Interactive Visualization of High-Dimensional Voxel Features in Volumetric Path Tracing using Dimensionality Reduction and HDBSCAN

Abstract

This paper introduces an interactive visualization framework for exploring complex volumetric microscopy data, using multi-dimensional feature extraction and dimensionality reduction to improve structural analysis. Each voxel is assigned an 11-dimensional feature vector to capture both local and global context. After uniform sampling, we apply UMAP and t-SNE for dimensionality reduction, followed by HDBSCAN for density-based clustering. The resulting clusters are mapped back to the original volume to construct an interactive transfer function, with Gaussian blobs representing cluster centers. Expert evaluation of troponin-labeled cardiomyocytes shows that this analysis offers greater sensitivity in detecting treatment-induced structural remodeling than visual inspection of 2D slices. Non-treated samples display higher morphological heterogeneity, with multiple small regions of varying intensity, while treatment with 500 nM carfilzomib for 24 hours leads to fewer, more dominant phenotypic states characterized by localized protein clustering and nuclear enlargement. Our results demonstrate that this unbiased, quantifiable workflow successfully reveals treatment-associated structural changes and protein redistributions that remain undetected in conventional 2D immunofluorescence evaluation.