Ultrasound Slice-to-Volume Registration via Artifact-Suppressed Ultra-Feature Learning

Accurate slice-to-volume registration (SVR) is essential for ultrasound (US)-guided liver interventions. Existing methods typically extract features from slice and volume images to estimate pose parameters. However, the inherent challenges of liver US imaging such as rib shadowing, gastrointestinal artifacts, and poorly visualized vasculature often compromise registration accuracy. To address these limitations, we propose the Artifact Suppressed Ultra-Feature guided SVR (ASUF-SVR), which enhances pose estimation by simultaneously suppressing image artifacts and improving the robustness of feature extraction. The framework integrates two key modules: 1) The Ultra-Feature Extraction (UFE) module, which is a dual-branch design tailored to mitigate the low signal-to noise ratio and low contrast of US images, enabling reliable anatomical feature extraction from both slice and volume data for pose prediction; and 2) the Artifact-Suppressed Evaluation (ASE) module, which supervises similarity measurement and encourages UFE to focus on true anatomical structures rather than artifacts. We validate ASUF SVR on datasets with varying initial offsets and across different organs. Experimental results demonstrate that our method consistently outperforms state-of-the-art baselines both quantitatively and qualitatively. By delivering superior accuracy in SVR, ASUF-SVR minimizes the risk of mis targeting during liver intervention, thereby enhancing over all clinical safety.