Hippocampus segmentation from MR infant brain images is indispensable for studying early brain development. However, most of hippocampus segmentation methods were developed for adult brain images, which are not suitable for infant brain images of the first year due to low image contrast and variable structural patterns of early hippocampal development. To address these challenges, we propose a boundary regression method to detect hippocampal boundaries in the infant brain images, and then use the obtained boundaries to guide the deformable segmentation. The advantages of our segmentation method are: (1) different from the recently-developed atlas-based hippocampus segmentation methods, our method does not perform time-consuming deformable registrations; (2) different from the conventional point-regression-based boundary detection methods, our boundary regression method can predict the whole hippocampal boundary by a single regression model. Experiments on MR infant brain images from 2-week-old to 1-year-old show promising hippocampus segmentation results.