Message Boards Message Boards

Analyze videos of moving color circles (cells) to extract rules/patterns?

GROUPS:

I have videos of cells moving around a dish (2D). The cells are in different colors, representing various physiological states, and the colors change with time. What I need is some sort of unbiased machine learning approach, to look for patterns and relationships between color and movement. For example, do cells of a specific color tend to touch each other more than different colors, or do specific kinds of contact lead to a color change shortly thereafter? Basically, to extract any possible relationships between movement, relative position, and color. What is an appropriate technique for that, and are there any existing packages that do this (or people interested in writing us some analysis code as a contractor project)?

POSTED BY: Michael Levin
Answer
18 days ago

I am no expert on this, but there are built-in functions that may help: ImageFeatureTrack, MorphologicalComponents.

POSTED BY: Gianluca Gorni
Answer
18 days ago

there are previous posts on the subject. you'd start with ListConvolve or something - perhaps finding hulls or outlines instead. MorphologicalComponents is a new feature that aids you to set it all up (though it was done in earlier versions without).

but you still have to identify movement v. color that "is interesting". while matrix algebra may do what you need (highlight changes), it might not. what it can and can't do has limits. it's just some multiplication operations on a grid ... it isn't anything other than that really. so if your answer doesn't pop out of that, you'll need to make something (perhaps machine learning will work for you as a quick way to exemplify what you seek: i don't know).

a question is of course why you wouldn't do this with your own eyes. if you don't know how to start the project are you sure it's not a (longer) project than you should start? and does existing lab equipment already have the feature? what dooes today's cytometers do?

POSTED BY: John Hendrickson
Answer
16 days ago

from what i understand. you may have to learn matrix algebra to understand Convolve (unless mathematica examples found by google suffice what you need). you'll have to do some programming to get colors, position, motion. and even if that pops out of an advanced matrix: you have to code an interpretation. it would give you position, color, motion and is possible: i didn't say easy, i haven't seen your "video". previous posts have asked about cell outlining before (you can search for those posts they HAVE EXAMPLES already made)

if you decide to do machine learning you don't get out of coding. you have to code it to recognize what your looking for in it's distinct language! (plus you have to code the niches needs of the ml itself)

(your other option, unless your crafty or find nueralnet code that does it or are a matrix wizard: is to write math script to check values exactly as you need: overlap (hit) per color, frequency stats, and show running stats perhaps)

it becomes a coding preference and if your serious about getting meaningful repeatable results: my guess is unless you google and find code for it, it will be allot of coding to do

POSTED BY: John Hendrickson
Answer
16 days ago

i suggest looking at this (note it is a paper, not a usable product like a cytometer or other lab equipment - which you should check for)

https://www.nature.com/articles/srep43693

POSTED BY: John Hendrickson
Answer
16 days ago

Group Abstract Group Abstract