P3-417: ACCURATE HIPPOCAMPUS SEGMENTATION USING FULLY CONVOLUTIONAL NETWORKS WITH RESIDUAL CONNECTIONS AND FUSED DEEP SUPERVISION

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 Alzheimer's disease. Although multi-atlas image segmentation and deep learning 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, efficient, and reproducible segmentation of the hippocampus, we develop a novel deep learning hippocampus segmentation method. Our deep learning based hippocampus segmentation method is built upon fully convolutional networks (FCNs), short/long-range residual connections, and fused deep supervision (Fig. 1). Particularly, 3D residual blocks with long-range connections across different scales are used to extract informative features for segmentation. Fused deep supervision from coarse to fine scales are adopted to guide the feature extraction, network training, and integration of segmentation results at multiple scales. Our deep learning 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 from EADC-ADNI. We also investigate the impact of different field strengths to the segmentation using 1.5T and 3T scans of 112 subjects from ADNI. Our method obtained mean Dice index values of 0.888/0.889 for the left/right hippocampi respectively (Fig. 2, top), while the mean Dice index values of an alternative deep learning segmentation method without fused deep supervision were 0.868/0.870 (Fig. 3). Correlation coefficients between the hippocampal volumes obtained by our method and manual segmentation were 0.970/0.955 for the left/right hippocampi respectively (Fig. 2, top). Our method also obtained highly reproducible segmentation results on 1.5T and 3T scans of the same subjects, with mean Dice index values of 0.930/0.929 and correlation coefficients of 0.985/0.976 between the volumes obtained on 1.5T and 3T scans respectively for the left/right hippocampi (Fig. 2, bottom).