Analysing Localisation Microscopy data

• Analysing data not generated by PYMEAcquire
• Using an image splitting device for multi-colour ratiometric imaging

Distributed analysis and queues

PYME has a distributed analysis model whereby a server process manages Task Queues and distributes groups of frames to multiple different worker processes. These can either all be on the same machine, or be distributed across a network/cluster. These server and worker processes need to be running before starting the analysis. The launchWorkers.py script simplifies this by starting the server and the same number of workers as there are cores on the current machine. To communicate with each other the server and worker processes use a package called Pyro, and a pyro nameserver needs to be running so workers and servers can find each other. If there is no name server running on the local network, launchWorkers starts one, but this is probably only suitable/robust if only one machine is being used for analysis.

If you’re running the analysis on the data acquisition computer, and have a decent number of cores available (e.g. 8), better performance can be achieved by reserving a core for each of the Acqusition and Server processes (ie limiting the number of workers to 6 in our 8 core case). This can be done by explicitly specifying the number of workers to launch as an argument eg: launchWorkers 6.

Distributing over multiple computers

Distributing the analysis over multiple computers (a small ad-hoc cluster) is now easy:

• Make sure that the same version of PYME is installed on all machines
• Run launchWorkers on each machine you want to use.

Loading data

Once the server and worker processes are running, the data should be opened using dh5view. For data spooled to a .h5 file this can be performed as one would expect, by either specifying the filename on the command line or by ascociating dh5view with .h5 files. For data saved directly to a queue, the easiest way is probably to click the Analyse button on the Spooling panel in PYMEAcquire. Many protocols will do this automatically after a the intial pre-bleaching phase has been performed.

For data not originating from PYMEAcquire the process is a little more complex (see Analysing data not generated by PYMEAcquire). Similarly, special attention is needed for analysing simultaeneous multi-colour data (see Using an image splitting device for multi-colour ratiometric imaging).

Analysis settings

With the data loaded in dh5view, one should see something like:

The Analysis and Point Finding panes in the left hand panel control the analysis settings. The settings are:

Setting	Description
Threshold	The threshold for event detection. This is not a strict threshold, but rather a scaling factor applied to a local threhold computed from the estimated pixel SNR. Will generally not need to be altered (except for some of the interpolation based fits)
Type	The type of fit to perform
Interp	The type of interpolation to perform (only methods with Interp in their name)
Start at	The frame # should to start our analysis at
Background	What range of frames should be used for background estimation. Uses python slice notation.
Debounce r	The radius (in pixels) within which 2 events cannot be reliably distinguished
Shifts	The shift field to use for chromatic shift correction (only methods with Splitter in their name.
PSF	The PSF measurement to use (only Interp methods
Estimate Drift	Whether to estimate drift using fiduciaries [BROKEN]
Subtract background	Should we subtract the background before fitting (rather than just event detection)

Fit types

PYME offers a number of different fit types, and is easily extensible to support more. The current ones are, with the ones you’d usually want to use in bold:

Type	Description
ConfocCOIR	Determines a 3D COI from confocal/widefield data
Gauss3DFitR	Fits a 3D gaussian to confocal/widefield data
InterpFitR	Fits an interpolated PSF to localisation Data (3D)
LatFitCOIR	Determines the position of events by taking their centroid. Fast but not as good as a proper fit.
LatGaussFitFR	Simple 2D gaussian fitting.
LatObjFindFR	Just perform the object finding part of the fitting process
LatPSFFitR	Fits a symplified model of a widefield PSF (3D). Use InterpFitR instead
SplitterFitCOIR	Determines centroids when two channels are split onto separate halves of the CCD
SplitterFitFR	Like LatGaussFitFR but for split data
SplitterFitInterpR	Like InterpFitR but for split data (3D)
SplitterFitQR	Faster version of SplitterFitFR (ommits background parameters)
SplitterObjFindR	Like LatObjFindFR, but for split data
SplitterShiftEstFR	Used for estimating shift fields

Starting the fitting

Testing the object detection

Whilst the threhold factor is fairly robust, it is generally worth testing the detection by clicking the Test button. This performs the object finding step on a selection of frames spaced throughout the sequence. If this fails one should check the Camera.ADOffset setting in the metadata (accessible through the Metadata tab) to see if this is reasonable before attempting to tweak the detection threshold. (The ADOffset is estimated by taking a number of dark frames before the acquisition starts, and can be fooled if the room lights are on and/or the laser shutters are misbehaving). Metadata parameters can be edited by right clicking the appropriate field in the Metadata tab.

Launching the analysis tasks

Once satisfied with the event detection, the analysis proper can be started by clicking the Go button. The results will automatically be saved, either under the PYMEDATADIR directory (if the environment variable was set earlier), or in a directory called PYMEData in the users home directory (c:\\Users\\<username>\\ under windows).