On the effect of training simulation point spread function in deep-learning based super-resolution ultrasound imaging

Deep learning (DL)-based super-resolution ultrasound microvascular imaging (SR-UMI) has gained interest in recent years owing to the success of DL in medical imaging applications. Mainstream DL techniques require an abundance of labeled training data. Since high-quality in vivo training data are challenging to obtain for medical imaging, most of the current studies have relied on simulations to generate training data of microbubbles (MBs) flowing in vasculature. This study investigates the effect of different MB point spread function (PSF) generation methods on the performance of DL models for SR-UMI. We used Gaussian shaped PSF (Gaussian-PSF), numerical simulation software (Field II-PSF), and experimentally collected PSFs (Exp-PSF) to generate DL training sets. The training sets were used to train DL models that performed both MB localization- and non-localization-based SR-UMI. The quality of the models trained with different datasets were evaluated based on their in vivo performance (mouse brain, chicken embryo), with optical imaging as ground truth for some applications. We discovered that Exp-PSF achieved the best performance overall, due to its high resemblance to real experimental data. Finally, we present a pipeline for obtaining Exp-PSF and integrating them with ultrasound MB simulation, with a publicly available Exp-PSF dataset provided via Github.