RapidVol: Rapid Reconstruction of 3D Ultrasound Volumes from Sensorless 2D Scans

Two-dimensional (2D) freehand ultrasonography is a widely used medical imaging modality, particularly in obstetrics and gynaecology. However, it only captures 2D cross-sectional views of inherently 3D anatomies, losing valuable contextual information. As an alternative to costly 3D ultrasound (US) scanners, 3D volumes can be artificially reconstructed from 2D scans, but this is usually prohibitively slow. Hence, we propose RapidVol: a neural representation framework to speed up slice-to-volume US reconstruction. We use tensor-rank decomposition to decompose the typical 3D volume into tri-planes, which are stored alongside a small neural network. With a set of 2D US scans and their estimated 3D orientation, RapidVol can achieve complete 3D reconstruction. To evaluate our method, we form reconstructions from real fetal brain scans, and then request novel cross-sectional views. Compared to prior fully implicit (e.g. neural radiance field) approaches, our method is over 3x quicker, 46% more accurate, and more robust to errors in pose estimation. We also demonstrate that further speed-up is achievable by reconstructing from a structural prior rather than from random initialisation.