End-to-end deep learning for tuning-free non-contrast ultrasound microvessel imaging

Microvessel (MV) imaging using non-contrast US is challenging due to low blood echo SNR, the presence of spectrally overlapping tissue clutter, and noise. Spatiotemporal clutter removal, e.g., via singular value decomposition (SVD), requires parameter tuning and are computationally expensive. In this work, we proposed an end-to-end deep learning approach to remove tissue clutter and extract blood signals from US ensemble in a purely data-driven fashion. Our deep learning approach accelerates the decluttering process and provides a generalizable framework to adapt to realistic scenarios via diversly augmented data. In this preliminary study, we scanned a day 8 developing chicken embryo to collect US ensembles of different regions of the embryo. Training data was created using a fine-tuned SVD for tissue clutter removal and to train a 2D U-net. The proposed deep learning approach can reconstruct MV from US ensembles (structural similarity index of 0.83 ± 0.02 compared to processing via SVD) implying good reconstruction. Our end-to-end deep learning approach is also 2x faster than the traditional SVD approach. Deep learning based MV imaging using non-contrast US shows promise in reconstructing MV in presence of different tissue motions and can be extended to perform under various SNR, noise, and motion conditions.