GLA-NeRF: global-local aligned neural radiance fields for multi-sweep freehand 3D ultrasound

Objective.Multi-sweep freehand 3D ultrasound is affected by two types of trajectory degradation: inter-sweep rigid bias and intra-sweep local tracking jitter, both of which hinder high-fidelity volumetric reconstruction from tracked 2D ultrasound slices. Although recent neural radiance field methods enable joint optimization of pose and neural representation in optical imaging, they are not directly transferable to ultrasound since limited slice overlap and strong view-dependent speckle of ultrasound make photometric supervision alone ill posed.Approach.This paper presents GLA-NeRF, a global-local aligned neural radiance field framework for joint optimization of pose and neural representation in multi-sweep freehand ultrasound. Within this framework, a unified pose model decouples sweep-wise global rigid bias from frame-wise local tracking jitter, allowing all slices to be mapped into a common canonical space. A global registration module is introduced to correct inter-sweep misalignment through appearance-based image retrieval, hierarchical matching, geometric inlier verification, and uncertainty-weighted Mahalanobis distance. A local denoising module is further incorporated to suppress intra-sweep tracking jitter through statistical sensor noise priors and cumulative Lie-group B-splines, which provide variance-reduced pseudo-labels and physically plausible kinematic constraints.Main results.Experiments on in vivo ultrasound datasets demonstrate that GLA-NeRF stabilizes joint optimization, reduces tracking errors at both global (mm/accuracy regardless of initialization) and local (frommm/tomm/) scales, and improves the anatomical fidelity of synthesized ultrasound views.Significance.Joint optimization of pose and implicit neural representations allow the reconstructed representation to become both more geometrically coherent and more stable for ultrasound novel view synthesis.