Analyzing Brain Morphology on the Bag-of-Features Manifold

• Published in: Information Processing in Medical Imaging Vol. 11492; pp. 45 - 56
• Main Authors: Chauvin, Laurent, Kumar, Kuldeep, Desrosiers, Christian, De Guise, Jacques, Wells, William, Toews, Matthew
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• ISBN: 3030203506; 9783030203504
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-20351-1_4

We propose a novel distance measure between variable-sized sets of image features, i.e. the bag-of-features image representation, for quantifying brain morphology similarity based on local neuroanatomical structures. Our measure generalizes the Jaccard distance metric to account for probabilistic or soft set equivalence (SSE), via a novel adaptive kernel density framework accounting for probabilistic uncertainty in both feature appearance and geometry. The method is based on highly efficient keypoint feature indexing and is suitable for identifying pairwise relationships in arbitrarily large data sets. Experiments use the Human Connectome Project (HCP) dataset consisting of 1010 subjects, including pairs of siblings and twins, where neuroanatomy is modeled as a set of scale-invariant keypoints extracted from T1-weighted MRI data. The Jaccard distance based on (SSE) is shown to outperform standard hard set equivalence (HSE) in predicting the immediate family graph structure and genetic links such as racial origin and sex from MRI data, providing a useful tool for data-driven, high-throughput genome wide heritability analysis.