MDNet: Multi-scale densely connected segmentation architecture with channel attention for automatic hippocampal segmentation

Morphological analysis of the hippocampus based on magnetic resonance imaging is of great significance for the computer-aided diagnosis of neuropsychiatric diseases. In this study, we propose a multi-scale densely connected segmentation architecture for automatic hippocampal segmentation. First, consider that single-scale feature extraction cannot accurately segment the hippocampus from the background. We design a multi-scale densely connected module to extract contextual information at different scales and levels and learn the spatial dependencies and details of the hippocampus through different scales. Then, channel attention is used to model the correlation between features at different scales, effectively highlighting the features of the hippocampus and suppressing irrelevant background information. Finally, multi-scale feature fusion is performed to achieve accurate segmentation of the hippocampus. Validation results on the ADNI hippocampus dataset show that the model has a DICE coefficient score of 0.9283, outperforming other classical segmentation methods.