Exploring the Generalizability of DeepNLCI in Out-of-Domain Scenarios

To address the limitations of contrast-enhanced ultrasound (CEUS) imaging, such as noise, motion artifacts, and a decreased frame rate, a single-pulse deep learning approach, namely DeepNLCI, was previously proposed. This study evaluates the network’s zero-shot robustness against changes of key imaging parameters, including voltage, focus depth, time gain compensation (TGC), and pulse length, as well as transmit, sampling, and demodulation frequencies. For this, datasets were acquired with individually varied parameters. These were tested directly on a pretrained network. The need for further training was evaluated by fine-tuning the pretrained network and a comparison to conventional multipulse CEUS was performed. While maintaining the network’s noise reduction ability, the results indicate that DeepNLCI maintains strong zero-shot performance when focus depth and sampling frequency are changed. For variations in the pressure, pulse length or TGC, the network remains robust mostly while achieving higher contrast metrics than conventional CEUS, and even surpassing fine-tuning scenarios in some cases. The network is most sensitive to changes in transmit and demodulation frequencies, however, its performance can be restored through targeted retraining. Overall, the study demonstrates that DeepNLCI reliably generalizes to most investigated parameter changes without additional training, highlighting the potential of zero-shot learning and the selective need for transfer learning.