Title: Automating Ultrasound Based Vascular Access

Abstract:

Timely care of trauma patients is important to prevent casualties in resource-limited regions such as the battlefield. In order to treat such trauma using point of care diagnosis, medical practitioners typically use an ultrasound for vascular access or detection of subcutaneous splinters for providing critical care. The problem here is two-fold: such ultrasound-based care requires dexterity to move the ultrasound probe on the skin and secondly, interpretation of ultrasound images requires significant prior knowledge and medical training. However, access to medical experts in austere environments is not always readily available. Thus, this thesis aims to automate the aforementioned processes. We (i) investigate the use of a robotic ultrasound system to locate vascular structures and control the ultrasound probe to maintain proper contact with skin, and (ii) we propose an ultrasound-based image segmentation method for interpretation of noisy ultrasound images.

Currently, robotic ultrasound systems rely on human experts to move the ultrasound probe to desired regions; instead, as an alternative to this supervised approach, we propose a method for finding blood vessels with no human expert  supervision. Our system controls the ultrasound probe to maintain proper contact with the skin using force feedback informed by the approximate surface curvature, thus capturing good quality ultrasound images. Once the blood vessels are located and a needle is inserted in the desired region (for vascular access), we develop a method to segment out needles in ultrasound images by exploiting needle motion. Recent  deep learning-based medical  image segmentation methods do not consider needle motion which we believe is an important signal to localize the needle. Our method combines a deep learning method called “U-Net” to capture visual features, and a novel convolutional neural network based on the Kalman Filter for capturing needle motion through time.

We demonstrate the efficacy of the ultrasound probe controller on two types of phantoms: a blue gel and a medical leg. On both phantoms, our method succeeds in finding the vascular dense structures and scanning along the phantom surface while maintaining proper contact. Our needle localization method surpasses state of the art image segmentation methods for needle localization on three different ultrasound datasets.

Committee:

Prof. Howie Choset (chair)

Prof. John Galeotti (co-chair)

Prof. Artur Dubrawski

Cecilia Morales Garza