Abstract
Neural radiance field (NeRF) has been proposed for 3D ultrasound reconstruction and has shown promising results. However, for freehand 3D freehand ultrasound, due to its reliance on human operators, i.e., sonographers or clinicians, can limit the use of NeRF. Since NeRF requires capture of multiple viewing angles of the underlying scene, different exerted pressure by the transducer during freehand scanning can cause pixel mismatches at the same location for different sweep sequences where NeRF requires highly accurate location information to learn the 3D representations. We first demonstrate this mismatch issue with freehand ultrasound scanning. We then propose to utilize a modified positional encoding method and demonstrate its utility on US freehand scans captured on a phantom. Quantitative and qualitative results show that our proposed method enables more consistent volume compounding for new viewpoints. Compared to the baseline, our proposed method achieves significantly improved results for both peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) metrics, with maximum improvements of 32.96% and 51.22%, respectively.