Try using octrees to improve performance

The volumes could be compressed using some sort of Octree data format (or trees in general) which should result in significantly less memory lookups (due to much more caching) and therefore probably improve the performance.
Octrees can be generated when loading the data the first time while parameters should be initialized dynamically depending on the histogram of the volume.
The raymarcher has to be modified to use a dynamic step size that it has to set each step when moving through the octree.

Start by locating the volume-loading path and raymarcher entry point mentioned in the issue, then inspect how volume data, histograms, and ray steps are currently represented. Done means an octree-based representation is generated during initial loading, parameters are initialized from the volume histogram, and raymarching uses a per-step dynamic size with performance evidence.

Written by the indexing model from the issue text.