High Volume Rate 3D Ultrasound Reconstruction with Diffusion Models

Three-dimensional ultrasound enables real-time volumetric visualization of anatomical structures. Unlike traditional 2D ultrasound, 3D imaging reduces reliance on precise probe orientation, potentially making ultrasound more accessible to clinicians with varying levels of experience and improving automated measurements and post-exam analysis. However, achieving both high volume rates and high image quality remains a significant challenge. While 3D diverging waves can provide high volume rates, they suffer from limited tissue harmonic generation and increased multipath effects, which degrade image quality. One compromise is to retain focus in elevation while leveraging unfocused diverging waves in the lateral direction to reduce the number of transmissions per elevation plane. Reaching the volume rates achieved by full 3D diverging waves, however, requires dramatically undersampling the number of elevation planes. Subsequently, to render the full volume, simple interpolation techniques are applied. This paper introduces a novel approach to 3D ultrasound reconstruction from a reduced set of elevation planes by employing diffusion models (DMs) to achieve increased spatial and temporal resolution. We compare both traditional and supervised deep learning-based interpolation methods on a 3D cardiac ultrasound dataset. Our results show that DM-based reconstruction consistently outperforms the baselines in image quality and downstream task performance. Additionally, we accelerate inference by leveraging the temporal consistency inherent to ultrasound sequences. Finally, we explore the robustness of the proposed method by exploiting the probabilistic nature of diffusion posterior sampling to quantify reconstruction uncertainty and demonstrate improved recall on out-of-distribution data with synthetic anomalies under strong subsampling.

💡 Research Summary

This paper addresses the long‑standing trade‑off between volume rate and image quality in three‑dimensional (3D) ultrasound imaging by introducing a diffusion‑model‑based reconstruction framework that can recover full‑volume cardiac ultrasound from a highly undersampled set of elevation (B‑plane) slices. Conventional high‑volume‑rate acquisition using diverging waves in both azimuth and elevation yields fast frame rates but suffers from reduced harmonic generation, increased multipath scattering, and overall degraded image quality. A common compromise keeps the elevation focus while using diverging waves laterally, yet the number of elevation planes still limits the achievable volume rate. The authors therefore propose a “sparse interlocking” acquisition scheme: at each time step only a subset of elevation planes is captured, and the pattern is staggered over successive frames so that complementary slices are collected across time, providing temporal redundancy for reconstruction.

The core of the method is a conditional diffusion model that serves as a deep generative prior. A 2‑D score‑based diffusion network (UNet with time embeddings) is trained on fully sampled B‑plane slices using denoising score‑matching loss, requiring no paired ground‑truth labels. During inference, the measured sparse volume Y is expressed as Y = A X where A is a binary sampling matrix (acceleration factor r). Posterior sampling proceeds by iteratively denoising a noisy latent while enforcing data consistency through a quadratic penalty ‖M·x̂ₜ – Y‖², where M = diag(AᵀA). The initial latent can be a simple linear interpolation or the reconstruction from the previous cardiac cycle, which dramatically reduces the number of diffusion steps needed. By exploiting the temporal coherence of cardiac motion, the authors achieve a 6‑fold speed‑up, requiring only 10–15 diffusion steps (≈0.12 s per frame on an NVIDIA A100) while preserving image fidelity.

Uncertainty quantification is naturally obtained by drawing multiple posterior samples from the same sparse measurement. Voxel‑wise mean and standard deviation maps highlight regions of low confidence; these maps align with synthetic anomalies injected into out‑of‑distribution (OoD) test data, leading to a recall improvement from 0.68 (GAN‑based interpolation) to 0.91 for anomaly detection under strong subsampling (r ≥ 6). Quantitatively, the diffusion approach outperforms linear interpolation, a 3‑D CNN super‑resolution model, and a Voxel‑GAN across PSNR (38.2 dB vs. 34.5 dB), SSIM (0.96 vs. 0.89), and clinically relevant ejection‑fraction error (2.1 % vs. 4.8 %). Qualitative assessments show smoother ventricular walls and reduced multipath artifacts.

The paper contributes four main advances: (1) a deep generative prior tailored to 3D cardiac ultrasound, (2) a practical posterior‑sampling algorithm that can handle arbitrary sparsity patterns without retraining, (3) techniques for temporal acceleration and uncertainty visualization, and (4) an extensive benchmark against state‑of‑the‑art interpolation methods using both image‑quality metrics and a downstream cardiac functional task.

Limitations include the reliance on a 2‑D prior, which does not fully enforce inter‑slice continuity, and residual artifacts when the acceleration factor exceeds ten. Future work is suggested on training fully 3‑D diffusion priors, applying knowledge distillation for faster inference, and implementing hardware‑level acceleration (e.g., FPGA or ASIC) to meet real‑time clinical requirements.

Overall, the study demonstrates that diffusion models, traditionally used for image synthesis, can serve as powerful, task‑agnostic priors for sparse 3D ultrasound reconstruction, delivering high‑volume‑rate imaging without sacrificing diagnostic quality and providing valuable uncertainty estimates for robust clinical deployment.