UltraGauss: Ultrafast Gaussian Reconstruction of 3D Ultrasound Volumes
Mark C. Eid, Ana I.L. Namburete, Jo\~ao F. Henriques
TL;DR
UltraGauss is a novel, ultrasound-specific Gaussian Splatting framework that enables fast, accurate 3D ultrasound volume reconstruction, improving clinical visualization with real-time performance on standard hardware.
Contribution
It introduces the first ultrasound-specific Gaussian Splatting method, extending view synthesis to ultrasound physics and optimizing computational efficiency and accuracy.
Findings
Achieves state-of-the-art 3D ultrasound reconstructions in 5 minutes.
Reaches 0.99 SSIM within 20 minutes on a single GPU.
Clinicians find UltraGauss reconstructions most realistic.
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…
Peer Reviews
Decision·ICLR 2026 Poster
- The rendering strategy is thoughtfully designed based on the principles of ultrasound imaging. - The paper provides a thorough discussion of related work. - It clearly introduces the background of 3D Gaussian Splatting and ultrasound imaging to aid understanding. - The datasets and experimental settings are described in detail, including a clinician survey for evaluation.
1. **Further clarification on efficiency**: UltraGauss is described as "efficient" in terms of both memory and computation time (lines 013, 044, 107, 122, 211， 477， etc.). However, this claim is not supported by any quantitative results in the current version of the paper. - Although the time consumption of various methods is presented in Fig. 4, Fig. B1 (Appendix), and Fig. D3 (Appendix), these results do not clearly support the claimed efficiency of UltraGauss. 2. **Clarification for the prop
1. The χ² ellipsoidal bounding and two-phase GPU pipeline are clearly motivated and provide a practical path to minute-level reconstructions with up to ~2M Gaussians. The exposition around Fig. 2–3 and Sec. 4.3–4.4 is crisp. 2. The lower-triangular precision factorization (Σ⁻¹=LLᵀ with positive diagonals) is a sensible alternative to quaternion-scaled covariances in this setting; it simplifies inversion and Gaussian resampling and appears faster in ablations. 3. Beyond 3D volume metrics, the r
1. The Beer–Lambert shadow term is extremely lightweight; while speed is compelling, the paper concedes deficits in speckle and multiple-scattering effects, where physics-informed baselines sometimes produce more realistic artifacts. The method’s robustness when probe geometry is imperfect is discussed qualitatively but not deeply quantified. 2. Core quantitative results hinge on 12 fetal brain volumes (Dataset A) and 3 freehand videos (Dataset B). That is narrow in anatomy, vendor, and probe d
The paper shows several notable strengths. First, its novelty is good — it is the first to adapt 3D Gaussian Splatting to ultrasound reconstruction, proposing a probe–plane intersection rendering mechanism that better captures ultrasound’s plane-based sampling and attenuation behavior. This represents a well-motivated and technically elegant reformulation of Gaussian splatting aligned with acoustic image formation. Second, the performance is solid and the experiments are very comprehensive. Th
(1) The paper lacks sufficient discussion and comparison with two highly related works, X-Gaussian and R2Gaussian. In particular, its Beer–Lambert–based attenuation formulation is conceptually similar to the radiative attenuation modeling in R2Gaussian, yet this connection is not explicitly acknowledged. The writing style and overall pipeline design also bear strong resemblance to these two prior works. Without a detailed discussion or quantitative comparison against them, the novelty claim of U
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsUltrasound Imaging and Elastography · Ultrasound and Hyperthermia Applications · Microwave Imaging and Scattering Analysis