Brain segmentation based on multi-atlas and diffeomorphism guided 3D fully convolutional network ensembles

In this study, we proposed and validated a multi-atlas diffeomorphism guided 3D fully convolutional network (FCN) ensemble model (M-FCN) for segmenting brain anatomical regions of interest (ROIs) from structural magnetic resonance images (MRIs). A novel based encoding block ROI patches with adaptive sizes were used. the block, both MRI intensity profiles expert priors deformed atlases encoded fed to FCN. Utilizing enabled more efficient training testing. To incorporate local global contextual information specific ROI, employed long skip connection between layer encoding-decoding block. relieve over-fitting FCN on data, adopted an strategy in learning procedure. Systematic evaluations performed two datasets, aiming respectively at 14 subcortical ventricular structures 54 whole-brain ROIs. Compared state-of-the-art segmentation methods including method existing model, exhibited superior performance.