fmi-faim / faim-wako-searchfirst Goto Github PK

Similar to these parameters in previous ImageJ macros:

// Define box in which for objects should be searched
bBx = 30; //x size of FOV is 2560 (for CV7000), 2000 (for CV8000)
bBy = 30; //y size of the FOV is 2160 (for CV7000), 2000 (for CV8000)
width = 2500;
height = 2100;

... we should implement these config parameters:

bounding-box:
  min_x: 0
  min_y: 0
  max_x: 2000
  max_y: 2000

This could potentially replace our usage of typer:

https://cyclopts.readthedocs.io/en/latest/index.html

When removing points of overlapping labels, we currently only remove the conflicting points themselves (of the label with the smaller area).

Instead, we should remove all points of the smaller object.

A simple parallelization of the for loop over all .tif files should go a long way:

with ThreadPoolExecutor(max_workers=num_threads) as executor:
    results = list(tqdm(executor.map(process_tif, tif_files), total=len(tif_files)))

An optional "pre-processing" step could be processing the list of files with an external command, e.g. a Fiji script called from the command line. This allows running an arbitrary segmentation while still using the filter and sample functionality of this repository.

The best entry point is probably the source folder (as it's a single parameter), but also the list of files would be possible.

Use case:

• Run Fiji script to call Weka Trainable Segmentation, write output masks to new folder
• Return the list of files, so we can hand it over to _process_tif, with an option to skip the segment step (i.e. only run label).
• Process filter and sample as usual.

Add new function segment.apply_weka_classifier() that takes a classifier.model file and applies it using pyimagej and trainableSegmentation.

/cc @SabineReither

Add new function segment.apoc_classifier() that takes a my_model.cl file and applies it using https://github.com/haesleinhuepf/apoc.

When acquiring objects that vary a lot in size (e.g., small objects fit in a single field of view, large objects require 3x2 tiles in second pass acquisition), we'd like to have a sampling that records:

• the object center for objects that fit into a single FOV,
• the centers of grid tiles (with definable overlap) centered on each object, for objects larger than a single FOV.

@nrepina what should be the behavior if two objects are close to each other? If we just sample per-object, we have the possibility of overlap (and therefore bleaching) between tiles belonging to two different objects.

I would be nice if we can keep the run script general, and have full flexibility in defining the processing via the config file.

steps:
  - segment: cellpose
  - filter:  # all functions should accept 'labels' and change inplace (and an optional other_channel argument?)
    - bounding_box
    - object_size
  - sample: grid

cellpose:
  - model: /path/to/model
  - diameter: 50.0
  - cellprob_threshold: 0.0
  - flow_threshold: 0.4

bounding_box:
  - min_x: 0
  - min_y: 0
  - max_x: 2000
  - max_y: 2000

object_size:
  - min_size: 100

The items in steps should correspond to functions in the faim_wako_searchfirst.segment, faim_wako_searchfirst.filter and faim_wako_searchfirst.sample modules, respectively.

In addition to steps, each function's arguments are listed under the function name. (This entry should be optional if the function requires no additional arguments.)