Feindel Brain and Mind Series: Making Medical AI Reproducible

From Hype to Hope: Making Medical AI Reproducible

Abstract: Artificial intelligence promises to transform medical imaging, but its impact will remain limited unless research can be made reproducible, transparent, and sustainable. In neuroimaging especially, small datasets, variable annotations, and scanner differences often erode the robustness of AI models. These challenges underscore that lasting progress in medical AI depends less on the next complex architecture, and more on how research practices are designed so that others can reliably reproduce and build upon them.

In this lecture, Julien Cohen-Adad shares lessons learned from developing AI tools for neuroimaging, highlighting both successes and failures. He emphasizes how open standards such as BIDS, interoperable ecosystems like MONAI, and sound data stewardship with tools like git-annex and DataLad provide a foundation for reproducibility at scale. Beyond algorithms, he discusses how reproducibility can be embedded directly into workflows—through transparent data pipelines, automated monitoring of models for drift, and software that remains lightweight, maintainable, and usable by the broader community.

Ultimately, Cohen-Adad argues that reproducibility is not a constraint but a catalyst: it fosters collaboration across disciplines, accelerates translation to the clinic, and ensures that medical AI builds trust by delivering results that can be verified, adapted, and extended. By re-centering efforts on reproducible research and leveraging unique strengths in MR physics and clinical partnerships, he charts a path where AI meaningfully advances both neuroscience and patient care.

Bio: Julien Cohen-Adad is a full Professor at Polytechnique Montreal, Director of the Neuroimaging Functional Unit at the University of Montreal, member of the Mila – Quebec AI Institute, and Canada Research Chair in Quantitative Magnetic Resonance Imaging. Julien Cohen-Adad has a background in MR physics, medical image analysis, A.I. and open-source software development. His research focuses on advancing MRI hardware and analysis methods to help characterizing pathologies in the central nervous system, with a particular focus in the spinal cord. He has published over 200 articles (h-index of 72). Cohen-Adad also dedicates efforts in bringing the community together by developing open source solutions ( https://neuro.polymtl.ca/software.html) and is a strong advocate of open science and reproducible research.

Unable to attend in person? Join us virtually

For more information, visit The Neuro's website