This is a collection of the software used in the Chung Lab's 2018 Shield paper. It is packaged as a docker container (see https://www.docker.com/) so that it can be easily and accurately deployed in labs with a minimum of IT administration support. The dockerfiles can also be used as recipes to build the components.
The software handles the following tasks:
-
Stitching and destriping volumes acquired from a SPIM microscope or similar. Destriping is performed using PyStripe (https://github.com/chunglabmit/pystripe). The stitching alignment is performed using Terastitcher (https://github.com/abria/TeraStitcher) and the aligned volume is written to disk using TSV (https://github.com/chunglabmit/tsv).
-
Alignment to the Allen Brain Atlas. This is done using Nuggt (https://github.com/chunglabmit/nuggt). The alignment pipeline that was used is outlined here: https://github.com/chunglabmit/nuggt#alignment-pipeline
-
Segmentation. The shield-2018 package includes a command, shield-2018-segmentation, that was used to find putative cell centers in the paper. These were then proofread using nuggt. The method robustly finds roughly spherical light objects against a dark background but finds false positives in areas of fiber. It is suggested that users supplant this with either manual editing or a machine-learning method that classifies true and false positive detections. The command parameters are detailed below.
-
Assignment of detections to the atlas. This is done using the count-points-in-region command.
The preferred method of installation of shield-2018 is as a docker container. If you have Docker installed, this is available as
> docker pull chunglabmit/shield-2018
A typical invocation might be
docker run -v /path-to-stack:/stack \
-v /path-to-output:/output \
chunglabmit/shield-2018 \
shield-2018-segmentation \
--input /stack/img_*.tiff \
--output /output/result.json
Commands from the Terastitcher and Nuggt packages can also be run from this docker. See the Nuggt documentation and the tutorial for more help.
The shield-2018-segment command finds cell centers in a stack of images. The output is a .json file containing a list of lists with each element of the inner list being the Z, Y and X coordinate of one of the cell centers. The format of the command is:
shield-2018-segment \
--input <input-stack-expression> \
--output <output-filename> \
[--block-size-x <block-size-x>] \
[--block-size-y <block-size-y>] \
[--block-size-z <block-size-z>] \
[--crop-x <crop-x>] \
[--crop-y <crop-y>] \
[--crop-z <crop-z>] \
[--padding <padding>] \
[--io-threads <io-threads>] \
[--processing-threads <processing-threads>] \
[--t1min <t1min>] \
[--t1max <t1max>] \
[--t2min <t2min>] \
[--t2max <t2max>] \
[--dog-low <dog-low>] \
[--dog-high <dog-high>] \
[--log-level <log-level>] \
[--log-file <log-file>] \
[--log-format <log-format>]
where
- input-stack-expression is a glob expression for listing the .tif files in the input stack, e.g. "/path-to/img_*.tif". The files are read in alphabetical order. Typically, the names are zero-padded, e.g. "img_0001.tif", so that they will be read in the correct order.
- output-filename the name of the .json file containing the cell centers.
- block-size-x, block-size-y, block-size-z The processing is done by breaking the volume into blocks in order to process teravoxel volumes on machines with less than terabyte memory sizes. These parameters specify the size of a block. Segmentation requires enough memory to hold (block-size-z + 2 * (padding)) * 32 bytes of memory for each pixel in one image plane and an additional block-size-x * block-size-y * block-size-z * processing-threads * 32 bytes to handle parallel processing.
- crop-x, crop-y, crop-z Processing can be done on a portion of the
stack (or processing can be broken up into manageable chunks and combined)
by cropping the stack to the dimensions given by these parameters.
Each one is specified as a minimum and maximum coordinate, e.g.
--crop-x 1000,3000processes the portion of the stack from x=1000 to x=3000. - padding is the amount of padding to add to the X, Y and Z boundaries
of a block. This should be roughly 3x the background sigma specified
by the
--dog-highswitch. The default is 30. - io-threads is the number of threads used when reading the image stack. The default is 4.
- processing-threads is the number of threads used during parallel processing. The default is the number of cores on the computer being used.
- t1-min, t1-max, t2-min, t2-max These are threshold limits for the seed-finding and segmentation thresholding step. T1 is the lower threshold used for segmentation and (T1+T2) / 2 is the threshold used for the seed finding step. These numbers can be adjusted if the method is finding false positive or false negative cell centers.
- dog-low, dog-high These are the sigmas for the difference of gaussians. dog-low is designed to pick the scale of the cells and should be roughly 1/4 of the cell width in voxels. dog-high is designed to pick the scale of the background (which is subtracted from the foreground signal). The defaults are 3 and 10.
- log-level is the minimum log message level. Possible values are "DEBUG", "INFO", "WARNING", "ERROR" or "CRITICAL". The default is "WARNING".
- log-file is the path to the file in which to record the log messages. The default is to display log messages on the console.
- log-format is the display format for log messages. See https://docs.python.org/3/howto/logging.html#changing-the-format-of-displayed-messages for help.
The alignment for the Shield paper was done using files derived from the Allen Brain Mouse Atlas (2018). These files are available via the GIT repository, https://github.com/chunglabmit/shield-2018-assets See the README.md file for more details.
The docker image stores these files at /allen-brain-mouse-atlas in its filesystem, so they are readily available for use in the Nuggt alignment commands.