Open Collections

Automated lumbar vertebral level identification using ultrasound

University of British Columbia

Spinal needle procedures require identification of the vertebral level for effectiveness and safety. E.g. in obstetric epidurals, the preferred target is between the third and fourth lumbar vertebra. The current clinical standard involves "blind" identification of the level through manual palpation, which only has a 30% reported accuracy. Therefore, there is a need for better anatomical identification prior to needle insertion. Ultrasound provides anatomical information to physicians, which is not obtainable via manual palpation. However, due to artifacts and the complex anatomy of the spine, ultrasound is not commonly used for pre-puncture planning. This thesis describes two machine learning based systems that can aid physicians to utilize ultrasound for lumbar level identification. The first system, LIT, is proposed to identify vertebrae, assigning them to their respective levels and tracking them in a sequence of ultrasound images in the paramedian plane. A deep sparse auto-encoder network learns to extract anatomical features from pre-processed ultrasound images. A feasibility study (n=15) evaluated performance. The second system, SLIDE, identifies vertebral levels from a sequence of ultrasound images in the transverse plane. The system uses a deep convolutional neural network (CNN) to classify transverse planes of the lower spine. In conjunction, a novel state-machine is developed to automatically identify vertebral levels as the transducer moves. A feasibility study (n=20) evaluated performance. The CNN achieves 88% accuracy in discriminating images from three planes of the spine. As a system, SLIDE successfully identifies all lumbar levels in 17 of 20 test scans, processed at real-time speed. A clinical study with 76 parturient patients was performed. The study compares level identification accuracy between manual palpation, versus SLIDE, with both compared to freehand ultrasound. SLIDE's level identification outperformed palpation with an odds ratio of nearly 3. A subset of recorded ultrasound (n=60) was labeled and used to retrain the CNN, improving classification accuracy to 93%. The systems showcase the utility of machine learning in spinal ultrasound analysis, with varied approaches to automatically identifying vertebral levels. The systems can be used to improve the accuracy of vertebral level identification compared to manual palpation alone.

Text

2018-08-31

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/62945

Biomedical Engineering

University of British Columbia

UBCV

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