Computational reproductibility: an overview illustrated with examples from the medical imaging research community

In the last few years, there has been a growing awareness of reproducibility concerns in many scientific fields. In [1] the analysis of a single neuroimaging dataset by 70 independent analysis teams revealed substantial variability in reported results, with high levels of disagreement across teams. Despite the increase in awareness, reproducibility is difficult to achieve because of the multiple sources of variability, from the computational environment and software versions, to the tool selection and the more global methodological approach. 

This presentation will introduce various concepts and definitions related to reproducibility, then focus on computational reproducibility defined as one's ability to get identical outputs when applying the same treatments to the same set of inputs. The causes for the lack of reproducibility on the computational environment level are mainly related to the library dependencies and their evolution over time, but also to numerical instability due to floating point arithmetic issues (rounding errors, hardware and compiler optimizations). We will illustrate these sources of irreproducibility through concrete examples ranging from papers to data and different levels of code. We will also present some tools and best practices to improve computational reproducibility. The presentation will be based on experience gained within the ReproVIP ANR project [2] and the tutorial organized at MICCAI 2023 [3].

[1] Variability in the analysis of a single neuroimaging dataset by many teams, R Botvinik-Nezer et al., https://doi.org/10.1101/843193

[2] https://www.creatis.insa-lyon.fr/reprovip/

[3] https://www.creatis.insa-lyon.fr/miccai2023/

[4] Reproducibility of tumor segmentation outcomes with a deep learning model, M Des Ligneris et al., https://hal.science/hal-04006057v1