Breathing-Compensated Neural Networks for Real Time C-Arm Pose Estimation in Lung CT-Fluoroscopy Registration

Abstract

Augmentation of interventional c-arm fluoroscopy using information extracted from pre-operative imaging has the potential to reduce procedure times and improve patient outcomes in minimally invasive peripheral lung procedures, where breathing motion, small airways, and anatomical variation create a challenging environment for planned pathway navigation. Extraction of the rigid c-arm pose relative to preoperative images is a crucial prerequisite; however, accurate 2D-3D fluoroscopy-CT soft tissue registration in the presence of natural deformable patient motion remains challenging. We propose to train a patient-specific neural network on synthetic fluoroscopy derived from the patient’s pre-operative CT, augmented by a generalized breathing motion model, to predict c-arm pose. Our model includes an image supervision path that infers the x-ray projection geometry, providing training stability across patients. We train our model on synthetic fluoroscopy generated from preclinical swine CT and we evaluate on synthetic and real fluoroscopy.