miseval: a metric library for Medical Image Segmentation EVALuation

The open-source and free to use Python package miseval was developed to establish a standardized medical image segmentation evaluation procedure. We hope that our this will help improve evaluation quality, reproducibility, and comparability in future studies in the field of medical image segmentation.

Guideline on Evaluation Metrics for Medical Image Segmentation

1. Use DSC as main metric for validation and performance interpretation.
2. Use AHD for interpretation on point position sensitivity (contour) if needed.
3. Avoid any interpretations based on high pixel accuracy scores.
4. Provide next to DSC also IoU, Sensitivity, and Specificity for method comparability.
5. Provide sample visualizations, comparing the annotated and predicted segmentation, for visual evaluation as well as to avoid statistical bias.
6. Avoid cherry-picking high-scoring samples.
7. Provide histograms or box plots showing the scoring distribution across the dataset.
8. For multi-class problems, provide metric computations for each class individually.
9. Avoid confirmation bias through macro-averaging classes which is pushing scores via background class inclusion.
10. Provide access to evaluation scripts and results with journal data services or third-party services like GitHub and Zenodo for easier reproducibility.

Implemented Metrics

Options for Boundary Distance computation

List of available distances:
  Bhattacharyya distance              bhattacharyya
  Bhattacharyya coefficient           bhattacharyya_coefficient
  Canberra distance                   canberra
  Chebyshev distance                  chebyshev
  Chi Square distance                 chi_square
  Cosine Distance                     cosine
  Euclidean distance                  euclidean
  Hamming distance                    hamming
  Jensen-Shannon divergence           jensen_shannon
  Kullback-Leibler divergence         kullback_leibler
  Mean absolute error                 mae
  Taxicab geometry                    manhattan, cityblock, total_variation
  Minkowski distance                  minkowsky
  Mean squared error                  mse
  Pearson's distance                  pearson
  Squared deviations from the mean    squared_variation

Distance Pooling (how to combine computed distances to a single value):
  Distance Sum                        sum
  Distance Averaging                  mean
  Minimum Distance                    amin
  Maximum Distance                    amax

How to Use

Example

# load libraries
import numpy as np
from miseval import evaluate

# Get some ground truth / annotated segmentations
np.random.seed(1)
real_bi = np.random.randint(2, size=(64,64))  # binary (2 classes)
real_mc = np.random.randint(5, size=(64,64))  # multi-class (5 classes)
# Get some predicted segmentations
np.random.seed(2)
pred_bi = np.random.randint(2, size=(64,64))  # binary (2 classes)
pred_mc = np.random.randint(5, size=(64,64))  # multi-class (5 classes)

# Run binary evaluation
dice = evaluate(real_bi, pred_bi, metric="DSC")    
  # returns single np.float64 e.g. 0.75

# Run multi-class evaluation
dice_list = evaluate(real_mc, pred_mc, metric="DSC", multi_class=True,
                     n_classes=5)   
  # returns array of np.float64 e.g. [0.9, 0.2, 0.6, 0.0, 0.4]
  # for each class, one score

Core function: Evaluate()

Every metric in miseval can be called via our core function evaluate().

The miseval eavluate function can be run with different metrics as backbone.
You can pass the following options to the metric parameter:

• String naming one of the metric labels, for example "DSC"
• Directly passing a metric function, for example calc_DSC_Sets (from dice.py)
• Passing a custom metric function

List of metrics : See miseval/__init__.py under section "Access Functions to Metric Functions"

The classes in a segmentation mask must be ongoing starting from 0 (integers from 0 to n_classes-1).

A segmentation mask is allowed to have either no channel axis or just 1 (e.g. 512x512x1), which contains the annotation.

"""
Arguments:
    truth (NumPy Matrix):            Ground Truth segmentation mask.
    pred (NumPy Matrix):             Prediction segmentation mask.
    metric (String or Function):     Metric function. Either a function directly or encoded as
                                     String from miseval or a custom function.
    multi_class (Boolean):           Boolean parameter, if segmentation is a binary or multi-class
                                     problem. By default False -> Binary mode.
    n_classes (Integer):             Number of classes. By default 2 -> Binary
    kwargs (arguments):              Additional arguments for passing down to metric functions.

Output:
    score (Float) or scores (List of Float)

    The multi_class parameter defines the output of this function.
    If n_classes > 2, multi_class is automatically True.
    If multi_class == False & n_classes == 2, only a single score (float) is returned.
    If multi_class == True, multiple scores as a list are returned (for each class one score).
"""
def evaluate(truth, pred, metric, multi_class=False, n_classes=2, **kwargs)

Installation

• Install miseval from PyPI (recommended):

pip install miseval

• Alternatively: install miseval from the GitHub source:

First, clone miseval using git:

git clone https://github.com/frankkramer-lab/miseval

Then, go into the miseval folder and run the install command:

cd miseval
python setup.py install

Author

Dominik Müller\ Email: dominik.mueller@informatik.uni-augsburg.de\ IT-Infrastructure for Translational Medical Research\ University Augsburg\ Bavaria, Germany

How to cite / More information

Dominik Müller, Dennis Hartmann, Philip Meyer, Florian Auer, Iñaki Soto-Rey, Frank Kramer. (2022)
MISeval: a Metric Library for Medical Image Segmentation Evaluation.
PubMed: https://pubmed.ncbi.nlm.nih.gov/35612011/
DOI: https://doi.org/10.3233/shti220391
arXiv e-print: https://arxiv.org/abs/2201.09395

@Article{misevalMUELLER2022,
  title={MISeval: a Metric Library for Medical Image Segmentation Evaluation},
  author={Dominik Müller, Dennis Hartmann, Philip Meyer, Florian Auer, Iñaki Soto-Rey, Frank Kramer},
  year={2022},
  journal={Studies in health technology and informatics},
  volume={294},
  number={},
  pages={33-37},
  doi={10.3233/shti220391},
  eprint={2201.09395},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

Thank you for citing our work.

License

This project is licensed under the GNU GENERAL PUBLIC LICENSE Version 3.\ See the LICENSE.md file for license rights and limitations.