tooluniverse-image-analysis

Use scikit-image to threshold the image with Otsu's method, remove small objects, then label connected components. For touching cells, apply a distance transform with watershed segmentation to separate merged nuclei before counting.

Segment objects from one channel (e.g., DAPI) to create a labeled mask, then use skimage measure.regionprops_table with each channel as the intensity image. Compute mean, max, and integrated intensity per object per channel, and derive channel ratios.

Otsu works best for high-contrast images with clear bimodal histograms, while Li's method is more sensitive for low-contrast or faint objects. For uneven illumination, use adaptive local thresholding instead of either global method.

Yes, scipy.stats.dunnett (scipy >= 1.10) replaces R's multcomp::glht for comparing multiple groups against a control. For natural splines, use patsy.cr with explicit quantile knots instead of R's ns().

Over-segmentation happens when min_distance between peaks is too small or smoothing is insufficient; increase min_distance or Gaussian sigma. Under-segmentation (merged cells) results from min_distance being too large or weak distance-transform peaks.

This workflow is not suited for phylogenetics, RNA-seq differential expression, or single-cell scRNA-seq analysis. For very dense or irregularly shaped cells, deep-learning tools like CellPose or StarDist may outperform threshold-based segmentation.