Metrics Reloaded

A Python implementation of Metrics Reloaded - A new recommendation framework for biomedical image analysis validation.

ℹ️ This is a fork of the Project-MONAI/MetricsReloaded repo. We created this fork because the original repo is no longer maintained (last commit on Sep 6, 2023).

ℹ️ In this fork, we have added some additional metrics, such as relative volume error (RVE), lesion-wise F1 Score, lesion-wise Sensitivity, and lesion-wise positive predictive value (PPV) and fixed some bugs (e.g., #1)

ℹ️ You can use the compute_metrics_reloaded.py wrapper script to compute metrics using the MetricsReloaded package. For the installation and usage of the script, see our quick start guide.

Installation

Using git

Create and activate a new Conda environment:

conda create -n metrics python=3.10 pip
conda activate metrics

Clone the repository:

git clone https://github.com/ivadomed/MetricsReloaded
cd MetricsReloaded

Install the package:

python -m pip install .

You can alternatively install the package in editable mode:

python -m pip install -e .

This is useful if you are developing MetricsReloaded and want to see changes in the code automatically applied to the installed library.

Overview

All functions used in this framework are documented here <https://metricsreloaded.readthedocs.io/en/latest/?badge=latest>

The repository is organised in three main folders:

• processes: this allows for the combination of multiple metrics in an evaluation setting and reflects the tasks tackled in the MetricsReloaded framework namely:
  1. Image Level Classification (ILC)
  2. Semantic Segmentation (SS)
  3. Object Detection (OD)
  4. Instance Segmentation (SS)
• metrics: this contains all the individual metrics reported and discussed in the MetricsReloaded guidelines. Those are classified as either:
  1. pairwise_measures - all metrics considering a binary or multiclass input
  2. prob_pairwise_measures - all metrics relying on multi threshold and/or probabilistic input
  3. calibration_measures - all metrics related to the evaluation of the calibration of probabilistic outputs
• utility: this contains all ancillary function relevant notably for aggregation of metrics, or preliminary tools required for complext assignments prior to metrics calculation notably in the case of Object Detection and Instance Segmentation.

Useful links to get started

To see examples on how to process different cases of tasks please look into the

example:	example_ss.py

Pictorial representation

Metrics Reloaded fosters the convergence of validation methodology across modalities, application domains and classification scales

Support

For any questions or remarks, please contact metrics-reloaded-package(at)dkfz-heidelberg.de.