Accurate and consistent hippocampus segmentation through convolutional LSTM and view ensemble

Abstract

In this work, a novel deep neural network is developed to automatically segment human hippocampi from MR images. To take advantage of the efficiency of 2D convolutional operations, as well the inter-slice dependence within 3D volumes, our model stacks fully convolutional neural networks (CNN) through convolutional long short-term memory (CLSTM) to extract voxel labels. Enhanced slice-wise label consistency is ensured, leading to improved segmentation stability and accuracy. We apply our model on ADNI dataset, and demonstrate that our proposed model outperforms the state-of-the-art solutions.