(Sunday, 7/14/2019) 2:00 PM - 3:00 PM
Room: 304ABC
Purpose: Metal artifacts often obscure image quality in cone-beam CT (CBCT) and confound the assessment of device placement. We present a metal artifact reduction (MAR) method that uses prior information of surgical instruments ("known components," KC), eliminating the segmentation step that often limits conventional MAR. The KC-MAR method was assessed for the first time in clinical studies of patients undergoing spine surgery.
Methods: The IRB-approved clinical pilot study involved 13 patients imaged after placement of spinal instrumentation, presenting a total of 76 screws varying in diameter, length, material, and manufacturer. Screws were automatically detected in CBCT projection data using a machine learning-based initialization followed by 3D-2D registration of the component model. This two-stage detection scheme avoids conventional segmentation and is hypothesized to provide fully automated localization of the devices with sub-pixel accuracy. Projection data were then corrected by simple inpainting (with more sophisticated polyenergetic methods in future work) and reconstructed by filtered back-projection (FBP). Images were evaluated in terms of artifact magnitude (reduction of blooming artifacts) and suitability to the clinical task of evaluating breach of the pedicle corridor.
Results: The 3D-2D registration demonstrated 3D screw localization error of ~0.5 mm and <1°. KC-MAR yielded 25-98% reduction in blooming artifact about the screw, enabling clear visualization of surrounding anatomical structures, such as the pedicle and adjacent nerves and vessels. Runtime of the initial implementation was <2 min for combined 3D-2D registration and FBP reconstruction.
Conclusion: KC-MAR provided strong reduction of metal artifacts and overcame segmentation errors that present a pitfall to conventional MAR. The method was demonstrated for the first time in a clinical study for a broad, realistic range of spine screw instrumentation. The method is compatible with standard FBP and may be consistent with routine clinical workflow.
Keywords
Cone-beam CT, Image Artifacts, Registration
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
