High Volume Rate 3-D Ultrasound Reconstruction With Diffusion Models

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Three-dimensional (3D) ultrasound offers real-time volumetric visualization, enhancing accessibility and analysis compared to 2D ultrasound.
• Current 3D ultrasound faces challenges balancing high volume rates with image quality, often requiring compromises in data acquisition.
• Existing methods using diverging waves achieve high rates but degrade image quality due to limited harmonic generation and multipath effects.

Purpose of the Study:

• To introduce a novel approach for 3D ultrasound reconstruction using diffusion models (DMs).
• To improve spatial and temporal resolution in 3D ultrasound by reconstructing volumes from sparsely sampled elevation planes.
• To evaluate the performance of DM-based reconstruction against traditional and deep learning interpolation methods.

Main Methods:

• Employed diffusion models (DMs) for reconstructing 3D ultrasound volumes from a reduced set of elevation planes.
• Compared DM-based reconstruction with traditional and supervised deep learning interpolation techniques.
• Utilized temporal consistency in ultrasound sequences to accelerate inference and explored probabilistic sampling for uncertainty quantification.

Main Results:

• DM-based reconstruction significantly outperformed baseline methods in image quality and downstream task performance on a 3D cardiac ultrasound dataset.
• The proposed method achieved accelerated inference by leveraging temporal consistency.
• Demonstrated robustness and improved recall on out-of-distribution data with synthetic anomalies under strong subsampling.

Conclusions:

• Diffusion models offer a powerful solution for high-resolution 3D ultrasound reconstruction from undersampled data.
• The novel approach enhances image quality and temporal resolution, overcoming limitations of existing 3D ultrasound techniques.
• The method shows promise for more robust and accurate 3D ultrasound imaging, with potential for uncertainty quantification.