Integrated ImageJ and Python workflow for super-resolution fluoresence image analysis.

Author: Eric Hidari

Version: 3.6

Issue Date: 02 June 2021

This workflow software uses GDSC SMLM ImageJ plugin for localisation fitting/fiduical drift correction/rendering super-res images, Redundant cross-correlation for correlation based drift correction, and DBSCAN for cluster analysis.

The software supports both GDSC SMLM 1 and GDSC SMLM 2, however some functions are not available for GDSC SMLM 2.

Make sure ImageJ and the GDSC SMLM plugin are installed. If you wish to use the redundant cross-correlation drift correction, make sure MATLAB (version not earlier than 2018b) is installed and included in system PATH.

Run the installer.exe file, fill in the requested paths and press Install. It will copy the source code into the destination folder, create configuration file, and set the PYTHONPATH to the destination folder.

• directory: dirctory containing the images to be analysed.

• image_condition: search criteria for images. If set to 'separate_images', images in each folder would be combined into a TIF stack before the analysis.

• pixel_size: pixel size in nanometer.

• sr_scale: default set to 8, scale for generating super-resolution images.

• camera_bias: in ADU, raw values recorded when no photons detected.

• camera_gain: ADUs/photon, see here for details.

• frame_length: ms per frame for exposure time.

• GDSC_SMLM_version: the version of GDSC SMLM used for the analysis.

• GDSC_SMLM_peak_fit: controls whether to run the GDSC SMLM peak fit in ImageJ or not, if False, only cluster analysis will be run.

• BG_measurement: controls whether to measure the background of the image stack (median pixel value). Value stored in median_intensity.txt.

• trim_track: run - whether to run the track trimmer. frame_number - how many frames remains for the analysis (the rest are trimmed). from_end - whether to keep the frames from the end or the beginning.

• signal_strength, precision, min_photons: parameters set for GDSC SMLM peak fit filters, see here for details. precision is in nm.

• spot_filter: only used with GDSC SMLM 2. filter_type can be either 'Single' or 'Difference', see here for details. smoothing is the size of the first smoothing filter, and smoothing2 is the size of the second filter for the 'Difference' spot filter.

• fiducial_correction: run - whether to run the drift correcter. correction_method - the method used to correct the drift. Four modes are supported for GDSC SMLM 1: 'auto_fid' (automatic fiducial correction), 'drift' (drift file), 'fid_file' (predefined fiducial coordinates), and 'corr' (redundant cross-correlation based drift correction); while only 'auto_fid', 'drift' and 'fid_file' are supported for GDSC SMLM 2. Furthermore, the drift plot and corrected image rendering are not supported for GDSC SMLM 2. smoothing_para and limit_smoothing control the smoothing of curve for the correction, see here for details.

• Here the 4 correction modes for fiducial_correction are briefly explained.

  • 'auto_fid': locate the fiducial markers in the FOV by thresholding out the pixels with intensity higher than fid_brightness, draw a square with fid_size around the center, and select fiducials lasting longer than fid_last_time in frames. Then it uses the GDSC SMLM builtin Drift calculator to correct the drift with the method setting Fiducial markers within an image.

  • 'fid_file': directly uses the predefined fiducial location in XY coordinates to correct the drift using the above method. Each image must be contained as the only image file in a folder, and the XY coordinates must be stored in the file Fiducials.txt within the folder, with the format as (delimited by tab):

  X    Y
  100    100
  200    200
  ...
  

  • 'drift': directly applies the predefined drift profile, and uses Drift calculator to correct the drift with the method setting Drift file. Each image must be contained as the only image file in a folder, and the drift file must be stored as Drift.txt within the folder, with the format as (delimited by tab):

  Time    X    Y
  1    0    0
  2    0.1    0.1
  3    0.2    0.2
  ...
  

  • 'corr': calculate the drift using redundant cross-correlation in MATLAB, save it in the drift file and correct the drift using the above method. bin_size sets the spatial bin pixel size in nm, and segpara sets the segmentation parameter in frame. See here for details and the original paper.

The rest of the parameters are for the python code of cluster analysis and characterisation. The detailed explanation of the analysis can be found in my thesis section 2.3.2 and 2.3.3, which can be downloaded here.

• fitresults_file_name: search criteria for fit results files. If set to 'default', it will be automatically set based on whether fiducial correction is enabled.

