Reconstruction of 3D interventional cardiovascular device geometries using biplanar X-ray fluoroscopic images

Accurate localization of interventional devices, like guidewires and catheters, is crucial for patient safety and the success of cardiovascular procedures. X-ray fluoroscopy, the most common imaging modality for guiding minimally invasive procedures, offers real-time device visualization but lacks soft tissue contrast and 3D spatial information, limiting precise guidance. This study addresses the need for 3D visualization by developing a deep learning based algorithm to estimate device geometry during cardiac catheterization using biplanar X-ray fluoroscopies. Extracted image frames from fluoroscopies were preprocessed and fed into a U-Net segmentation model to predict device masks, while their DICOM metadata was used to create a geometric model in Cartesian space, where labelled points from the masks were mapped onto detector planes as projections. X-ray vectors were traced between sources and detectors, and their intersections were found to give 3D points. Due to lack of volumetric data for comparison, the 3D model was back-projected to each perspective plane and compared with its original projections. Results showed Mean Absolute Errors of 1.19-5.25mm between original and back-projected points, with an average accuracy of 0.63, precision of 0.8, F1-score of 0.75, and recall of 0.72. The algorithm performed well for simple device structures, but faced limitations with complex shapes and multiple guidewires. Broader and multi-labelled datasets can further improve segmentation and reconstruction accuracy, while volumetric data and comparison with other reconstruction methods can provide better assessment. With its effectiveness against almost all simple structures, and a scalable framework for other interventional procedures as well, the algorithm shows promise in successfully reconstructing 3D geometry from biplanar fluoroscopies.