Ultrasound imaging is a crucial tool for guiding radiation therapy, particularly for cancers such as pancreatic cancer, where tumors exhibit respiration-induced motion. While flexible ultrasound transducers offer improved anatomical conformity and reduced compression-induced distortion compared to rigid probes, their variable geometry presents significant challenges for conventional beamforming. In this study, we investigate a deep learning-based beamforming framework that directly predicts delayed RF data from raw RF input, bypassing explicit transducer shape estimation and traditional delay-and-sum computations. Building upon an artificial curvature simulation strategy, we systematically analyze the impact of curvature-induced variation and inherent RF noise on model performance and generalizability. We further introduce frequency-domain analysis to quantify RF-level signal variation that may not be apparent in spatial-domain image comparisons. Our results demonstrate that although noise-augmented training improves prediction consistency, reconstruction performance remains limited under the current prototype noise conditions. These findings highlight the importance of RF data diversity and noise characterization in developing clinically robust deep learning beamformers for flexible transducer-based ultrasound-guided radiation therapy.
Feng et al. (Sun,) studied this question.