Hierarchical Dimensionality Reduction for Scalable Multivariate Volume Rendering

Abstract

Multivariate volumetric datasets are becoming increasingly large and complex, making transfer function design for direct volume rendering more difficult. Recent work has shown that flat dimensionality reduction techniques, such as t-SNE, can support transfer function design by projecting high-dimensional voxel attributes into a two-dimensional embedding space. However, flat dimensionality reduction methods become difficult to scale to datasets containing millions of voxels. They produce visually cluttered transfer function domains and require large nearest-neighbor structures for mapping rendering samples to the embedding. In this work, we use Hierarchical Stochastic Neighbor Embedding (HSNE) as a scalable alternative for dimensionality reduction-based transfer function design in multivariate volume rendering. Instead of defining the transfer function over all voxels, we select a level of the HSNE hierarchy and use its landmarks as a reduced domain, and integrate this representation into the rendering pipeline. Our method is implemented and evaluated in the ManiVault framework using large multivariate tissue datasets. The results show that the HSNE-based approach significantly reduces preprocessing and rendering times compared to a t-SNE-based baseline, while also reducing visual clutter in the transfer function space. Higher hierarchy levels further improve runtime performance and simplify interaction, although they may lose fine detail. These results demonstrate that hierarchical dimensionality reduction can improve the scalability and usability of dimensionality reduction-based transfer function design for large multivariate volumetric datasets.