UltraGauss: Ultrafast Gaussian Reconstruction of 3D Ultrasound Volumes

TL;DR

UltraGauss addresses the challenge of turning 2D ultrasound data into accurate 3D volumes by adopting an ultrasound-specific Gaussian Splatting framework that models probe-plane intersections rather than optical ray casting. It introduces an efficient, numerically stable covariance parameterization and a two-phase GPU rasterization strategy to accelerate reconstruction, achieving state-of-the-art quality within minutes on a single GPU. Clinician surveys demonstrate that UltraGauss reconstructions are more realistic than competing methods, underscoring its potential to standardize and speed up fetal ultrasound interpretation. The approach supports end-to-end clinical pipelines from freehand cine-sweeps to volumetric reconstructions, with broad implications for clinical workflow and diagnostic research.

Abstract

Ultrasound imaging is widely used due to its safety, affordability, and real-time capabilities, but its 2D interpretation is highly operator-dependent, leading to variability and increased cognitive demand. 2D-to-3D reconstruction mitigates these challenges by providing standardized volumetric views, yet existing methods are often computationally expensive, memory-intensive, or incompatible with ultrasound physics. We introduce UltraGauss: the first ultrasound-specific Gaussian Splatting framework, extending view synthesis techniques to ultrasound wave propagation. Unlike conventional perspective-based splatting, UltraGauss models probe-plane intersections in 3D, aligning with acoustic image formation. We derive an efficient rasterization boundary formulation for GPU parallelization and introduce a numerically stable covariance parametrization, improving computational efficiency and reconstruction accuracy. On real clinical ultrasound data, UltraGauss achieves state-of-the-art reconstructions in 5 minutes, and reaching 0.99 SSIM within 20 minutes on a single GPU. A survey of expert clinicians confirms UltraGauss' reconstructions are the most realistic among competing methods. Our CUDA implementation will be released upon publication.

Figures (13)

• Figure 1: Input images of the fetal brain are acquired (blue Input Video frames) and their poses estimated (blue probes/planes). UltraGauss then uses these to form a total 3D reconstruction within 5 minutes. A few of the 3D gaussians are shown as grey ellipses. Cross-sectional views at previously unseen poses can then be sampled from the reconstructed 3D volume to form a complete cinesweep or be viewed individually. These can be seen in the green Output Video cinefilm frames.
• Figure 2: Reconstruction results for 5 different models (ours in bold) applied on 3 different training datasets, and shown at 3 time points over the duration of reconstruction. Higher SSIM is better. Errors bars show $\pm 1$ standard deviation amongst fetuses
• Figure 3: Illustration of the occlusion mechanism in gaussian splatting, for images in the visual spectrum. The rendering equation (\ref{['eq:rgb-render']}) accumulates transmittance $T$ over a pixel's ray, and this is used to model occluding gaussians (in this example, the blue gaussian occludes the yellow one).
• Figure 4: Illustration of the proposed gaussian rendering mechanism for ultrasound (\ref{['sec:image-formation']}). Instead of raycasting, the image formation model is based on the intersection of the gaussians with the probe plane (which maps 1:1 to the captured ultrasound pixel grid). This means that occlusions do not play as large a role as for the visual spectrum (\ref{['fig:intersection']}), and so are not modeled.
• Figure 5: Side view of the gaussian bounding box intersection and projection (\ref{['sec:boundaries']}). Boxes that do not intersect with the probe plane are rejected early (\ref{['sec:load-balance']}). Those that intersect it are rasterized only within the intersecting 2D bounding box. In this example, pixels that are not iterated over are marked with crosses.