From Slices to Structures: Unsupervised 3D Reconstruction of Female Pelvic Anatomy from Freehand Transvaginal Ultrasound

Volumetric ultrasound has the potential to significantly improve diagnostic accuracy and clinical decision-making, yet its widespread adoption remains limited by dependence on specialized hardware and restrictive acquisition protocols. In this work, we present a novel unsupervised framework for reconstructing 3D anatomical structures from freehand 2D transvaginal ultrasound sweeps, without requiring external tracking or learned pose estimators. Our method, TVGS, adapts the principles of Gaussian Splatting to the domain of ultrasound, introducing a slice-aware, differentiable rasterizer tailored to the unique physics and geometry of ultrasound imaging. We model anatomy as a collection of anisotropic 3D Gaussians and optimize their parameters directly from image-level supervision. To ensure robustness against irregular probe motion, we introduce a joint optimization scheme that refines slice poses alongside anatomical structure. The result is a compact, flexible, and memory-efficient volumetric representation that captures anatomical detail with high spatial fidelity. This work demonstrates that accurate 3D reconstruction from 2D ultrasound images can be achieved through purely computational means, offering a scalable alternative to conventional 3D systems and enabling new opportunities for AI-assisted analysis and diagnosis.

💡 Research Summary

The paper introduces TVGS, an unsupervised framework that reconstructs three‑dimensional female pelvic anatomy from freehand transvaginal ultrasound (TVS) sweeps without any external tracking hardware or pre‑trained pose estimators. The core innovation is the adaptation of Gaussian Splatting—a recent real‑time rendering technique that represents a scene with a set of anisotropic 3D Gaussians—to the specific physics and geometry of ultrasound imaging. Because ultrasound slices are thin planar cross‑sections with virtually no overlap, the authors design a slice‑aware differentiable rasterizer that projects each Gaussian onto the slice plane using the slice’s rigid‑body transformation. This rasterizer makes pixel coordinates a differentiable function of the pose parameters, allowing gradients to flow not only to the Gaussian attributes (position, covariance, opacity, intensity) but also to the slice poses themselves. Consequently, the method jointly optimizes anatomical structure and probe motion in a single end‑to‑end pipeline driven solely by image‑level photometric loss and regularization terms (Gaussian pruning, opacity constraints, etc.).

Each Gaussian is defined by a center µ, a covariance Σ factorized into a diagonal scale matrix S and a rotation matrix R derived from a unit quaternion, an opacity o∈