From geometry to intensity: a coarse-to-fine pipeline for unsupervised 3D ultrasound stitching

Three-dimensional ultrasound (3DUS) stitching aims to expand the field-of-view (FOV) by registering partially overlapping 3DUS volumes acquired from different probe positions. This task is highly challenging due to several intrinsic characteristics of 3DUS imaging, such as sector-shaped FOVs, low image quality, substantial noise, and imaging artifacts, as well as large inter-volume motion. These factors make it difficult for conventional registration methods to accurately identify both global and local geometrical correspondences between image pairs. To address these challenges, we propose a simple yet effective unsupervised pipeline for 3DUS stitching, which consists of three main stages: (1) unsupervised geometric feature extraction, (2) point cloud (PCD)-based coarse registration, and (3) intensity-based fine registration. Extensive quantitative and qualitative evaluations on a 3DUS placenta dataset demonstrate that our method significantly outperforms existing state-of-the-art 3DUS registration frameworks in both accuracy and robustness.