Enhancing Precision in Image-Guided Spine Surgery through the Prediction of Occluded Fiducials Utilizing ResNet Architecture

With advancements in imaging technologies, the C-Arm fluoroscopy imaging modality is essential for intra-operative visualization in Image Guided Spine Surgery(IGSS). That necessitates the development of computational techniques for calibrating and correcting distortions in C-Arm images. For this purpose, calibration drums with embedded metallic fiducials are typically attached to the C-Arm detector. Our research introduces a novel algorithm designed to predict occluded distortion fiducials, thereby enhancing precision in IGSS. The fundamental strategy is to minimise reprojection error(RPE) by correcting image distortion using these fiducials. Fiducial occlusion often arises from surgical instrument placement, reduced contrast in the C-Arm machine, pre-existing screws in patients, or deficiencies in imaging algorithms. Our approach employs an in-house dataset to train a ResNet architecture for retrieving occluded fiducials by image registration, achieving similar image detection accuracy between 86% to 90%. Once occluded fiducials are identified, a thin-plate spline fitting algorithm is applied for distortion correction. This method, utilizing retrieved fiducials and spline fitting, resulted in a 9% reduction in reprojection error.