Automatic Guide-Wire Detection for Neurointerventions Using Low-Rank Sparse Matrix Decomposition and Denoising

In neuro-interventional surgeries, physicians rely on fluoroscopic video sequences to guide tools through the vascular system to the region of interest. Due to the low signal-to-noise ratio of low-dose images and the presence of many line-like structures in the brain, the guide-wire and other tools are difficult to see. In this work we propose an effective method to detect guide-wires in fluoroscopic videos that aims at enhancing the visualization for better intervention guidance. In contrast to prior work, we do not rely on a specific modeling of the catheter (e.g. shape, intensity, etc.), nor on prior statistical learning. Instead, we base our approach on motion cues by making use of recent advances in low-rank and sparse matrix decomposition, which we then combine with denoising. An evaluation on 651 X-ray images from 5 patient shows that our guide-wire tip detection is precise and within clinical tolerance for guide-wire inter-frame motions as high as 6 mm.