SIP-UNet: Sequential Inputs Parallel UNet Architecture for segmentation of Brain Tissues from Magnetic Resonance Images

Our work introduces a novel architecture for extracting information from neighboring slices in the image segmentation. The Proposed architecture consists of parallel UNets for each slice/image. These UNets are fused using Residual Network for late fusion. We passed features from central slice only in the residual network so that the features from the image to be segmented will not be lost. We performed segmentation using this architecture on the brain Magnetic Resonance Images (MRI). The processed data is obtained from the Open Access Series of Imaging Studies (OASIS) dataset. Data of each subject is in 3D. We extracted 2D slices/images from each subject and train/test for each view (axial, coronal, sagittal). As the 3D data has neighboring slices, that are similar to each other, we hypothesize and considered before and after slices as neighboring slice for segmentation of central slice.