A Keypoint-based Morphological Signature for Large-scale Neuroimage Analysis

We present an image keypoint-based morphological signature that can be used to efficiently assess the pair-wise whole-brain similarity for large MRI datasets. Similarity is assessed via Jaccard-like measure of set overlap based on the proportion of keypoints shared by an image pair, which may be evaluated in O(N log N) computational complexity given a set of N images using fast nearest neighbor indexing. Image retrieval experiments combine four large public neuroimage datasets including the Human Connectome Project (HCP), the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Open Access Series of Imaging Studies (OASIS), for a total of N=7536 T1-weighted MRIs of 3334 unique subjects. Our method identifies all pairs of same-subjects images based on a simple threshold, and revealed a number of previously unknown subject labeling errors.