Beyond the LUMIR challenge: The pathway to foundational registration models

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Medical image challenges drive innovation and set performance benchmarks.
• Image registration is a fundamental neuroimaging task, advanced by initiatives like Learn2Reg.

Purpose of the Study:

• Introduce the Large-scale Unsupervised Brain MRI Image Registration (LUMIR) challenge, a novel benchmark for unsupervised brain MRI registration.
• Encourage biologically plausible deformation modeling using self-supervision on unlabeled data.

Main Methods:

• Utilized 4,014 unlabeled T1-weighted MRIs for training in a self-supervised manner.
• Evaluated methods on 590 in-domain test subjects and zero-shot tasks across diverse populations, protocols, and species.
• Compared deep learning approaches against optimization-based methods.

Main Results:

• Deep learning methods achieved state-of-the-art performance in unsupervised brain MRI registration.
• Generated anatomically plausible and diffeomorphic deformation fields.
• Demonstrated robustness to domain shifts and outperformed traditional optimization methods.

Conclusions:

• Deep learning is maturing in neuroimaging registration, offering robust and plausible solutions.
• LUMIR serves as a next-generation benchmark for advancing unsupervised medical image registration.
• These findings suggest deep learning's potential as a foundation model for general medical image registration.