A Deep Reinforcement Learning Based Region-Specific Beamformer for Sparse Arrays 3D Ultrasound Imaging.

Abstract

Sparse arrays offer several advantages over other element reduction techniques for 3D ultrasound imaging. However, the large inter-element spacing in these arrays results in high sidelobe-related artifacts, which significantly degrade image quality and limit their application in 3D ultrasound imaging. Adaptive beamformers have been proposed to mitigate sidelobe-related artifacts, but they often degrade speckle texture quality, resulting in unnaturally dark images. To overcome these limitations, we propose RSB-Net, a region-specific beamformer based on deep reinforcement learning. RSB-Net adaptively selects the most suitable beamformer for each pixel of the image, applying adaptive beamforming in regions dominated by sidelobe artifacts and delay-and-sum beamforming in regions where speckle texture should be preserved. The effectiveness of RSB-Net was validated on both simulated and experimental synthetic transmit aperture RF datasets with a new designed sparse array prototype. On simulated data, RSB-Net achieved significant gains, with improvements of 52.81 dB in contrast ratio and 0.65 in generalized contrast-to-noise ratio compared to DAS beamforming. In experimental tissue-mimicking phantom data, RSB-Net demonstrated similar performance, achieving gains of 51.01 dB and 0.64 respectively. These results highlight the potential of RSB-Net as a robust and effective solution for high-quality B-mode 3D ultrasound imaging using 2D sparse arrays, advancing the standardization of 3D ultrasound in clinical settings by enhancing anatomical visualization, reducing operator dependency, and improving measurement accuracy for lesions and calcifications.