Grayscale-weighted distance transform

[0x00b1]

From @LeeKamentsky on October 31, 2014 15:46

From @dlogan :
The following paper describes a grayscale-weighted distance transform that helps highlight neurites.

Xiao, Hang, and Hanchuan Peng. “APP2: Automatic Tracing of 3D Neuron Morphology Based on Hierarchical Pruning of a Gray-Weighted Image Distance-Tree.” Bioinformatics 29, no. 11 (June 1, 2013): 1448–54. doi:10.1093/bioinformatics/btt170.

The algorithm:

• compute a conservative background (the pixels that are surely background). The paper suggests the mean intensity value - it could be a quartile too, e.g. 25%.
• compute the distance transform
• the pixels adjacent to background pixels retain their value
• other foreground pixels have a transformed value of their value plus the transformed value of the adjacent pixel closest to the background (i.e. the sum of all pixels along the shortest path to the background). I would do the average of all adjacent pixels at the same distance.

Not so hard to do - maybe 1/2 day including a test. I can see how it would give high weights to the spine of a neurite, even if it were a little foggy. The background method introduces some instability and parameterization - you probably need the full thresholding settings to give the user enough flexibility.

Copied from original issue: CellProfiler/CellProfiler#1249