Physics-Guided Learning for Ultrasound Image Reconstruction and Inverse Problems

Ultrasound imaging provides real-time, safe, and portable visualization of biological tissue, but its performance depends strongly on how acoustic measurements are converted into images and quantitative tissue properties. Conventional reconstruction methods often rely on simplified assumptions about wave propagation, particularly a constant Speed of Sound (SoS), which can limit image quality and quantitative accuracy. This dissertation develops physics-guided learning methods for ultrasound image reconstruction and inverse problems, progressing from learned enhancement of conventional beamforming to adaptive estimation of acoustic properties and, ultimately, physics-constrained quantitative reconstruction.

The first part of the dissertation focuses on pulse-echo imaging, beginning with a learning-based framework for reconstructing high-quality images from a single plane-wave transmission. By combining radio-frequency data normalization across transmission angles with learned pixel weighting and image enhancement, the method improves upon conventional single-plane-wave beamforming while preserving the temporal advantages of ultrafast imaging and demonstrating generalization to independently acquired datasets. Next, the constant-SoS assumption is relaxed through an adaptive beamforming framework that learns spatially varying SoS distributions and channel apodization weights directly within the reconstruction process. Without requiring ground-truth SoS maps or multi-plane-wave reference images, the method adapts beamforming delays and receive weighting to the underlying propagation conditions and achieves image quality comparable to multi-angle coherent compounding with substantially reduced computational cost. Building on this, a self-supervised framework is developed to directly estimate spatial SoS from full-transmit pulse-echo measurements. By coupling differentiable beamforming with an acoustic forward model and enforcing consistency in both the image and radio-frequency domains, the approach extends SoS from an adaptive reconstruction parameter to a quantitative tissue property recovered directly from the measured data.

The second part addresses quantitative SoS reconstruction using ultrasound computed tomography. A physics-constrained inverse learning framework couples an inverse neural network with a differentiable wave-propagation model. This formulation incorporates acoustic physics directly into the inverse process and enables reconstruction of spatial SoS distributions without paired ground-truth supervision. Evaluation on simulated phantoms, including anatomically realistic, simulated breast phantoms, demonstrates improved reconstruction performance relative to conventional full-waveform inversion and supervised learning methods. The proposed framework also achieves more than two orders of magnitude computational speedup compared with full-waveform inversion. In addition, it requires substantially less training data than supervised methods and remains robust to measurement noise.

The work presented in this dissertation shows how ultrasound reconstruction can progress from learned enhancement of conventional beamforming, through physics-adaptive image formation and quantitative estimation of acoustic properties, to physics-constrained solution of the acoustic inverse problem. The dissertation demonstrates how integrating learning with acoustic physics can improve the speed, accuracy, data efficiency, and generalizability of ultrasound reconstruction across both pulse-echo imaging and ultrasound computed tomography.

 

Share this event:

facebook linked in twitter email

Media Contact: Emmalia Lutz