Neighborhood-Preserving Dimensionality Reduction for Multivariate Volume Rendering
R. Snellenberg (TU Delft - Electrical Engineering, Mathematics and Computer Science)
T. Höllt – Mentor (TU Delft - Computer Graphics and Visualisation)
C.A. Raman – Graduation committee member (TU Delft - Pattern Recognition and Bioinformatics)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Multivariate volumetric datasets, which are volumetric datasets that contain more than one variable per voxel, are becoming increasingly prevalent, such as transcriptomic datasets, which can contain up to hundreds of dimensions. This high dimensionality has proven to be a difficult obstacle to overcome when trying to visualize such datasets, especially in creating user-friendly transfer functions, which are needed to guide the direct volume rendering pipeline into visualizing information relevant to the user.
This work proposes the use of neighborhood-preserving dimensionality reduction, such as t-SNE and UMAP, to define a transfer function that simplifies highlighting highly similar points in data with an arbitrarily high number of dimensions.
However, since direct interpolation between points in the resulting non-linear embedded space does not provide correct results, we propose and evaluate several interpolation methods to address this limitation. The most accurate of which is the use of approximate nearest neighbor algorithms to approximate the position an interpolated high-dimensional point would have in the embedded space. We also propose an adaptation of a two-step TF that can simulate some of the surface-based effects whose scalar-based implementations cannot be directly applied to multivariate data, such as gradient-based opacity and surface shading.
Finally, we implement a plugin in the ManiVault framework, which is an existing framework made to analyze multivariate data, to test and analyze our method.