An Automatic Segmentation of Subregions of the Hippocampus with Unet Architecture

The automatic segmentation of the human hippocampus is a promising area of research in cognitive neuroscience. Our objective is to develop an automatic and robust technique that segments the hippocampus from MR images in real time. This article introduces a fully automatic method utilizing CNN model priors for hippocampus MRI segmentation. The method leverages variable training datasets from the NeuroImaging Tools & Resources Collaboratory (NITRC), annotated by ITK-SNAP software. The proposed CNN architecture, inspired by the U-net model, minimizes the number of layers and feature maps. A key advantage of our method is the elimination of the region of interest (ROI) extraction phase, as our datasets contain only the hippocampus, unlike other approaches that use full brain images. Additionally, our method can detect more than six hippocampal sub-regions in coronal and axial views within a test time of approximately 23 seconds. Experimental results suggest that our approach achieves promising outcomes by detecting hippocampal sub-regions rather than treating the hippocampus as a single entity.