Abstract
PURPOSE: To create an accurate 3D reconstruction of the vascular trees, it is necessary to know the exact geometrical parameters of the angiographic imaging system. Many previous studies used vascular structures to estimate the system's exact geometry. However, utilizing interventional devices and their relative features may be less challenging, as they are unique in different views. We present a semi-automatic self-calibration approach considering the markers attached to the interventional instruments to estimate the accurate geometry of a biplane X-ray angiography system for neuroradiologic use. METHODS: A novel approach is proposed to detect and segment the markers using machine learning classification, a combination of support vector machine and boosted tree. Then, these markers are considered as reference points to optimize the acquisition geometry iteratively. RESULTS: The method is evaluated on four clinical datasets and three pairs of phantom angiograms. The mean and standard deviation of backprojection error for the catheter or guidewire before and after self-calibration are [Formula: see text] mm and [Formula: see text] mm, respectively. The mean and standard deviation of the 3D root-mean-square error (RMSE) for some markers in the phantom reduced from [Formula: see text] to [Formula: see text] mm. CONCLUSION: A semi-automatic approach to estimate the accurate geometry of the C-arm system was presented. Results show the reduction in the 2D backprojection error as well as the 3D RMSE after using our proposed self-calibration technique. This approach is essential for 3D reconstruction of the vascular trees or post-processing techniques of angiography systems that rely on accurate geometry parameters.