Alzheimer’s disease (AD) is one of the most common brain dementia, which effects human memory, thinking and behavior. It has been proven that hippocampus is an important region related to AD diagnosis. Most of the existing methods on hippocampus analysis are based on the shape and volume analysis of the bilateral hippocampi. However, the 3D structural magnetic resonance images (MRI) can capture more useful information of hippocampus and its adjacent regions. In this paper, we propose a new method based on deep 3D convolutional neural network (3D CNN) for hippocampus analysis using 3D MR images for AD diagnosis. First, two hippocampi are segmented from other regions and the centers of hippocampus regions are calculated. Then, based on each hippocampus center, a local 3D image patch is extracted from the 3D MR image to cover each hippocampus region. Next, a deep 3D CNN model is constructed to extract the hierarchical imaging features for each hippocampus, followed by a softmax layer to generate a class prediction score for AD diagnosis. Finally, the classification is made by combination of the prediction scores from two hippocampi. Our method is evaluated using T1-weighted structural MR images on 231 subjects including 101 AD patients and 130 normal controls (NC) from Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Experimental results show the proposed method achieves an accuracy of 86.98% for classification of AD vs. NC, demonstrating the promising classification performance.