Room: Karl Dean Ballroom C
Purpose: Accurate segmentation of the hippocampus from magnetic resonance (MR) brain images is important in neuroimaging studies of brain disorders, such as brain cancer, epilepsy, and Alzheimerâ€™s disease. Although multi-atlas image segmentation methods have exhibited promising hippocampus segmentation performance, the hippocampus segmentation remains a challenging task owing to poor image intensity contrast between the hippocampus and its surrounding brain tissues in MR brain images. To achieve accurate and efficient segmentation of the hippocampus, we develop a novel deep learning hippocampus segmentation method.
Methods: Our deep learning based hippocampus segmentation method is built upon fully convolutional networks (FCNs) and residual convolutional neural networks (CNNs). Particularly, 3D residual network blocks are used to extract informative features for segmentation. We also develop an image registration based data augment method to generate a large number of training data by registering images with manual segmentation labels to other images without the segmentation labels. Our deep learning segmentation model was trained based on 35 MP-RAGE T1 MRI scans with manually labeled hippocampi and validated based on a set of 99 scans with manual hippocampal labels. We also compared our method with multi-atlas based image segmentation methods, including non-local patch based label fusion (NLP), local label learning (LLL), joint label fusion (JLF), and metric learning based label fusion (ML).
Results: Our method obtained mean Dice index values of 0.889/0.886 for the left/right hippocampi respectively, while the mean Dice index values of alternative methods were 0.861/0.867 (NLP), 0.870/0.878 (LLL), 0.874/0.878 (JLF), and 0.875/0.880 (ML). Correlation coefficients between the hippocampal volumes obtained by our method and the manual segmentation method were 0.979/0.968 for the left/right hippocampi.
Conclusion: The experimental results have demonstrated that our deep learning method could achieve significantly better hippocampus segmentation performance than state-of-the-art multi-atlas based image segmentation methods.