• temporal_grouping: group neighbouring localisations (in time and space) into the same burst. run - whether to run the temporal grouping. dThresh and min_loc control the DBSCAN parameters for grouping localisations in space as molecules; while tThresh and min_frame are the ones for grouping localisations in the time dimension as bursts. max_mol_area, min_on_prop and min_burst are extra filters can be applied: max_mol_area (in nm²) filters out localisations incorrectly grouped in space and thus occupying too large of an area; min_on_prop filters out localisations that are incorrectly grouped in time and thus leaving too long between bursts. 'min_burst' filters out bursts that blink less than the threshold during the track.

• cluster_analysis: group neighbouring subjects (cluster_subject, either 'burst' for bursts generated in temporal grouping, or 'loc' for the raw localisations) into clusters that represent protein aggregates. run - whether to run the cluster analysis. DBSCAN_eps_nm and DBSCAN_min_samples - DBSCAN parameters.

• length_measure: run - whether to run the length measurement. algorithm - algorithm to treat holes in skeletonised aggregate structures. 'blur' or 'close'; if 'blur', sigma is the Gaussian sigma; if 'close', sigma is the closing square size.

• eccentricity_measure and convexhull_measure: whether to run the eccentricity and convex hull area measurements.

• Rendering_SR: method to render the super-resolution images after cluster analysis. Set to False for no rendering, 'loc' for rendering with localisations only, 'loc_pre' for rendering with localisations (width = precision). See here for details.

• save_histogram: whether to save the information of individual molecules and bursts (generated in temporal grouping) into summary tables. Set to False for large imaging experiments can significantly save time and storage space.

Issued: 02-06-2021

Changes over v3_5:

1. Add a GUI installer, including automated installation of Python dependencies.

2. Add support for cross-correlation fiducial correction.

3. Add support for reading images of different folder.

4. Save drift profile plots.

5. Fix the conflict with newer versions of ImageJ.

6. Remove symlink and variable background functions.

7. Document the parameter settings.

Issued: 28-04-2019

Changes over v3_4:

1. Add temporal grouping: group localisations into molecules, then group them into individual bursts (using 2D and 1D DBSCAN, respectively.) Filters can be applied on burst number of molecule, molecule size and on time proportion of burst (blinking molecules can be removed). The X, Y positions and precisions in each group are averaged to enable further clustering.

2. Add DBSCAN clustering with molecules or bursts. DBSCAN with bursts will be suitable with STORM/PALM, while molecule suits with PAINT.

3. Remove the burst filter and burst analysis because temporal grouping does the same job.

1. Move to_ImageJ.json to individual analysis folders, allowing multiple analysis conducted at the same time.

2. Simplify the compatiablity code with GDSC SMLM 2.

3. Simplify the data saving code.

Issued: 10-03-2019

SR_toolkit_v3_5 is under development with a more accurate clustering strategy and therefore more punctuate reconstruction of super-resolution imge thanks to @StevetheChemist's advice. If you would like to use this toolkit for publication, please wait for v3_5 (before end of April 2019)

Changes over v3_3:

1. Add support of GDSC SMLM 2, although the package is still an unstable release, it gives the difference Gaussian filter which could be useful. The other parameters have not yet been included in SR_toolkit however if any of them is significant it will be added in a future version.

2. Add setup file for easy installation.

3. Simplify the code of generating fit header file.

Issued: 01-Feb-2019

Changes over v3_2:

1. Give the option of running the fiducial correction based on given fiducial coordinates.

2. Add fiducial correction parameters (fiducial last time, smoothing parameter, limit_smoothing).

3. Add camera bias, gain in the settings.

4. Add labelled images for clusters.

5. Add background calculation and variable background correction.

6. Rendering SR images after cluster analysis.

7. Add closing algorithm to calculate length. Instead of blurring the skeletion, the holes between the localisations are filled so that no loops will appear in the skeleton.

8. Fix some bugs that cause error_log to malfunction.

Note: Please put Rendering_SR.py in the same folder as GDSC_SMLM_peak_fit.py

=============================================================================

SR_toolkit_v3_2

Issued: 17-Jan-2019

Changes over v3_1:

1. Add compatibility with v3_0, i.e. run cluster fit only.

2. Add fiducial markers removal.

3. Save fiducial marker number in the summary.

4. Fix some bugs in the library.

=============================================================================

SR_toolkit_v3_1

Issued: 20-Dec-2018

Major changes over v3:

1. Incorporate ImageJ GDSC SMLM peak fit and the brightness-based fiducial
correction code into the analysis. Re-organise the main run_fit function
so it workflows the ImageJ part and the python part. The ImageJ part is
'lazy'--- it only runs if a previous analysis using the same parameters
is not found in the same root path. Notice that if only fitresults file are 
available (images are not available), this script will not run properly. 
Please use SR_toolkit_v3_0 for that. 

