NeuroPack: Advanced Learnable Packaging Toolkit Towards Ultra-High Compression Ratio for Volumetric Neuroimaging

NeuroPack, a hyperprior-driven 3D deep-learning packaging toolkit for mesoscale microscopy volumes, achieves 100- to 1000-fold compression while delivering average relative gains of 73.4% (BD-PSNR) and 50.8% (BD-MSSSIM) over the best competing codec.

• Why it matters: Terabyte to petabyte 3D microscopic brain-volume datasets create major storage bottlenecks, and existing learned codecs can suffer decoding drift across heterogeneous hardware and microscopy sources.
• What they did: NeuroPack uses a compact hyperprior stream conditioned on local voxel-wise variance for rate–distortion optimization, and integerizes the hyperprior side-information pathway by replacing floating-point hyper-decoding with deterministic fixed-point integer convolutions.
• The result: The integerized design yields bit-exact entropy models across CPUs, GPUs, workstations, and cloud servers, preserving rate–distortion behavior and maintaining microscopy fidelity validated on confocal whole-brain Danio rerio images with human annotation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.