Program Information
Real-Time, Image-Based 3D-2D Registration for Ultrasound-Guided Spinal Interventions Using Haar Feature Matching
T De Silva1*, A Uneri1 , J Punnoose1 , M Ketcha1 , X Zhang1 , R Han1 , J Goerres1 , M Jacobson1 , S Vogt2 , G Kleinszig2 , J Wolinsky1 , J Siewerdsen1 , (1) Johns Hopkins Univeristy, Baltimore, MD, (2) Siemens Healthcare XP Division, Erlangen, Germany
Presentations
WE-G-708-4 (Wednesday, August 2, 2017) 4:30 PM - 6:00 PM Room: 708
Purpose: Spine injections are commonly performed in pain clinics worldwide, with back pain affecting 80% of the population. Although ultrasound (US) is a low-cost, real-time, portable imaging modality suitable for these procedures, suboptimal image quality and artifacts confound visualization of deep bony anatomy and have limited its widespread use. Real-time fusion of US images with pre-procedure MRI would provide valuable assistance to guide needle targeting in 3D. Toward this objective, we propose a fast, entirely image-based 3D-2D rigid registration that operates without external hardware tracking.
Methods: Registration of 2D US images is performed via a fast search algorithm that determines probe pose within a predefined set of pose configurations. 2D slices are extracted from a static 3D US image (alternatively, from a 3D US computed from MRI by a forward simulation developed in related work) to construct a feature dictionary comprising different probe poses. Haar features are computed in a four-level pyramid that transforms 2D image intensities to a 1D feature vector, which are in turn matched to the 2D target image. The method was validated in experiments (80 test images) conducted in a lumbar spine phantom with known translations and compared to tracker-based registration.
Results: The Haar feature method demonstrated registration accuracy (mean ± std) = 2.5 ± 2.2 mm and fairly broad capture range (17.1 ± 8.8 mm). The capture range in this context gives an estimate for the admissible spacing between slices when constructing the dictionary. In comparison, mean-squared error (MSE) showed inferior accuracy (6.5 ± 7.7 mm) and capture range (11.0 ± 6.3 mm). Registration run-time with Haar features outperformed MSE by an order of magnitude: (16 ± 2 ms) versus (175 ± 11 ms).
Conclusion: The Haar feature method provided accurate 3D-2D US registration in real-time and could support broader utilization of US-guided needle interventions.
Contact Email: