Medical imaging research is increasingly built on complex computational pipelines: data curation and preprocessing, model training (often at scale), hyperparameter optimization, evaluation under shift, and downstream clinical translation. While these advances have accelerated performance, they have also amplified persistent reproducibility challenges: (i) incomplete reporting of experimental details (data splits, preprocessing, augmentation, and training schedules), (ii) non-portable software environments and dependency drift, (iii) hidden sources of randomness and evaluation leakage, (iv) restricted data access and privacy constraints, and (v) fragile benchmark comparisons that do not generalize across sites, scanners, and protocols.
The MICCAI community has already made important progress through challenge-driven benchmarking, where standardized datasets, metrics, and centralized evaluation workflows can substantially improve comparability and reproducibility. Yet, most MICCAI papers still “live” outside challenge settings: they rely on custom pipelines, partially specified training/evaluation details, and frequently on private or institution-specific datasets. As a result, even when methods show strong reported performance, they can be difficult to reproduce independently and hard to maintain as usable scientific artifacts over time.
At MICCAI—a uniquely interdisciplinary venue connecting medical image computing (MIC) and computer-assisted intervention (CAI)—there is a timely opportunity to establish domain-specific reproducibility practices that respect clinical constraints while raising scientific rigor and long-term reusability. R2MI 2026 is designed to complement and extend challenge-based reproducibility by promoting challenge-independent best practices, tools, and policies that can be adopted across the broader body of MICCAI research.
R2MI 2026 aims to provide a structured discussion forum and a practical, actionable outcome for the MICCAI community. The workshop objectives are to: