Awesome

CHAOS-evaluation

This repo contains evaluation metrics of CHAOS challenge. The evaluation metrics are;

1. Sørensen–Dice coefficient (DICE)
2. Relative absolute volume difference (RAVD)
3. Average symmetric surface distance (ASSD)
4. Maximum symmetric surface distance (MSSD)

For further information about metrics and scoring system please refer to CHAOS_Metrics_and_Evaluation_Method.pdf file and https://chaos.grand-challenge.org/Evaluation/

The evaluation code is shared with sample submission. Evaluation of CHAOS is handled via MATLAB language. In addition, Python and Julia versions of the code are presented.

Files and Folders

We provide an example evaluation code for sample segmentation submission. Besides, we prepared a mini-experiment to compare metrics (only in MATLAB).

MATLAB Version

There are two evaluation codes.

Requirements

The evaluation codes were written and tested with MATLAB R2018a. Image Processing Toolbox is necessary to run.

1. Evaluation of sample submission (evaluate3D.m)

This is an evaluation of a sample submission placed in Data_3D\Segmentation. The segmented and reference data are stored in a series of PNG files. They are imported as 3D volumes. DICOM files are used for their header info which is necessary for transforming voxel values into real-world values.

The script is ready for use. After downloading the repo, just running of evaluate3D.m file is enough to perform the evaluation. (natsort.m and natsortfiles.m files are used to import files in the right order. https://uk.mathworks.com/matlabcentral/fileexchange/47434-natural-order-filename-sort)

After evaluation the result will be:

DICE=0.978 RAVD=0.665 ASSD=0.734 MSSD=14

2. Comparison of metrics with different artifacts (evaluate2D_metric_compare.m)

In CHAOS challenge, we have received many questions about why we are using multiple metrics instead of using just a single metric (such as DICE as many papers in literature). This mini-experiment was designed to analyze the output of evaluation metrics under different segmentation artifacts. There are reference data(ref.png), an original segmentation(seg1.png) and some modified versions of seg1.png with different kind of artifacts(seg2.png ... seg7.png). As it can be observed from metric results, each metric has advantages and disadvantages about determining particular errors in segmented volumes. Also, DICE does not generate significant results in many cases. We hope that this experiment clarifies some questions about usage of multiple metrics.

Python Version

We have implemented the evaluation code in Python because of high demands from many researchers. evaluate3D.py and CHAOSMetrics.py files are Python implementation of evaluate3D.m and CHAOSMetrics.m files.

Requirements

The evaluation code was written and tested with Python 3.7, and it needs the libraries below:

(They are presented with used versions.)

• numpy (1.16.4)
• scipy (1.3.0)
• Pydicom (1.2.2)
• opencv-python (4.1.0.25)
• SimpleITK (1.2.0)
• sklearn2 (2.0.0.13)

You may install them using PyPI with the commands below:

pip install numpy
pip install scipy
pip install pydicom
pip install opencv-python
pip install SimpleITK
pip install sklearn2

After downloading the repo and installing necessary libraries, you may run evaluate3D.py file to perform the evaluation. The result will be:

DICE=0.978 RAVD=0.665 ASSD=0.734 MSSD=14

Julia Version

Julia is a high-level, general-purpose dynamic programming language designed for high-performance numerical analysis and computational science. Since Julia is a new language, it doesn't have as many communities and libraries as Python or MATLAB. We would like to make a small contribution to Julia community because we believe that Julia will be a popular programming language in the future. evaluate3D.jl and CHAOSMetrics.jl files are Julia implementation of evaluate3D.m and CHAOSMetrics.m files.

Requirements

The evaluation code was written and tested with Julia 1.1.1, and it needs the packages below:

(They are presented with used versions.)

• FileIO (v1.0.7)
• Images (v0.18.0)
• ImageMagick (v0.7.4)
• NearestNeighbors (v0.4.3)
• LocalFilters (v1.0.0)
• DICOM (v0.3.1)

You may install them using Julia's package manager (Pkg) with the commands below:

import Pkg
Pkg.add("FileIO")
Pkg.add("Images")
Pkg.add("ImageMagick")
Pkg.add("NearestNeighbors")
Pkg.add("LocalFilters")
Pkg.add("DICOM")

Important note: In Windows, there may be some errors while installing ImageMagick. This error is caused by some modifications on the download links of some packages in the install script. You may need to edit install script to install.

After downloading the repo and installing necessary packages, you may run evaluate3D.jl file to perform the evaluation. The result will be as same as MATLAB version:

DICE=0.978 RAVD=0.665 ASSD=0.734 MSSD=14

Contact

If you have a question, please consider looking web page of CHAOS: https://chaos.grand-challenge.org/News_and_FAQ/.

If you cannot find the answer, you may contact with us: emrekavur@gmail.com