NoGL-Net: a Deep Learning Approach to Grating Lobe Rejection Using Sparse RF Channel Interpolation in Synthetic Radial Aperture Focusing

Despite the great potential of high-frame-rate ultrasound imaging, image quality is often degraded by grating lobe (GL) artifacts due to sparse spatial sampling. While denser spatial sampling can alleviate these artifacts, it is often not allowed due to array and scanning configurations dictated by clinical applications. Moreover, it requires increased data transmission and storage demands, computational cost, and challenges for real-time imaging. We propose NoGL-Net, an encoder-decoder network optimized for GL suppression in synthetic radial aperture focusing. Specifically, we perform a comparative study between the networks trained for channel interpolation using raw RF or delay-compensated RF data. In proof-of-concept studies, NoGL-Net effectively reduced artifacts from doubled array pitch, with delay-compensated training achieving the best performance (SSIM = 0.95), outperforming raw RF training (SSIM = 0.74). These results demonstrate the feasibility of deep-learning-based interpolation for GL suppression and highlight the added benefit of delay compensation in improving channel interpolation and image reconstruction.