UCLA

This dissertation investigates the application of deep learning to clinical ultrasound imaging, progressing from task-specific models for thyroid nodule diagnostics to large-scale self-supervised foundation models for ultrasound. The first part begins with a systematic review of AI methods in the thyroid nodule diagnostic pipeline, identifying problematic gaps in benchmarking rigor and external validation that serve as a driver for following work. Building on those findings, we develop a joint detection and segmentation framework that addresses nodule localization in a multitask model, and ThyGraph, a graph-based architecture for multi-view feature aggregation across an ultrasound image study to improve study-level malignancy classification. These task-specific approaches demonstrate strong performance, but external validation reveals a generalization gap and supervised, single-task training demonstrates representational limits. These observations motivate the second part of the dissertation, which proposes self-supervised pretraining as a path toward more informative ultrasound representations. We develop US-JEPA, a Joint Embedding Predictive Architecture adapted for ultrasound. In parallel, we curate the Open Ultrasound (OPUS) dataset, the largest public ultrasound corpus assembled to date, for model pretraining. The model achieves state-of-the-art performance on various classification tasks, and outperforms competing models in low-label regimes and under test set image corruption. Finally, we introduce UltraBench 2.0, a standardized benchmark and evaluation protocol for ultrasound foundation models. Dataset coverage spans multiple anatomies and clinical tasks, addressing the absence of a common framework that has made meaningful evaluation across models difficult. Collectively, these contributions outline a grounded transition from task-specific deep learning to foundation model development and rigorous benchmarking, establishing both practical tools and conceptual standards for future ultrasound-based AI methods.