Improving Interpretability of 2-D Ultrasound of the Lumbar Spine

Ultrasound-guided anesthesia uses a safe, portable imaging modality to provide visual feedback during the needle injection. Widespread adoption of ultrasound-guided anesthesia has been primarily limited by a lack of access to advanced ultrasound technology and a lack of ultrasound training for anesthesiologists. We sought to address these limitations by introducing a method that aids the interpretability of cross-sectional ultrasound from conventional (2D) machines. We propose a constrained registration of a 3D active shape model constructed from computerized tomography (CT) scans of the lumbar spine to a specific set of targets automatically extracted from 2D B-mode ultrasound images with machine learning models. The registration results in an overlay of the entire bone cross-section of the lumbar spine onto the ultrasound image. Our proposed registration achieved a mean squared error of $1.4 \pm 0.3$ mm on a set of 43 ultrasound images, which is smaller than the key anatomical features, suggesting that the overlay is suitable for interpretation.