Segmentation Evaluation after border thinning

Python implementation of Rand error and Variation of Information

The following Jupyter Notebook allows you to evaluate the performance of your segmentation method after thinning the borders of the image segments to 1-pixel width.

Retrospective evaluation of the original ISBI-2012 segmentation challenge scoring system revealed that it was not sufficiently robust to variations in the widths of neurite borders. After evaluating all of these metrics and associated variants, it was found that specially normalized versions of the Rand error and Variation of Information best matched our qualitative judgements of segmentation quality:

• Foreground-restricted Rand Scoring after border thinning
• Foreground-restricted Information Theoretic Scoring after border thinning

Further details about the metrics can be found in the challenge publication.

Prerequisties

This code has been implemented in Jupyter. You need to install pyimagej which is a python wrapper for ImageJ.

To install pyimagej:

$ pip install pyimagej

To install scikit-image:

$ pip install scikit-image

You will also need to install Fiji. If you have already installed Fiji, then we are good!

Run

1. This repository includes a folder Fiji.app which is imported to access the segmentation metric libraries of ImageJ. If you are interested in importing your installed version of Fiji, change the path in Line 2 of the code to the desired Fiji.app folder path on your local machine.
2. The Groundtruth and Results (for example segmentation outputs of your network, method, etc.) in this repository are in .png format. If the formats of your data are different, change the formats in Line 3 accordingly.
3. Change the paths of the macros in Line 4 to the current working directory where this notebook runs in.