densereconstruction's Introduction

Dense Reconstruction of neurons from skeletons

Repo for constructing dense neuron segments from CATMAID skeletons, given probability maps.

To run this, you need:

vigra (you need the current master!)
sklearn
scipy

Input data:

membrane probability maps for the area of interest as tif files
json file with skeleton coordinates
bounding box coordinates to map from skeleton coordinates to probability maps
- these have to be stored in the first line of a file, in the order: x_min, x_max, y_min, y_max, z_min, z_max
pretrained random forest for oversegmentation edges (I will provide this)

There are two scripts, which implement different functionality:

make_dense: Constructs the dense segmentation for all skeletons and saves it.
get_statistcs: Performs the reconstruction and extracts relevant statistics over the segments.

Usage:

For make_dense:

python make_dense.py path_to_probability_maps path_to_exported_skeletons
path_to_rf bounding_box.txt folder_for_results

The results will be stored as tif slices.

For get_statistics:

python get_statistcs.py path_to_probability_maps path_to_exported_skeletons
path_to_rf bounding_box.txt folder_for_results --debug_folder folder_for_images

The results will be stored as csv (one csv for each skeleton), containing the following statistics (in units of pixel) for every slice:

Area
x-Radius
y-Radius
Radius

Values for a skeleton id, which is not present in a given slice, are zero.

The argument debug_folder is optional, if given, the reconstructions will be saved in the given folder.

Possible Issues

You need proper cutouts of probability maps
So far, I haven't come around to look at the interpolation of virtual nodes, so there are some segments missing, but I will do this once I work on learning from skeletons.
Reconstruction might fail for some cells, but for regular KCs it should work ok
- Also I can improve this with more features for the Random Forest or a suited agglomeration procedure.

densereconstruction's People