Weakly-Supervised 2D-3D Cardiac Ultrasound Registration with ROI Segmentation and Test-Time Optimization

2D-3D medical image registration plays a critical role in cardiac image-guided interventions by aligning preoperative 3D volumes with intraoperative 2D images to provide more comprehensive anatomical context. However, the task remains highly challenging due to large modality and dimensional gaps, low-quality ultrasound images, and the lack of sufficient supervisory signals. In this paper, we propose WSCU-Reg, a segmentation-guided optimization framework for Weakly Supervised 2D-3D Cardiac Ultrasound Registration. WSCU-Reg is a hybrid approach that integrates weakly supervised anatomical segmentation with optimization-based refinement to improve the accuracy of cardiac ultrasound registration. WSCU-Reg builds upon the original CU-Reg framework by incorporating an additional segmentation-guided branch and post-prediction optimization. Our method first extracts features from 2D images, 3D volumes, and corresponding segmentation masks, which are then processed by an attention-based pose predictor to estimate the rigid transformation. To enhance alignment accuracy, we introduce a weakly supervised segmentation branch to predict the anatomical mask of the resampled 2D slice. A segmentation loss is computed between the predicted and ground truth masks and used along with the image similarity loss. During the inference stage, the predicted transformation is further refined through a segmentation-driven optimization step to improve the alignment quality. Experimental results on a cardiac ultrasound dataset demonstrate that our method improves registration performance. Compared to the CU-Reg, the distance error is reduced from 3.91mm to 2.87mm, and NCC(Normalized cross-correlation) of ground-truth slice and aligned slice improves from 88.07% to 90.17%, confirming the effectiveness of the proposed segmentation-aware refinement baseline.