Open Collections

Machine learning in ultrasound-guided spinal anesthesia

University of British Columbia

Acute and chronic pain treatment with topically or orally administered agents alone is often insufficient and requires injection of analgesic or anesthetic agents directly at the nociceptive site. Examples of target site injections include epidural or facet joint blocks in treatment of acute labour pain or chronic back pain, respectively. In this thesis, machine learning models and algorithms are proposed that aim to facilitate spinal injections through automatic identification of the anatomy in 2D and 3D ultrasound (US). In particular, the proposed techniques detect the anatomical landmarks of the needle target in paramedian and transverse planes for lumbar epidural and facet joint injections. Such methods can be used to identify the correct needle puncture site before injection. A local-directional feature extraction method is first proposed in order to recognize the patterns of the US echoes from the vertebrae in paramedian planes. Later, a supervised convolutional model is proposed to localize the needle target for epidural anesthesia in the paramedian US images. The model uses a combination of hand-engineered local-directional features and convolutional feature maps that are automatically learned from the images. In the transverse plane, a deep classifier is designed to identify the symmetry in the image, and accordingly, classify the midline from off-center images. Later, a conditional generative model along with an adaptive data augmentation algorithm is proposed, which synthesize transverse US images based on the performance of the classifier to improve its accuracy. Finally, an unsupervised model is proposed that learns the variations of the data without the need for labels. Unsupervised learning from US images is valuable because there is a major cost associated with data annotation, which can be avoided by unsupervised learning. The above mentioned methods were tested on US images collected in vivo and demonstrated promising performance to be a useful guide for injections. Moreover, the proposed systems can be utilized as training tools to familiarize the novices with the spine anatomy in US.

Text

2019-09-16

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/71689

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/