deep point correspondence estimation and syn-based refinement for multimodal brain image registration

Accurate registration of preoperative magnetic resonance imaging (mri) with intraoperative ultrasound (us) is essential for effective neuronavigation, particularly in brain tumor surgeries where brain shift compromises anatomical fidelity. this study proposes a hybrid framework integrating a deep learning-based multi-layer perceptron (mlp) with an optimization pipeline to enhance mr-to-us registration. the mlp is trained on paired anatomical landmarks extracted from the bite and resect datasets to predict us coordinates from corresponding mri points. an ensemble of five mlps, weighted by inverse validation errors, is employed to estimate dense point correspondences, which are used to initialize an affine transformation. this transformation is refined using symmetric normalization (syn) within the ants registration toolkit to model non-linear deformations. quantitative evaluation demonstrates a mean squared error (mse) of 0.1954 and a mean euclidean distance of 4.97 mm—significantly outperforming a baseline rigid registration approach with 60% improvement in spatial alignment. the proposed pipeline executes in under 4 minutes per case on standard hardware, indicating potential for clinical integration. the results suggest combining learning-based correspondence prediction and classical registration yields accurate and computationally efficient multimodal registration.