Thanks for your help Octavio. It occurred to me that I can also adjust the digital gain and offset with my PMT. Now I'm thinking I could tweak these parameters to keep my pixels in the dynamic range (none saturated, none black). Which to use, gain or offset? Would anyone know if there is an empirical method to optimize the gain/offset for dynamic imaging?

Hi Steve,

You wrote "initially dim neurons might weigh in more heavily". This suggests that you estimate "F0" (or what you call the background, F) from an *initial* baseline period? If so, I think you will have better results by doing what many other in vivo GCaMP experimentalists do: estimate F0 (aka baseline) over time, during the entire recording, for each neuron, using one of several common approaches. I can't tell you which approach is best, but if you dig into the methods of any recent in vivo GCaMP paper, you'll see that while the details vary from lab to lab, there are probably just 3-4 fairly common fundamental approaches.

I would say that generally one may be faced with slow-timescale drifts in baseline (which could be due to bleaching, although with GCaMP this is typically not a problem), and on a more rapid timescale, the challenge is to identify event-free recording periods wherein one can measure the baseline.

One approach is to look at the *distribution* of the signal in e.g. a 30 second window, and identify the baseline as some low percentile; or some other feature of the distribution.

Another approach is to identify Ca events, and then average across all samples outside of these events to obtain the baseline. Of course this begs the question, how do you identify events without knowing the baseline? So typically this is an iterative approach, where you start with some estimate of the baseline and iterate until reaching a suitable criterion. But there are also papers where event-free periods are manually identified, and the baseline is measured from these.

David Tank (see Dombeck papers, ca. 2009) introduced an approach that further classifies Ca events to detect false positives, based on the observation that negative events are false, and therefore the distribution of negative event amplitudes and durations can be used to predict the corresponding distribution of positive events. Essentially this allows you to assign significance values to positive events based on their duration and amplitude.

Z-score can be a useful tool as well.

Most of these approaches assume that you have a "baseline", that is to say, that the neuron is mostly quiet, and only has Ca events infrequently. This is true for almost all excitatory neurons. However, certain types of interneurons, for example, can have very high firing rates in vivo - combined with the kinetics of the Ca sensor, this can result in a trace that has no obvious baseline. If you nevertheless must use GECIs to study neurons with high firing rates in vivo, other approaches are used - I'm not particularly familiar, but a simple approach is to look at deviations around a moving-window average.

Feel free to post some short example traces in raw A/D units.

Thanks Ben. Yes you are correct that for my preliminary data I am deriving F0 from an initial baseline. Fortunately, my cells are rather quiet and return to baseline after stimulation, so this does not seem to be an issue. But yes, to be on the safe side, our lab is implementing some of the methods you suggested to account for noisy baselines.

In my original message, I was more concerned about dead (pure black) pixels. In some cases the baseline might be deceivingly quiet in dim cells because they are outside the dynamic range of the PMTs. I think the solution to this problem is to adjust the digital offset — likely adding a new set of challenges.

I'm imaging spinal cord, so one of the biggest struggles is getting a strong enough signal through myelin, elastomer, glass, and water!

Here is the prep I am using:

https://www.jove.com/video/52196/a-procedure-for-implanting-spinal-chamber-for-longitudinal-vivo