u00a0of \u00a0data \u00a0sets, \u00a0because \u00a0the model can\u2019t \u00a0solve \u00a0the exact segmentation of the details. We proposed Residual and Dense Fully Convolutional Network (RDFCN) that \u00a0consist \u00a0of \u00a0Residual \u00a0Connection \u00a0Block \u00a0and \u00a0Dense \u00a0Connection \u00a0Block, \u00a0which \u00a0makes \u00a0up \u00a0for \u00a0the shortcomings \u00a0of \u00a0U-Net. \u00a0The \u00a0dataset \u00a0we \u00a0used \u00a0for \u00a0training \u00a0and \u00a0testing \u00a0comes \u00a0from \u00a0iSeg-2017 \u00a0challenge (http://iseg2017.web.unc.edu). This dataset is comprised of infant(between 6 and 9 months of age) brain MR images. After the testing, our model outperforms the U-Net and some of its improved models in evaluation of WM, GM and CSF.