Ultrasound Imaging in the Era of Deep Generative Modeling

Ultrasound imaging is a cornerstone of modern clinical diagnostics due to its real-time capability, safety, and portability. Despite its broad utility, from obstetric to cardiovascular imaging and beyond, challenges remain, particularly in maintaining high image quality under physical and anatomical constraints. The modality relies on transmitting acoustic pulses into tissue and reconstructing images from returning echoes. However, accurate interpretation is often hindered by inherent limitations of the imaging system, such as multipath scattering, aberration, and attenuation. Conventional clinical systems typically assume greatly simplified statistical models and decouple acquisition (action) from reconstruction (perception), limiting their ability to exploit spatial and temporal dependencies in the ultrasound data. While executing more transmit events can enhance image quality by better resolving or compounding signals, it comes at the cost of reduced temporal resolution. Recent advances in deep generative models offer a powerful solution to these limitations. Probabilistic inference, diffusion models, and reinforcement learning open new pathways for intelligent, adaptive ultrasound systems that integrate prior knowledge, uncertainty quantification, and data-driven decision-making. This thesis explores how these tools can be used to enhance both the acquisition and interpretation of ultrasound signals. On the perception side, we introduce diffusion-based methods for (1) removing haze in cardiac ultrasound, (2) achieving high volume rate 3D ultrasound imaging by reconstructing high-quality images from heavily undersampled data, and (3) accelerating posterior sampling for real-time imaging. On the action side, we develop adaptive acquisition strategies using reinforcement learning and Bayesian experiment design to optimize transmit patterns in applications such as needle tracking, as well as intravascular and cardiac imaging. As a step toward realization, we present an open-source toolbox that integrates these methods into a unified framework. This toolbox provides a flexible and efficient platform for implementing closed-loop ultrasound imaging, enabling real-time reconstruction, dynamic parameter adjustment, and advanced generative modeling. Together, these contributions demonstrate how deep generative modeling and active sensing can transform ultrasound into a fully adaptive, data-driven modality, laying the groundwork for future cognitive ultrasound systems.