Regscore: A Foundation Model-Driven Universal Metric for Multi-Modality Medical Image Registration
Abstract
Purpose
Accurate evaluation of multi-modality deformable image registration remains a critical challenge in radiotherapy. Traditional metrics such as Normalized Mutual Information (NMI), Mean Absolute Error (MAE), and Normalized Cross-Correlation (NCC) are based on intensity similarity measurement, lacking anatomical structure understanding, and their effectiveness is limited in cross-modality tasks where intensity correspondence does not exist. We propose RegScore, a universal registration evaluation metric based on 3D cross-modality landmark matching, leveraging foundation models to provide anatomy-based evaluation robust to intensity variations across modalities.
Methods
RegScore evaluates registration quality by automatically matching cross-modality anatomical landmarks. We extended the 2D image-matching foundation model MINIMA to 3D, by fine-tuning it on 200 multi-modality patient cases (CT, T1, T2, FLAIR). The framework includes: (1)a pre-trained SuperPoint network that extracts slice-wise landmarks along with their local anatomical descriptors for subsequent matching; (2)3D position encoding that provides spatial position information, complementing anatomical descriptors by incorporating positional plausibility; and (3)a Transformer network performs cross-modality landmark matching via information extracted above. RegScore quantifies registration quality via confidence-weighted distances between matched landmark pairs. To compare different similarity metrics, we applied incrementally increasing known deformations to initially registered image pairs and computed the Spearman correlation between each metric and deformation magnitude, where values closer to 1 indicates better sensitivity to registration errors. We evaluated eight cross-modality registration tasks using images outside the training set.
Results
RegScore achieved a mean(±s.d.) Spearman correlation coefficient of 0.9979±0.0991 across eight cross-modality registration tasks, outperforming traditional metrics including NCC (0.7604±0.5437), NMI (0.8188±0.1859), and MIND (0.8133±0.3245). In an abdomen T1-T2 registration task, RegScore also substantially outperformed the segmentation-based Dice coefficient metric (1.0000 vs. 0.4371).
Conclusion
RegScore provides a universal, anatomy-based metric for evaluating multi-modality registration with exceptional accuracy and robustness. This advancement enables reliable quality assessment for cross-modality registration, facilitating adaptive planning, dose accumulation, and image-guided treatment workflows.