[Paper Review] Automated scalable segmentation of neurons from multispectral images

Uygar Sümbül, Douglas H. Roossien+6 more|arXiv (Cornell University)|Nov 1, 2016

Cell Image Analysis TechniquesBiochemistry, Genetics and Molecular Biology27 references18 citations

TL;DR

This paper presents a scalable, automated method for segmenting neurons in multispectral Brainbow images by leveraging denoising, supervoxelization, and graph-based segmentation to reduce computational load by four orders of magnitude. The approach achieves accurate, connectivity-preserving segmentation of complex neuronal arbors despite color inconsistencies and noise, validated on simulated and real mouse hippocampal datasets with high segmentation accuracy and robustness to noise and channel count.

ABSTRACT

Reconstruction of neuroanatomy is a fundamental problem in neuroscience. Stochastic expression of colors in individual cells is a promising tool, although its use in the nervous system has been limited due to various sources of variability in expression. Moreover, the intermingled anatomy of neuronal trees is challenging for existing segmentation algorithms. Here, we propose a method to automate the segmentation of neurons in such (potentially pseudo-colored) images. The method uses spatio-color relations between the voxels, generates supervoxels to reduce the problem size by four orders of magnitude before the final segmentation, and is parallelizable over the supervoxels. To quantify performance and gain insight, we generate simulated images, where the noise level and characteristics, the density of expression, and the number of fluorophore types are variable. We also present segmentations of real Brainbow images of the mouse hippocampus, which reveal many of the dendritic segments.

Motivation & Objective

• To address the challenge of automated, scalable segmentation of neurons in multispectral images with high noise and variable color expression.
• To overcome limitations of existing methods that fail on complex, interwoven neuronal arbors and noisy, pseudo-colored images.
• To develop a computationally efficient pipeline that reduces voxel-level problems by four orders of magnitude through supervoxelization.
• To enable accurate segmentation by modeling spatio-color relationships in a graph-based framework that preserves neuronal connectivity.
• To evaluate performance under varying noise levels, fluorophore counts, and expression densities using simulated and real datasets.

Proposed method

• Applying BM4D collaborative filtering to denoise individual color channels in 3D image stacks, assuming Gaussian noise.
• Generating supervoxels via superpixel-like segmentation to reduce the number of units from voxels to supervoxels, decreasing problem size by four orders of magnitude.
• Constructing a graph where nodes are supervoxels and edge weights encode spatial and color similarity, penalizing discontinuities and color inhomogeneities.
• Using spectral clustering on the graph to segment neurons, with parameters tuned for connectivity and color consistency.
• Employing a simulation framework based on real neuron reconstructions to generate synthetic images with tunable noise, expression density, and fluorophore types.
• Applying the same processing pipeline—denoising, supervoxelization, merging, and clustering—across both simulated and real datasets with consistent parameters.

Experimental results

Research questions

• RQ1 How does the inclusion of additional color channels affect segmentation accuracy in multispectral neuron imaging?
• RQ2 To what extent does noise in color expression degrade segmentation performance, and how can it be mitigated?
• RQ3 Can a supervoxel-based approach preserve neuronal connectivity while reducing computational complexity by four orders of magnitude?
• RQ4 How does the accuracy of segmentation vary with over- or under-estimation of the number of neurons?
• RQ5 Can the proposed method reliably segment complex, interwoven neuronal arbors in real Brainbow images, including expanded microscopy datasets?

Key findings

• Segmentation accuracy improves significantly with the addition of more color channels, aligning with ongoing experimental efforts to increase spectral diversity.
• Performance degrades gradually with increasing neuron-color noise (σ₁), but remains robust under mild overestimation of neuron count, with rapid decay when the count is underestimated.
• The method successfully segments dendritic processes in real mouse hippocampal Brainbow images, preserving connectivity and enabling downstream tracing.
• The 4× expanded microscopy dataset shows more fragmentation, likely due to fewer color channels, but most neurites remain identifiable and segmentable.
• Supervoxelization reduces the problem size by four orders of magnitude, enabling scalable processing of large 3D image stacks.
• The approach correctly clusters most supervoxels into connected neuronal components, with only minor errors from aggressive merging or spurious clusters.