Friday 12 June 2009

Quantifying in Scientific Images

I regularly teach histologists and biologists how to extract quantitative 3D information from 2D images and the issue of image dequantification is a key part of the program.

There are a number of linked issues;

(1) Often in experimentally derived images the x and y magnifications can be different due to non square CCD pixels, optical distortions etc. This can be overcome by taking images of known gratcules in both directions to check magnification.

(2) Using calculated "magnifications" derived from the nominal magnification on the side of the objective is often not accurate.

(3) Most histological images are 2D projections of a 3D space (admittedly thin) that has been subjected to considerable tissue processing, shrinkage etc since it was a live piece of tissue. This should be experimentally investigated and accounted for by the scientists. Note that relatively modest linear shrinkage becomes significant 3D volumetric shrinkage. Furthermore not all tissue shrinks the same amount. Famously Herbert Haug a German anatomist found in the 1960's that the brain tissue of young humans shrank more than the brain tissue of old human brains. This led to an apparent "loss" of neurons with age (they had a lower numerical densitry but after correcting for differential shrinkage the same total number).

(4) The act of taking a thin histological (or optical) section leads to an obersved reduction in feature dimension. This is not widely reflected upon and there are no accepted ways to indicate this in standard hsitological images.

(5) Often the histology stains used do not stain tissue compartments uniformly.

(6) Histological images are in fact 2D samples (real Flatland stuff) from 3 space and in common with all statistical sampling they suffer from the "Central Paradox of Sampling" = simply by looking at the picture you have no idea if it has been randomly sampled or carefully selected (i.e. it is an unbiased or a biased sample).

(7) The short advice I give for all quantitative image analysis tasks is always is to think Outside-In not Inside-Out. i.e. Think about what was done to obtain the image you have in front of you and whether this is appropriate for the scientific task at hand NOT get obsessed by the image details some of which are misleading.

Matt