[Paper Review] Gradient-Domain Processing for Large EM Image Stacks

Michael Kazhdan, Randal Burns+4 more|arXiv (Cornell University)|Sep 30, 2013

Advanced Optical Imaging Technologies18 references3 citations

TL;DR

This paper proposes a two-stage gradient-domain method for removing inter-slice discontinuities in large-scale electron microscopy (EM) image stacks while preserving intra-slice detail. It uses anisotropic diffusion along the slice axis followed by screened-Poisson reconstruction per slice, accelerated by a novel out-of-core, multigrid solver that processes tera-voxel datasets in under 60 hours on 12 cores with 45GB RAM.

ABSTRACT

We propose a new gradient-domain technique for processing registered EM image stacks to remove the inter-image discontinuities while preserving intra-image detail. To this end, we process the image stack by first performing anisotropic diffusion to smooth the data along the slice axis and then solving a screened-Poisson equation within each slice to re-introduce the detail. The final image stack is both continuous across the slice axis (facilitating the tracking of information between slices) and maintains sharp details within each slice (supporting automatic feature detection). To support this editing, we describe the implementation of the first multigrid solver designed for efficient gradient domain processing of large, out-of-core, voxel grids.

Motivation & Objective

• To address inter-slice discontinuities in large EM image stacks caused by variations in lighting, camera parameters, and sectioning artifacts.
• To preserve high-frequency anatomical details within individual slices during processing.
• To enable efficient, scalable gradient-domain processing of tera-voxel, out-of-core 3D volumetric data.
• To develop a multigrid solver tailored for large-scale, out-of-core voxel grids to support gradient-domain techniques.
• To extend 2D gradient-domain image processing to 3D volumetric data for applications in connectomics and neuroscience.

Proposed method

• The method first applies anisotropic diffusion along the z-axis (slice axis) to smooth inter-slice discontinuities while preserving intra-slice gradients in x and y directions.
• It formulates the anisotropic diffusion as a minimization of the weighted L2 norm of the difference between the gradient of the output and a target gradient field with reduced z-component weight.
• A screened-Poisson equation is then solved independently within each slice to re-introduce high-frequency details from the original data while maintaining continuity across slices.
• The solution uses a weighted gradient field where the z-component is suppressed, and the screened-Poisson equation balances fidelity to the smoothed data and the original gradients.
• A new parallel, out-of-core multigrid solver is implemented to efficiently solve the large linear systems arising from the screened-Poisson equations on tera-voxel datasets.
• The solver supports both in-memory and disk-based computation, with adaptive V-cycle strategies and memory-efficient data layout for large-scale voxel grids.

Experimental results

Research questions

• RQ1 Can gradient-domain processing effectively remove inter-slice discontinuities in large EM image stacks without blurring intra-slice anatomical details?
• RQ2 How can anisotropic diffusion be applied along the slice axis to suppress cross-slice artifacts while preserving in-plane gradients?
• RQ3 What is the performance and scalability of a multigrid solver for out-of-core, large-scale 3D voxel grids in the context of gradient-domain processing?
• RQ4 Can the combination of anisotropic diffusion and screened-Poisson reconstruction achieve both inter-slice continuity and intra-slice sharpness in EM data?
• RQ5 How does the choice of solver (constant, hybrid, linear) affect convergence and runtime in processing large 3D volumes?

Key findings

• The proposed two-stage method successfully removes inter-slice discontinuities while preserving fine anatomical details, as demonstrated in visual comparisons of original, diffused, and reconstructed data.
• The anisotropic diffusion step reduces residual norms to approximately 4.7×10⁻⁴ for both AD and SP phases, with the hybrid solver achieving the best balance of speed and convergence.
• The multigrid solver completed processing of a 1850×21496×25792 EM stack in under 60 hours using 12 cores and 45GB of RAM, demonstrating scalability for tera-voxel datasets.
• The hybrid multigrid solver achieved a residual decay rate of 0.53 per 10⁵ seconds, outperforming the constant and linear variants in terms of efficiency when accounting for runtime.
• The method was successfully extended to non-linear filtering of a 1920×1080×742 video, reducing noise by suppressing small gradients, with a peak memory footprint of only 185 MB per channel.
• The implementation supports general gradient-domain filtering, including HDR compression, sharpening, and non-photorealistic rendering, on out-of-core 3D data.