Skip to content

Segmentation Methods

SPIMquant provides two built-in segmentation methods for detecting pathology signals in SPIM data, plus support for intensity correction pre-processing. The method and correction are configured independently per stain via the snakebids.yml config file.

Overview

flowchart LR
    A[Raw SPIM] --> B[Intensity Correction]
    B --> C{seg_method}
    C -->|threshold| D[Threshold]
    C -->|otsu+k3i2| E[Multi-Otsu + k-means]
    D --> F[Binary Mask]
    E --> F
    F --> G[Clean: remove edge / small objects]
    G --> H[Segmentation Output]

After segmentation the binary mask is used to compute field fraction, object count, and per-object region properties.

Intensity Correction

Bias-field correction is applied before segmentation to compensate for the spatially varying illumination typical of lightsheet microscopy.

Gaussian (correction_method: gaussian)

A Gaussian blur is applied to the raw image at a coarse downsampled level to estimate the low-frequency illumination profile, which is then divided out. This is fast and requires no extra memory, making it suitable for quick iteration or large datasets.

Config key: correction_method: gaussian

N4 (correction_method: n4)

ANTs N4BiasFieldCorrection is applied. N4 fits a smooth B-spline model of the bias field and is more accurate than Gaussian correction, especially for thicker sections or strong illumination gradients.

Config key: correction_method: n4

Tip

Use n4 when high quantitative accuracy is required. Use gaussian for faster exploratory runs or when the illumination gradient is mild.

Segmentation Methods

Threshold (seg_method: threshold)

A fixed intensity threshold is applied to the (corrected) image. Voxels above the threshold are classified as positive; all others are negative.

The threshold value is configured per-stain:

stain_defaults:
  abeta:
    seg_method: threshold
    seg_threshold: 500   # intensity value; adjust to your data

When to use: Works well when the pathology signal is clearly brighter than background and the intensity scale is stable across subjects. Simple to interpret and debug.

Limitations: Sensitive to residual intensity non-uniformities and to between-subject intensity variation. May require manual threshold tuning per dataset.

Multi-Otsu (seg_method: otsu+k3i2)

An unsupervised thresholding method that adapts to the intensity distribution of each image.

The method string encodes two parameters: k — the number of Otsu classes, and i — the threshold index to use for the binary classification. For otsu+k3i2:

  1. Multi-Otsu thresholding — Otsu's method is extended to find k-1 thresholds that minimise within-class variance, splitting the histogram into k classes. With k=3 the image is divided into background, low-signal, and high-signal classes (2 thresholds).
  2. Binary classification — the threshold at index i (1-based) is applied to produce the final binary mask. With i=2 this selects the second (higher) threshold, classifying only the brightest voxels as positive.

Config key:

stain_defaults:
  abeta:
    seg_method: otsu+k3i2

When to use: Preferred when the staining intensity varies across subjects or imaging sessions, because the threshold adapts automatically to each image. Also more robust to residual illumination gradients.

Limitations: Can fail on images with unusual histograms (e.g. very sparse pathology that does not form a distinct peak) or when the background is very noisy.

Post-Segmentation Cleaning

After the initial binary mask is produced, a cleaning step removes two classes of artefact:

  1. Edge objects — connected components that touch the border of the field of view are removed, as these are typically cut-off tissue or slide artefacts rather than true pathology.
  2. Small objects — objects below a minimum volume threshold (configured via regionprop_filters) are discarded, eliminating small noise specks.

Scale convention

The output mask is stored on a 0–100 scale (not 0–1). This is deliberate: when the mask is spatially downsampled to compute field fraction, the resulting values are directly interpretable as a percentage (0–100 %).

Per-Stain Configuration

Each stain is configured independently. A typical snakebids.yml block looks like:

stain_defaults:
  abeta:
    seg_method: otsu+k3i2
    correction_method: n4
    regionprop_filters:
      min_area: 50       # voxels; objects smaller than this are discarded
  Iba1:
    seg_method: threshold
    seg_threshold: 300
    correction_method: gaussian

Stains not listed in stain_defaults fall back to the top-level seg_method and correction_method keys.

Further Reading