Continuous Bernoulli Distribution for More Realistic Ultrasound Reconstruction with NeRF

Three-dimensional visualization offers a significantly improved depiction of the underlying anatomical structure; however conventional ultrasound typically provides two-dimensional B-mode images of anatomical regions in real time. Recent breakthroughs in computer graphics have leveraged the ability of Neural Radiance Fields (NeRF) to reconstruct three-dimensional scenes implicitly. This is achieved by formulating the three-dimensional scenes as continuous signals using multi-layer perception (MLP). The reconstructed images can be viewed from novel perspectives using a differentiable rendering method. However, current methods cannot recover complex ultrasound anatomical structures and can introduce artifacts, which impact the reconstruction quality. We propose a probabilistic and differentiable approach to estimate backscattered data to provide a more realistic three-dimensional ultrasound reconstruction with NeRF. Our proposed method achieves an improved peak signal-to-noise ratio (PSNR) score of 27.73±0.98 dB compared to the baseline method with 26.31±1.91 dB, along with an improved structural similarity index measure (SSIM) value of 0.47±0.21 to 0.54±0.19 for NeRF.