Reducing Ultrasound Localization Microscopy acquisition time for rat renal arterial tree using deeplearning model DBlink
O.Y. He (TU Delft - Mechanical Engineering)
C.S. Smith – Mentor (TU Delft - BN/Nynke Dekker Lab)
K. Uğurlu – Mentor (TU Delft - Team Carlas Smith)
D. Maresca – Graduation committee member (TU Delft - ImPhys/Medical Imaging)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Ultrasound imaging is a non-invasive imaging method, which uses ultrasound waves to produce images of the internal organs in the body. Since sound waves can interfere with each other, the ultrasound images are diffraction limited. Ultrasound localization microscopy (ULM) is a processing technique, which is able to bypass this diffraction limit by localizing individual spatially isolated contrast agent microbubble (MB)s in the low resolution ultrasound frames. These MBs acts as a point source and appears as a blurry point in the ultrasound frame also known as Point spread function (PSF) whose centroids can be localized with a precision beyond the diffraction limit. By localizing these MBs and tracking their paths over thousands of consecutive ultrasound frames and accumulating their tracks, a super resolution image of the vasculature can be reconstructed. While these super resolution images significant benefits to biomedical applications, they require long acquisition times.
This thesis investigates whether the deep learning model DBlink, a bidirectional convolutional long short-term memory (LSTM) with a Convolutional Neural Networks (CNN) head can reduce the long acquisition time of ULM. An in silico rat renal arterial tree was simulated to provide the data required for training and evaluating the deep learning model. Two different input type were explored for the DBlink model: Localization maps (summed frames of super resolved localizations) and velocity tracks (maps containing super resolved velocity tracks) of the MB. The effect of different receptive field (RFd) sizes were also examined.
The performance of the DBlink model was compared to the conventional ULM method and showed a reduced acquisition time of 8.7 seconds for large radii vessels in silico. However, the reduction in acquisition time diminishes for small radii vessel, where the passage of MB is still limited by low blood flow rate. Although DBlink reduces acquisition time, it introduces hallucinations in the reconstruction of vessels, especially in regions containing dense small vessels.
Overall, this research highlights the use of the deep learning model DBlink in ULM and the use of different input type to reduce the acquisition time in ULM. However, further research is needed in order to apply this deep learning model in vivo.