Skip to content

How LightSuite Works

LightSuite maps central-nervous-system samples into a standard atlas and detects and quantifies labelled cells. Whether you start from a whole-brain lightsheet volume, a spinal cord, or a series of wide-field sections, the backbone is the same:

downsample → register to the atlas → (optionally) detect cells → export everything in atlas space.

This page explains the shared ideas once. The workflow pages give the click-by-click steps for each modality.

The atlas and coordinate space

  • Brains are registered to the Allen Mouse Brain Common Coordinate Framework (CCF v3, 2020 parcellation; Wang et al., 2020). Spinal cords use the Fiederling et al. (2021) atlas.
  • All registration runs on volumes downsampled to 20 µm isotropic — fast to handle, and still finer than the >500 µm spatial scale of the non-rigid fit. The atlas template and annotations are kept at 10 µm.
  • Results are reported in atlas voxel coordinates. The exact column order depends on the module and is stated with each output (for example, slice-module cell coordinates are [ML, AP, DV], whereas the probe probe_ccf uses [AP, DV, ML]).

Registration: from sample to atlas

B-spline deformable registration is non-convex, so a good initialization is critical. Instead of relying on image similarity alone — which can fail on damaged, distorted, or autofluorescent tissue — LightSuite initializes from sample-derived 3D point clouds and then refines in three stages:

  1. Point-cloud initialization. High-saliency structural points are extracted locally (so uneven illumination does not bias them) and matched to a point cloud from the atlas with Bayesian Coherent Point Drift. A similarity transform (rotation, uniform scale, translation) is fit first — scale matters because clearing can shrink or swell tissue.
  2. Affine. A global affine transform is fit from matched points (automatic correspondences plus any you add).
  3. B-spline (Elastix). A smooth, local deformation is computed with a multi-metric objective: image mutual information plus a corresponding-points term that pulls your landmarks onto their matches. Both the forward (atlas→sample) and inverse (sample→atlas) transforms are saved; the inverse is what maps detected cells into atlas space.

Each modality adapts this backbone: the spinal cord is straightened first, and the slice module registers each 2D section before stacking them into a volume.

Refining registration with control points (active learning)

A shared GUI lets you add corresponding landmarks between your sample and the atlas (the brain and spinal-cord modules use the same 3D interface; the slice module has a 2D version). You click matching points on the sample and atlas side by side, and the atlas region boundaries are overlaid on your sample.

The overlay is the feedback loop. Whenever the sample and atlas landmark counts match, a 3D affine is re-fit live from all pairs so far and the overlay is redrawn — so every new point is placed against a progressively better alignment. Candidate views are sampled evenly along all three axes in randomized order (and expose one quadrant at a time) to encourage uniform coverage rather than clustering.

Tips that apply everywhere:

  • Place at least 16 pairs spread across the full extent of the sample; the fit only activates at 16, and 100+ helps on heavily deformed tissue.
  • Watch the MSE in the banner — lower means a tighter fit.
  • Good landmarks are ventricle boundaries, major fiber tracts, and distinct nuclei outlines.
  • In damaged or missing regions, place a point where the structure should be. This stops the deformation from bending the atlas into the void.

Cell detection and artifact classification

Detection borrows from spike-sorting: extract candidates everywhere, then triage them.

  1. Bandpass filter. A Butterworth filter centered on the expected cell size highlights cell-sized objects and suppresses noise and slow background.
  2. Signal-to-background ratio (SBR). Subtracting the bandpass result from the original gives a local background; dividing by it converts raw intensity into an SBR. SBR is robust to the threefold intensity differences seen across brain areas and to one-sided lightsheet illumination.
  3. Local maxima. Maxima above the primary SBR threshold become candidate centers, grown out to a secondary threshold.
  4. Morphological filtering. Candidates that are too elongated (axis ratio > 2.5), too small, or too large/bright (tissue folds, bubbles) are discarded.
  5. CNN artifact classifier. A lightweight network classifies each candidate from three maximum-intensity projections (cheaper than a 3D network) and removes false positives such as bubble edges, tissue–solution interfaces, and bright neurites. It reaches >98% accuracy and can be retrained on your own labels.

The whole-brain pipeline runs this in 3D; the slice module uses a 2D adaptation. celldiam sets the filter scale (the single most important parameter) and the threshold pair sets sensitivity.

What LightSuite produces

  • Registration transforms (transform_params.mat) — forward and inverse, reused by every downstream step including implant tracing.
  • Registered volumes for every channel, warped into atlas space.
  • Cell positions in atlas space, plus regional cell counts and fluorescence intensities per area and hemisphere (exportable to CSV).
  • Implant localization for optical fibers / GRIN lenses and Neuropixels probes — see Probe & implant tracing.