New algorithm facilitates volumetric image compression

October 25, 2005

Researchers in Slovenia have proposed a new algorithm for the lossless compression of volumetric data sets, a finding that may allow CT and MR modalities to compress voxel data during acquisition.

Researchers in Slovenia have proposed a new algorithm for the lossless compression of volumetric data sets, a finding that may allow CT and MR modalities to compress voxel data during acquisition.

Unlike existing approaches based on octree encoding, the new method builds a quadtree representation of each slice (Comput Med Imaging Graph 2005;29(4):305-312).

The solution addresses the problematic issue of compression of volumetric data sets.

"Few algorithms exist that are strictly oriented toward the compression of volumetric data, and most of these are lossy, which is inappropriate in medical applications," said Gregor Klajnsek, Ph.D., of the Laboratory for Geometric Modeling and Multimedia Algorithms at the University of Maribor.

Most existing lossy algorithms use octree encoding for compression, a disadvantage in that they do not exploit the coherence between neighboring slices but instead create a common voxel space. This requires large onboard memory, Klajnsek said.

Instead, Klajnsek uses a different compression approach: processing two, consecutive, slices at a time.

"We try to determine coherent areas between the two slices," Klajnsek said.

Some properties of voxel data sets, such as homogeneity and coherence, can then be used more efficiently to produce better compression ratios.

"With this approach, only two slices have to be loaded into memory at a given time, implying the algorithm does not require a lot of working storage. This makes it far more attractive for possible hardware implementation," he said.

The quadtree method is similar to the interframe-based digital video compression method used by MPEG-2, in that it exploits similarities between neighboring slices. In contrast to MPEG-2, which forms a set of equally sized blocks of pixels (macro-blocks) and predicts where they would move in the next frame, Klajnsek's algorithm uses macro-blocks of different sizes and does not make any prediction. The positions and sizes of the macro-blocks are coded by a quadtree.

Klajnsek said this method enables progressive visualization, which is especially important when transferring volumetric data over the Internet.

"The exchange of volumetric data between radiologists and referring physicians remains limited due to lengthy transfer times," he said. "Any general-purpose compression techniques can be used to reduce the size of the data, but they do not support progressive visualization, which is of critical importance when quick insight is required."