Robust automatic detection and removal of fiducial projections in fluoroscopy images: An integrated solution

Automatic detection and removal of fiducial projections in fluoroscopy images is an essential prerequisite for C-arm calibration. This paper presents an integrated solution to fulfill this task. A custom-designed calibration cage with a two-plane pattern of fiducials is utilized in our solution. The cage is attached to the C-arm image intensifier and acquired images are calibrated automatically by a three-step on-line calibration algorithm including fiducial projection detection, image calibration, and fiducial projection removal. A sequence of carefully designed image processing operations consisting of image binarization, connected-component labeling, region classification, adaptive template matching, and shape analysis, are developed for an accurate and robust localization of fiducial projections. A similarity measure that is proposed previously for image-based 2D–3D registration is employed in the adaptive template matching to improve the detection accuracy. Shape analysis based on the design information of the calibration cage is used to further improve the robustness of the detection. Thin-plate spline based vector transforms are used to correct the image distortion. The detected fiducial projections are then removed by an image inpainting technique based on the fast marching method for level set applications. Our in vitro experiments show on average 4 s execution time on a Pentium IV machine, a zero false-detection rate, a miss-detection rate of 1.6 ± 2.3%, and a sub-pixel localization error. Using a custom-made tool for checking accuracy, a forward projection error of 1.0 ± 0.4 pixels and a backward projection error of 0.3 ± 0.1 mm were found. We are confident that our solution is fast, robust, and accurate enough for image-guided interventional applications.