Calibration and Bias
A software for evaluating model calibration, volume bias and a correlation between the two
Description:
We provide software to measure per-image expected calibration error (ECE), per-image volume bias (defined as difference between true volume and prediction based on a segmentation system) and the correlation between them.
More Information:
Originates from the AI-Driven Data Science research challenge (FAIR 2019–2023).
Responsible Research Lead: Matthew B. Blaschko
Authors: Teodora Popordanoska, Jeroen Bertels, Matthew B. Blaschko
Programming Language: Python
Categories: Supervised Learning, Images
Publications:
T. Popordanoska, J. Bertels, D. Vandermeulen, F. Maes, M. Blaschko, "On the relationship between calibrated predictors and unbiased volume estimation", 24th international conference on medical image computing and computer assisted intervention (MICCAI 2021), 2021.