3D neuron soma segmentation for light-sheet microscopy

Problem

Deep networks can segment neuronal somas in 3D brain images, but they need voxel-level ground truth: slow to label, and rarely released, especially for light-sheet (LSFM) images of cleared rat brains. Our capstone team built MIND, an open-source plugin for the Vaa3D image-analysis platform, to label somas quickly, analyse their shape and simulate extra training data.

Constraints

• Terabyte-scale images. The plugin works with Vaa3D’s TeraFly viewer, which streams whole-brain volumes.
• Anisotropic voxels. 1.8 × 1.8 × 4.0 µm at 3.6× magnification and 0.71 × 0.71 × 1.0 µm at 9×.
• Isolated somas only. A soma must fit in a sphere around its center without overlapping its neighbours. Dense regions are out of scope.
• No labelled LSFM data to score against during the capstone, so its evaluation was qualitative.

What I built

My part was the segmentation module:

• Three thresholding modes: iterative thresholding from the reference plugin, global Otsu, and local Otsu computed inside each marker’s sphere, which is the default.
• 3D region growing from each marker’s center, constrained to the marker’s sphere by default (the marker constraint), with an option to turn it off.
• A histogram-based 3D median filter (default window 3) that removes salt-and-pepper noise before segmentation without blurring soma edges.
• Per-soma override. Each soma’s threshold is saved in its marker; editing it and re-running re-segments only that soma.

1. One marker per somacenter and radius
2. 3D median filterremoves salt-and-pepper noise
3. Local Otsu thresholdcomputed inside the marker’s sphere
4. 3D region growingfrom the center, constrained to the sphere
5. Binary 3D maskthreshold saved in the marker for per-soma override

Since the capstone, I built a Dice-based evaluation against hand-labelled gold standards. The rest of the platform (PCA shape analysis, the probabilistic shape model and the simulation pipeline) was built by teammates.

Results

• Local Otsu followed the visible soma boundaries in qualitative inspection.
• Global Otsu under-flooded dim somas and over-flooded bright ones. Iterative thresholding under- or over-flooded depending on its parameters and the soma’s shape.
• The platform built a probabilistic shape model from 10 segmented somas and generated synthetic image–label pairs from 10 labelled somas.

What went wrong and how I fixed it

• Over-flooding that a heuristic can’t see. Iterative thresholding stops when the segmentation’s center of mass drifts. A soma that floods evenly in every direction keeps its center of mass, so the error goes undetected, and odd-shaped somas trip the rule too early. I made local Otsu inside each marker’s sphere the default and constrained growth to that sphere.
• One threshold can’t fit every soma. A global threshold is too generic for somas of different brightness. Per-marker thresholds, with a manual override, handle that.
• No numbers in the capstone. Without labelled LSFM data, the capstone could only inspect results visually. I followed up with the Dice-based evaluation against hand-labelled gold standards.

What I’d do next

• Extend segmentation to dense regions, where somas can’t be isolated in a sphere.
• Train deep networks with and without the simulated data, and measure the difference with the Dice evaluation.

Stack

• C++
• Vaa3D
• TeraFly

Links

Related