Anybody knows methods to infer dynamic spatial patterns evaluating signal intensity and topological distribution in 2D from a time lapse recording?

[Quique]

We are biologist and take movies of developmental processes, employing a microscope. We can monitor through time the activation of a particular molecule by increments in the intensity of signal (not on/off but gradual and not limited to single pixels). This occurs in the plane in different positions (xy coordinates) over time (t) and we like to identify if these occurrences follow a specific patterns (temporal or spatial). We can repeat these recordings in different animals and we would like to test how robust it could be any pattern detected. Would it be reproducible?
We do not know how to do this?
I’m not sure if this is the right forum but I will appreciate any help

[crackwitz]

can you show us such pictures?

[jsxyhelu]

picture and example please

[Quique]

Thanks for the reply

The images are huge (movies of 5000 time points).

I’m enclosing 1 gif with a sequence of images (200 time points each) showing different activation patterns. I had to reduce the resolution 5 times to be able to post them. They are much much better.

The idea is to see if in the 5000 time points a pattern is present or if the activation is stochastic. If there is a pattern or if a pattern is generated overtime. then we are in business.

Activation are the lightning flashes that can be observed with a duration of 2-4 time point. They appear in different positions (these are muscles in which signalling by calcium becomes transiently active). The images represent 2 segments of the abdomen of a fly (top and bottom), each one with a full set of muscles

[Quique]

Thanks for the reply

The images are huge (movies of 5000 time points).

I’m enclosing 1 gif with a sequence of images (200 time points each) showing different activation patterns. I had to reduce the resolution 5 times to be able to post them. They are much much better.

The idea is to see if in the 5000 time points a pattern is present or if the activation is stochastic. If there is a pattern or if a pattern is generated overtime. then we are in business.

Activation are the lightning flashes that can be observed with a duration of 2-4 time point. They appear in different positions (these are muscles in which signalling by calcium becomes transiently active). The images represent 2 segments of the abdomen of a fly (top and bottom), each one with a full set of muscles

[crackwitz]

curious data!

this would take serious statistics, machine learning, deep learning perhaps.

I’d recommend looking for collaboration at your research institution or a nearby one.

I’d also recommend sharing the dataset, if legally feasible.

you’d probably want to reduce the dimensionality first, i.e. go from “soup of pixels” to determining “areas”, each of which lights up as an integral thing.

some segmentation in time and space would give you descriptions of these flashes, their extent in time and space.

you probably know better than I the literature on detecting repeating patterns. if this was halfway regular, I’d say autocorrelation.

you probably don’t just mean repeating patterns but correlations between activations that have some distance in space and time.

there might be literature on this in neuroscience and speech recognition.

there may also be in “process mining”. their situation is that they have “logs” of events and they try to recover an automaton or other temporal description from these logs. this is “squishy”, i.e. the complexity/fidelity of the recovered model can be controlled.

[Quique]

@crackwitz

curious data!

this would take serious statistics, machine learning, deep learning perhaps.

I guess so

I’d recommend looking for collaboration at your research institution or a nearby one.

Looking for that

I’d also recommend sharing the dataset, if legally feasible.

Legally feasible but where do you recommend to share?

you’d probably want to reduce the dimensionality first, i.e. go from “soup of pixels” to determining “areas”, each of which lights up as an integral thing.

We created rows for each of the individual topological units (they are conserved over time) and calculate numerical signal values for each over time. I think these should be the data to analyse

some segmentation in time and space would give you descriptions of these flashes, their extent in time and space.

I guess we can get this from the data extracted for the ROIs. The question is how to integrate the information

you probably know better than I the literature on detecting repeating patterns. if this was halfway regular, I’d say autocorrelation.

It does not look very regular but we could try to perform autocorrelation

you probably don’t just mean repeating patterns but correlations between activations that have some distance in space and time.

what I mean with patterns is to see if after activation of let’s say Roi 1, we get immediate activation of Roi 2 and delayed simultaneous activation of Roi 3 and 4 (this is an example). I think that some machine learning, deep learning or convolutional network analyses could extract individual sets of information from the whole process. These individual sets could be further combined.
May be I’m crazy!

there might be literature on this in neuroscience and speech recognition.

I do not know any but I will keep searching

there may also be in “process mining”. their situation is that they have “logs” of events and they try to recover an automaton or other temporal description from these logs. this is “squishy”, i.e. the complexity/fidelity of the recovered model can be controlled.

This is a good point. I never thought about it

Any further recommendation?

many thanks

[laurent.berger]

Have you make some progress?