SegmentationTools.jl

This package provides a set of visualization tools for segmentation. It is geared towards neuron segmentation, but many of the tools can be used for other segmentation problems as well.

Prerequisites

This package requires you to have previously installed the FlavellBase.jl, ImageDataIO.jl, UNet2D.jl, NRRDIO.jl, WormFeatureDetector.jl, and SegmentationStats.jl packages from the flavell-lab github page, and that you have succesfully configured WebIO if you're using this package's visualization tools with Jupyter. It is designed to interface with the pytorch-3dunet package, also in the flavell-lab github page.

Additionally, the example code provided here requires the ImageDataIO package is loaded in the current Julia environment.

Cropping the worm

Before performing any other operations on an image, it is useful to crop out non-worm regions; this will speed up the subsequent operations, and can also improve UNet output:

crop_parameters = Dict()
cropping_errors = crop_rotate!(param_path, param, t_range, [ch_marker], crop_parameters, save_MIP=true)

Making weighted HDF5 files

The make_unet_input_h5 method can generate appropriately-weighted HDF5 files from raw NRRD data and labeled NRRD files:

make_unet_input_h5(param_path, path_nrrd_crop, t_range, ch_marker, get_basename)

These HDF5 files can then be fed as input to the UNet.

Visualizing the UNet's predictions

The display_predictions_3D method can display the UNet's predictions in comparison with the raw and labeled data.

# load the data (say it's dataset 50)
raw, label, weight = load_training_set("/path/to/data/50.h5")
# load the UNet's predictions
predictions = load_predictions("/path/to/data/50_predictions.h5")
# display the predictions
# the order of plots will be raw data + label, weights + label, predictions + label, predictions vs label match
display_predictions_3D(raw, label, weight, [predictions])

Instance segmentation

After the UNet has been verified to be giving reasonable output, the next step is to turn the UNet's semantic segmentation into an instance segmentation. The instance_segmentation_watershed function does this instance segmentation and outputs the result to various files, which can be used during later steps in elastix registration:

results, errors = instance_segmentation_watershed(param, param_path, nrrd_crop_dir, t_range, get_basename, save_centroid=true, save_signal=true, save_roi=true)

Visualizing instance segmentation

Assuming that you have successfully instance segmented the data, you can view the resulting ROIs, in comparison with the raw and predicted data, to ensure that instance segmentation was successful:

view_roi_3D(raw, predictions, img_roi)