A Guarded Hybrid Pipeline for Rigid FLAIR–T1 MRI Registration: Learned Initialisation of Iterative Mutual-Information Refinement

Background: Rigid alignment of multimodal MRI is a prerequisite for neurological post-processing, yet classical iterative optimisers are accurate only when allowed to converge, at tens of seconds per volume, while learning-based predictors are fast but far less precise. This study asks whether a learned predictor can serve as the initialisation of an iterative optimiser so that both speed and accuracy are obtained. Methods: A supervised dual-path convolutional network with a cross-modal attention block (AttentionReg) was trained on BraTS 2020 under a subject-level split and a synthetic rigid-transformation protocol. Its prediction initialises a brain-masked Mattes Mutual-Information optimiser, forming a guarded hybrid pipeline in which a refinement is accepted only when it improves the metric. Eight methods were compared under one protocol on a common subset of 990 held-out pairs from 99 subjects, using a landmark-based Target Registration Error expressed in physical millimetres. Results: The proposed hybrid attained a median TRE of 0.81 mm over the evaluated pairs (IQR 0.57–1.12; P95 1.67 mm; 0.3% above 3 mm) at 0.79 s per volume, against 0.57 mm at 25.3 s for the fully converged iterative reference. The subject-level paired difference was 0.23 mm (95% CI 0.19–0.28), statistically significant yet equivalent within a 0.50 mm margin, obtained with a 32-fold reduction in computation time. Used alone, the network reached 3.10 mm; iterative refinement improved it by a mean subject-level paired reduction of 2.49 mm (95% CI 2.32–2.67). An ablation showed the attention block does not change median accuracy relative to a matched network without it. Conclusions: A learned initialiser converts an accurate but slow iterative method into an accurate and fast one, providing a proof of concept for real-time rigid multimodal MRI registration.