2. Trimming the frame to a certain frame_number is supported. This allows a
certain number of frames to be processed. 

3. Cluster analysis measurement can be turned off. This allows only GDSC SMLM
to run and therefore a quick check of the localisation number.

4. A symbolic link to the original image stack can be generated in the 
analysis folder. This feature needs python to be run in administrator mode.

Some minor changes:

1. All the analysis now are by default copied. This means no analysis files
will be appended to the raw image folder. 
2. Use JSON to output the logging information rather than xml. More human
readable.
3. Fix some bugs in the library.

Potential features update:

1. Integrate more ImageJ analysis including GDSC SMLM, background correction, 
find maxima (ThT counting), rendering SR images into the main analysis.
2. Doing DBSCAN on bursts (events lasting multiple frames count as one burst)
rather than localisations can be more precise.
3. Use burst analysis for fiducial correction. (need to check papers)
4. Length measurement is slow. Could be improved.
5. Integrate sqlite to pack the outputs into database automatically. (ongoing)

===============================================================================

SR_toolkit_v3_0

First published: 03-Dec-2018

Major changes over v2:

1. Pack the functions in the v2 scripts into 3 classes and their methods:
a. Analysis class - logs the analysis parameters, collects and saves the
outputs, with the flexibility of packing more analysis in the class.
A log file is saved to keep track of the parameters used in the analysis.

b. SR_fit class - represents a fit result file aka. an image.
Records the parameters passed by the analysis class, runs the 
DBSCAN cluster and the subsequent characterisations of the clusters.

c. cluster_track class - for each cluster identified in the cluster 
analysis. Stores the characteristics of the cluster.

2. Add functionalities: burst filter and burst analysis
These methods analyse the time dimension of each cluster with a certain 
fill_gap parameter to account for dye blinking.
   
Burst filter (set with a higher fill_gap) tries to remove clusters that 
have less than N bursts. This is assuming the real events are recurring 
and the non-specific bindings are random, non-recurring. One could play 
around the balance between min_burst and DBSCAN_min_samples options to 
see how the results changes.

Burst analysis (usually set with a lower fill_gap) measures the on and 
off time of each cluster. This is related with qPAINT analysis, which 
determines how many labelled molecules are at one spot. See Jungmann paper
for details. The remove_single option removes bursts that only lasts for
1 frame. They are likely to be camera noises.
    
3. Remove Nearest neighbour analysis: The NN analysis made assumption such 
as defining 5 times nearest neighbour distances as the 'neighbourhood';
the nearest neighbour distance is not really informative since the nearest 
neighbours usually come from the same binding spanning in mulitple frames.

Since the anlysis was intended to measure the density of localisations,
convex hull area is added here so that the 2D density of the aggregates
can be estimated. 

More elegant methods, such as a 'Concave hull' (with a smoothing constant)
or Ripley's k function could be added in the features.

Some minor changes:

1. Remove the generate SR_mage, ThT maxima, cropping functionality. 
The former two are more convenient done in imageJ, the latter will be added
in a future version, cropping either space or time.

2. The new analysis makes some uses of the Image.results.xls file which 
records the frame number, width and height. However it will throw an error
if the file does not exist. 

3. Timestamp is used to log the analysis, giving each image a unique ID 
which will be handy in SQL operations. A relationship can be easily built 
between the summary and histograms file. 

4. A folder will be generated to hold the analysis outputs everytime an 
Analysis object is generated. There is also an option of copying all analysis 
generated files to the results folder. This avoids overwriting the previous 
results.

5. More to cover...

Potential features update: 1. Create symlink for raw images. 2. Integrate ImageJ analysis including GDSC SMLM, background correction, find maxima (ThT counting), fiducial correction into the main analysis. 3. Doing DBSCAN on bursts (events lasting multiple frames count as one burst) rather than localisations can be more precise. 4. Use burst analysis for fiducial correction. 5. Length measurement is slow. Could be improved. 6. Integrate sqlite to pack the outputs into database automatically. 7. See minor changes 1, major changes 3. 8. Output JSON log file