Dual dense context-aware network for hippocampal segmentation

Abstract The morphological analysis of hippocampus based on magnetic resonance imaging is significant for computer-aided neurological disease diagnosis. The segmentation of the hippocampus is a prerequisite for calculating its volume and measuring its shape. In this study, we propose a dual dense context-aware network (DDCNet) for automatic hippocampal segmentation, mainly including a multi-scale input approach and multi-resolution feature fusion. First, considering that a single-scale network cannot accurately segment the hippocampus and background with large difference in volume, we design a multi-scale input module (MSIM) to extract the hippocampus features of different receptive fields. Specifically, large scales are used to learn spatial dependencies, while smaller scales are used to learn more hippocampal details. Taking advantage of the fact that the features from different resolution layers offer semantically different information for segmentation, we develop a multi-resolution feature fusion module (MRFFM) to effectively aggregate complementary features and to simultaneously utilize the high-resolution features from the encoder to guide the low-resolution hippocampus edge segmentation in the decoder phase. In addition, many dense connections are introduced to further extend the feature fusion and contextual information of various resolutions between the encoder and decoder. The proposed approach achieves state-of-the-art performance, outperforming existing hippocampal segmentation approaches on the ADNI dataset.