Hippocampus segmentation for preterm and aging brains using 3D densely connected fully convolutional networks

Abstract

Efficient and accurate segmentation of hippocampi from preterm and aging brain MR images is one of the most fundamental steps in understanding hippocampal growth and development or diagnosing and monitoring various clinical conditions. Current hippocampus segmentation methods for preterm and aging brain are limited due to: 1) they can rarely achieve preterm infant hippocampus segmentation; 2) the computation cost is high; 3) current deep learning models cannot well handle the hippocampal feature learning; 4) they are not open obtainable. To deal with these problems, we propose an efficient, open-source algorithm, 3D densely connected fully convolutional network (3D-DCFCN) for the infant and aging hippocampal segmentation. Specifically, we search for a suitable distribution of the hierarchical receptive field size and a joint loss function to balance local and global information and lead to better optimization. In addition, we use image cross-registration for vast augmentation of the infant training data and incorporate multi-modality infant brain information. We compare the performance of our algorithm with those of several state-of-the-art methods. The results show that our method outperforms all comparison methods on infant and aging datasets and achieves much faster speed (less than 0.11s per image). We also provide a notably comprehensive evaluation of the method. Our experiments further demonstrate our model can 1) well generalize to the dataset with different magnetic fields; 2) satisfactorily find hippocampal atrophy in cognitive-decline groups compare with normal controls.