Deep learning based tracked X-ray for surgery guidance

Authors:
Publisher website
Google Scholar
Abstract
Minimally invasive interventional cardiology procedures can be guided by fluoroscopic 2-D projections from a C-arm device. However, 2-D projections typically offer a restricted view on the anatomical structures of interest. An augmented reality (AR) view could realises an enhanced understanding of the anatomy. . Therefore, we propose a deep-learning framework for automated detection of semi-opaque markers in fluoroscopy, which enables prediction of the C-arm pose with respect to the preinterventional CT. A convolutional neural network is trained on datasets from both synthetically generated X-ray and real X-ray images to label the markers. The 3-D/2-D point correspondences of the detected markers are used to determine the pose of the C-ARM. The detection of semi-opaque markers yielded an accuracy of 0.60 ± 0.74 mm on real 2D fluoroscopy. The pose of the X-ray source was estimated within a translation and rotation error of 1.01 ± 0.17 mm and 0.30 ± 0.39◦ respectively. Finally, the target registration error was 1.88 ± 0.30 mm for a realistic percutaneous-epicardial-access scenario. The presented approach can in real-time estimate the C-arm pose with respect to the pre-interventional CT and intra-interventional X-ray image which is a key element in the practical application of AR in fluoroscopy guided interventions.

This publication has 25 references indexed in Scilit: