A Robust Algorithm for Characterizing Anisotropic Local Structures

摘要: This paper proposes a robust estimation and validation framework for characterizing local structures in positive multi-variate continuous function approximated by Gaussian-based model. The new solution is against data with large deviations from the model margin-truncations induced neighboring structures. To this goal, it unifies statistical parametric fitting multi-scale analysis based on scale-space theory. unification realized formally extending mean shift-based density towards signals whose structure characterized an anisotropic fully-parameterized covariance matrix. A method analyzing residual error of chi-square also proposed to complement framework. strength our aforementioned robustness. Experiments synthetic 1D 2D clearly demonstrate advantage comparison γ-normalized Laplacian approach [12] standard sample [13, p.179]. applied 3D volumetric lung tumors. implementation evaluated high-resolution CT images 14 patients 77 tumors, including 6 part-solid or ground-glass opacity nodules that are highly non-Gaussian clinically significant. Our system accurately estimated spread orientation 82% total tumors correctly rejected all failures without any false rejection acceptance. processes each 32-voxel volume-of-interest average two seconds 2.4GHz Intel CPU. generic can be blob-like various other applications.