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

TL;DR

The paper tackles the challenge of constructing high-fidelity 3D pelvic anatomy from freehand 2D transvaginal ultrasound without external tracking or pose supervision. It introduces TVGS, an unsupervised Gaussian Splatting framework with a slice-aware differentiable rasterizer that optimizes anisotropic Gaussian primitives directly from 2D slices and jointly refines slice poses. Key contributions include a CUDA-accelerated forward and backward pass, density-based pruning of inactive primitives, and a robust evaluation on synthetic and real TVS data showing competitive accuracy with substantial speed advantages over implicit baselines. The approach offers a scalable, hardware-free pathway toward AI-assisted 3D ultrasound analysis in gynecology, with potential for workflow integration and real-time feedback, while acknowledging limitations in multi-view alignment and physical realism. Overall, the work demonstrates that purely computational reconstruction can achieve accurate 3D US volumes from sparse, uncalibrated data, enabling broader clinical adoption and future data-driven imaging research.

Abstract

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.

Figures (15)

• Figure 1: Top: Simulated sagittal sweep. Bottom: Simulated transversal sweep. The image only shows some of the generated probe rotations.
• Figure 2: Synthetic ultrasound slices generated from a simulated uterus model using a renderer. Left: Sagittal view showing longitudinal anatomy. Right: Transversal view perpendicular to the sagittal plane. These slices can be sampled at arbitrary resolution.
• Figure 3: Two frames sampled from the real data sweeps. On the left, a frame taken from the sagittal view. On the right, a frame taken from the transversal view.
• Figure 4: Overview of our differentiable Gaussian Splatting pipeline for ultrasound volume reconstruction. A set of 3D Gaussians—each defined by a position, shape, intensity, and opacity—is rendered onto input ultrasound slice planes via a custom rasterizer. A reconstruction loss drives gradient-based optimization of all Gaussian parameters. The framework includes strategies for initialization, pruning, and performance optimization.
• Figure 5: Example of a reconstructed 3D view scene after uniform random initialization of means.