Interactive demo · Statistical Methods for Modern Biology · Chapter 11

Voronoi tessellation on an image manifold: segmenting cell bodies

Every pixel is given to the nucleus that is closest to it, where the length of a path is measured on a landscape whose height is the actin intensity. A step from a pixel to one of its eight neighbours costs ds = √((dz² + λ(dx² + dy²)) / (1 + λ)), the metric (11.4) of Holmes and Huber (2019) up to a constant factor: for large λ only lateral distance counts and the result is the ordinary Voronoi tessellation; for small λ climbing costs, and cell boundaries settle in the dark valleys between cells. Distances are computed for all pixels at once by Dijkstra's algorithm started from every nuclear pixel. Click a cell to see its area.

Trade-off λ (log scale)

Landscape z

Height increment dz

Region

Show

Scoreboard

Try these

The seeds are the 43 nuclei of Figure 11.12 and the landscape is the actin channel smoothed with a Gaussian of standard deviation 3, both computed by the R code. With the 3 × 3 window increment (the sum over the window of |z(p + d) − z(q + d)|) the page reproduces EBImage::propagate; the single-pixel increment |z(q) − z(p)| is the literal reading of (11.4). Each pixel takes the label of the neighbour on its shortest path.

In the lecture notes