Multi-Plane Wave Signal Inpainting with CNNs: A Framework for Reducing RF Data Volume in Ultrafast Ultrasound

Reducing radio frequency data volume prior to beamforming is crucial for enabling efficient 3D imaging and ultraportable devices. In previous work, we introduced a convolutional neural network (CNN)-based inpainting method designed to estimate missing radio frequency (RF) data from single plane waves (PWs) acquired with sparse arrays in reception. Motivated by the superior image quality achievable through compounding multiple PWs, we extend the inpainting method for multi-PW acquisitions. To ensure fair comparisons with single-PW approaches, we maintain a consistent data volume by increasing the downsampling factor (DF) proportionally to the number of PWs. We evaluate two downsampling strategies on raw signals from five PW acquisitions: uniform transducer element selection and random spatio-temporal masking. Results show that random masking significantly outperforms uniform downsampling in image quality metrics and visual assessment, though neither fully replicates the quality of the target images. To benchmark performance, we compare the extended multi-PW method to our original single-PW inpainting approach with sparse arrays. Our findings reveal that the CNN-based inpainting method is more effective on single PW acquisitions with low downsampling factors than on multi-PW acquisitions with higher downsampling factors. These results underscore the potential of single-PW strategies for resource-constrained applications, offering reduced computational complexity while preserving high-quality image reconstruction.