Using Atom-Like Local Image Features to Study Human Genetics and Neuroanatomy in Large Sets of 3D Medical Image Volumes
Authors:
Laurent Chauvin
Abstract:
The contributions of this thesis stem from technology developed to analyse large sets of volumetric images in terms of atom-like features extracted in 3D image space, following SIFT algorithm in 2D image space. New feature properties are introduced including a binary feature sign, analogous to an electrical charge, and a discrete set of symmetric feature orientation states in 3D space. These new p…
▽ More
The contributions of this thesis stem from technology developed to analyse large sets of volumetric images in terms of atom-like features extracted in 3D image space, following SIFT algorithm in 2D image space. New feature properties are introduced including a binary feature sign, analogous to an electrical charge, and a discrete set of symmetric feature orientation states in 3D space. These new properties are leveraged to extend feature invariance to include the sign inversion and parity (SP) transform, analogous to the charge conjugation and parity (CP) transform between a particle and its antiparticle in quantum mechanics, thereby accounting for local intensity contrast inversion between imaging modalities and axis reflections due to shape symmetry. A novel exponential kernel is proposed to quantify the similarity of a pair of features extracted in different images from their properties including location, scale, orientation, sign and appearance. A novel measure entitled the soft Jaccard is proposed to quantify the similarity of a pair of feature sets based on their overlap or intersection-over-union, where a kernel establishes non-binary or soft equivalence between a pair of feature elements. The soft Jaccard may be used to identify pairs of feature sets extracted from the same individuals or families with high accuracy, and a simple distance threshold led to the surprising discovery of previously unknown individual and family labeling errors in major public neuroimage datasets. A new algorithm is proposed to register or spatially align a pair of feature sets, entitled SIFT Coherent Point Drift (SIFT-CPD), by identifying a transform that maximizes the soft Jaccard between a fixed feature set and a transformed set. SIFT-CPD achieves faster and more accurate registration than the original CPD algorithm based on feature location information alone, in a variety of challenging.
△ Less
Submitted 25 August, 2022;
originally announced August 2022.
Efficient Pairwise Neuroimage Analysis using the Soft Jaccard Index and 3D Keypoint Sets
Authors:
Laurent Chauvin,
Kuldeep Kumar,
Christian Desrosiers,
William Wells III,
Matthew Toews
Abstract:
We propose a novel pairwise distance measure between image keypoint sets, for the purpose of large-scale medical image indexing. Our measure generalizes the Jaccard index to account for soft set equivalence (SSE) between keypoint elements, via an adaptive kernel framework modeling uncertainty in keypoint appearance and geometry. A new kernel is proposed to quantify the variability of keypoint geom…
▽ More
We propose a novel pairwise distance measure between image keypoint sets, for the purpose of large-scale medical image indexing. Our measure generalizes the Jaccard index to account for soft set equivalence (SSE) between keypoint elements, via an adaptive kernel framework modeling uncertainty in keypoint appearance and geometry. A new kernel is proposed to quantify the variability of keypoint geometry in location and scale. Our distance measure may be estimated between $O(N^2)$ image pairs in $O(N~\log~N)$ operations via keypoint indexing. Experiments report the first results for the task of predicting family relationships from medical images, using 1010 T1-weighted MRI brain volumes of 434 families including monozygotic and dizygotic twins, siblings and half-siblings sharing 100%-25% of their polymorphic genes. Soft set equivalence and the keypoint geometry kernel improve upon standard hard set equivalence (HSE) and appearance kernels alone in predicting family relationships. Monozygotic twin identification is near 100%, and three subjects with uncertain genoty** are automatically paired with their self-reported families, the first reported practical application of image-based family identification. Our distance measure can also be used to predict group categories, sex is predicted with an AUC=0.97. Software is provided for efficient fine-grained curation of large, generic image datasets.
△ Less
Submitted 16 September, 2021; v1 submitted 11 March, 2021;
originally announced March 2021.
