Geometry Aware Neural Radiance Fields for Freehand Ultrasound Reconstruction

Le résumé fourni par la source

Many freehand 3D ultrasound methods map 2D images directly into a 3D volume, but they are sensitive to transducer pose errors, leading to poor reconstructions. Implicit representation methods, such as Neural Radiance Fields (NeRF), offer an alternative by modeling the underlying 3D scene as a continuous volumetric function, learned from 2D images. While NeRF has shown promise in other imaging domains, its application to ultrasound (US) remains limited due to its difficulty in refining transducer poses, which leads to severe reconstruction artifacts. To address this challenge, we propose Geometric-Aware Ultrasound NeRF (GAU-NeRF), a novel framework based on Bundle Adjusting NeRF (BARF). GAU-NeRF introduces a regularization strategy during the joint optimization of the NeRF representation and transducer poses, leading to more accurate pose refinement and improved reconstruction quality. Our method significantly outperforms existing baselines on both simulated and in vivo US datasets, achieving substantial gains across multiple metrics, including up to 132% increase in peak signal-to-noise ratio (PSNR), 133% improvement in structural similarity index measure (SSIM), and 350% reduction in learned perceptual image patch similarity (LPIPS).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.