Addressing Domain Shifts Across Sparse Array Strategies for Deep Learning Inpainting in Ultrafast Ultrasound Imaging

Deep learning shows great promise in enhancing ultrasound acquisitions using sparse arrays. We previously introduced a deep learning-based in-painting method to estimate missing raw radio frequency signals in single plane wave acquisitions with sparse arrays in reception. This study evaluates its robustness with fully sparse arrays, where both transmission and reception are undersampled. We compare three training strategies: (i) training from scratch on reception-only sparse array data, (ii) training from scratch on fully sparse array data, and (iii) fine-tuning a model pretrained on reception-only sparse array data using a small fully sparse array dataset. For fully sparse arrays with transducer elements reduced by factors of 2 and 4, all models reduce artifacts and improve image quality. However, models trained solely on reception-only sparse array data yield lower image quality metrics. Fine-tuning with a small fully sparse array dataset achieves quality metrics closer to those obtained by training from scratch on a large fully sparse array dataset. These results highlight the importance of dataset selection and fine-tuning to mitigate domain shifts, offering an effective strategy for improving ultrafast ultrasound acquisitions with sparse arrays in both transmission and reception.