Abstract
Deep learning-based medical image registration often suffers from poor generalization across different modalities and domains. This paper presents a robust strategy developed for the LUMIR brain registration challenge, utilizing the Modality Independent Neighborhood Descriptor (MIND) as an input feature space rather than a loss function. By combining a symmetric, diffeomorphic backbone (SITReg) with MIND-transformed inputs and an ensemble strategy, the proposed method achieves state-of-the-art performance, particularly in multi-modal (T1 to T2) registration, without compromising in-domain accuracy.
Introduction
Deep learning has revolutionized medical image registration, yet these models remain notoriously "brittle." A model trained to align T1-weighted MRIs often fails when presented with T2-weighted scans or data from different manufacturers—a phenomenon known as domain shift. The LUMIR challenge targets this issue by evaluating models on their ability to register images with varying contrasts and modalities. This paper proposes a method to move away from overfitted models toward robust ones that generalize across diverse clinical scenarios.
Methods
The proposed approach relies on a combination of a strong architectural backbone and a novel preprocessing strategy.
The SITReg Backbone
The authors utilize SITReg, a symmetric and diffeomorphic architecture. Symmetry ensures that the registration is independent of which image is designated as "fixed" or "moving," while diffeomorphism guarantees smooth, physically plausible deformations, preventing digital tissue "tears."
MIND as Input Representation
The core innovation is the use of MIND (Modality Independent Neighborhood Descriptor) features as the primary input to the network. Unlike raw pixel intensities, MIND describes local structural relationships. Since anatomy remains constant even when image contrast changes, MIND features are highly consistent across T1 and T2 scans. By "stripping away the paint" (modality) and leaving only the "architecture" (anatomy), the network focuses on structural alignment.
Training and Ensembling
During training, the network processes MIND features but is optimized using a standard intra-modality loss (Normalized Cross-Correlation) on the original images. This forces the model to learn structural alignment while ignoring modality-specific noise. Finally, an ensemble of five models is used, with predictions averaged in the B-spline weight space to maintain physical plausibility.
Results
The impact of the MIND transformation was most dramatic in multi-modal scenarios. As shown in the ablation study, the Dice overlap score for T1-to-T2 registration jumped from a failing 28.4% in the baseline to a robust 74.5% with the full strategy.
Crucially, this gain in robustness did not come at the cost of specialized performance; in-domain T1-to-T1 accuracy remained high at approximately 78%. The method also significantly reduced the "non-diffeomorphic volume" (NDV), indicating fewer unrealistic image folds. At the time of submission, this strategy held 1st place on the LUMIR 2025 challenge leaderboard.
Discussion and Conclusion
The authors demonstrate that using structural descriptors like MIND as input features, rather than just as loss functions, provides a powerful mechanism for domain generalization. While a small performance gap remains between intra-modal and multi-modal registration (78% vs 74.5% Dice), the stability and robustness offered by this approach are significant. Future work may explore even more sophisticated unified representations and test the strategy on other anatomical regions and modalities like CT scans.
Summary updated