Random-Accessible Volume Data Compression with Regression Function

Typical volumetric data sets are 3D regular grids captured from computerized tomography (CT) or magnetic resonance (MR) scanners. There has been a constant tension between volume data resolution and memory usage. In this paper, we introduce an effective 3D compression scheme for volume data, which exploits the power of regression function. A regression function represents a non-linear mapping from positions of voxels to scalar values, and is modeled as a multi-layer feed-forward neural network. Our approach is feature preserving and resolution efficient, and supports real-time random access at decoding stage, via a simple GPU pixel shader. We applied our technique on multiple medical and scientific data sets, demonstrating that it is well suited for encoding and decoding volumetric data including quality and compression ratios.