Implicit neural representation for scalable 3D reconstruction from sparse ultrasound images

Although volumetric ultrasound is limited by cost and availability of 2D arrays, 3D volumes can be reconstructed from 2D slices if transducer position is known, which is not usually the case. Even with position data, existing algorithms for reconstruction are impractical due to their discrete nature that struggles with scale. We propose a 1D array on a programmable motor for scanning and implicit neural representations for continuous reconstruction. Our network's ability to sample at arbitrary positions was compared to classic algorithms, achieving x7.9 performance while maintaining accuracy. Based on these, a reconstruction pipeline was tested on simulated data with 93% accuracy using only 36 B-mode images. This was evaluated in-vivo to measure tumor volumes in mice, with 6.3% mean error. Our findings suggest implicit neural representations can reduce data needed to recreate volumes from 2D slices and replace interpolation methods to enable interactive analysis.