Automated Cell Nuclei detection for Large-Volume Electron Microscopy of Neural Tissue

This is a supplementary to the ISBI 2014 paper "Automated Cell Nucleus Detection for Large-Volume Electron Microscopy of Neural Tissue".

Data and ground-truth annotations

Raw data

The raw data (HDF5 file, 4 GB, md5sum 8b1f88fd0cd57874dd8f0f0b74be61ba) can be requested using the form at the bottom of this page. The hd5 file is split up into 3 tar files for convenience.

This file uses the HDF5 data format for chunked and compressed storage. Regions of interest can be read from the file using Matlab's h5read command, using Python and h5py library as well as many other languages (see the wikipedia article for a list).

The data is stored in the group G1/20130722_132814 as a dataset of size 1024 x 768 x 7552. It was written in chunks of 64 x 64 x 64 using the deflate-1 OPT compression filter. Note that the data is in x,y,z axis order, which requires a transpose when read from C or python (see example below), and also from Matlab (because it uses y,x,z order).

An example python script to read a slice from the raw data and save it as a PNG image can be found here.

Ground-truth

Ground-truth annotations are provided as CSV files (separator: comma)with x,y,z and r columns.
x,y,z specify the manually marked approximate center position of each neuronal or glial nucleus;
r the manually estimated radius. Because nuclei are in 3D and not always spherical these are only rough estimates.
Two files are provided: In the first, the complete set of neuronal and glial nuclei annotations are given. In the second, we have removed those annotations for which the sphere touches the volume border in order to simplify evaluation of automated detection algorithms (see paper).

• Original ground-truth (all neuronal and glial nuclei)
• Edge clean ground-truth (annotations touching border removed)

Source Code

• Code for block-wise thresholding and connected-component labeling can be found in the blockedarray github repository.
  Example usage from C++ can be found in the ccpipeline.cpp file.
• Code for block-wise component accumulator can be found in the blockWiseComponentAccumulator.m
  This matlab function works on a labelled volume dataset stored in an HDF5 file (created by the blockedarray). It calculates bounding box, centroid, and coordinate list for all components except the label zero (assumed background). The output is written to a .mat file.
• Code for component-wise morphological filtering can be found in the file componentWiseFiltering.m
  This matlab function works on a component list (created by the blockWiseComponentAccumulator.m) stored in a .mat file. It calculates and writes the new components to another .mat file.