Deep-Learning-Based Video Frame Interpolation Method for Ultrasound Localization Microscopy: A Preliminary Study

Ultrasound Localization Microscopy (ULM) achieves super-resolution vascular imaging by accurately localizing the positions of individual ultrasound contrast agents (microbubbles) across multiple consecutive frames. However, traditional ULM methods require ultrafast plane wave imaging to acquire microbubble trajectories. This requirement poses high demands on the equipment and is difficult to achieve with traditional focused wave-based ultrasound devices, thus hindering the clinical application of ULM to some extent. To address this issue, we propose a modified deep learning-based method for Ultrasound Video Frame Interpolation (VFI) with 3× interpolation. We extracted the feature maps from the original input of two frames and subsequently generated the feature maps for the intermediate two frames. This method aims to increase the frame rate in a software-based manner and is specifically designed f or U LM. To evaluate the effectiveness of our proposed method, we conducted experiments using a dataset of rat brain images. The obtained super-resolution frames achieved an average Peak Signal-to-Noise Ratio (PSNR) of 55.16dB and an average Structural Similarity Index (SSIM) of 0.9957. Notably, the resulting images allowed visualization of vessels with a minimum diameter of 26.25 micrometers. These findings highlight the potential of our proposed method in advancing ultrasound-based imaging techniques. By overcoming the limitations imposed by the need for ultrafast plane wave imaging, our method has the potential to pave the way for broader clinical application of ULM, enabling its utilization in a wider range of diseases and conditions.