Noise reduction in fluoroscopic image sequences for joint kinematics analysis

Analysis of dynamic videofluoroscopic can provide spine kinematic data with an acceptable low X-ray dose. Estimation of the kinematics relies on accurate recognition of vertebrae positions and rotations on each radiological frame. In previous works we presented a procedure for automatic tracking of vertebra motion by smoothed gradient operators and template matching in fluoroscopic image sequences. A limitation to the accurate estimation of the kinematics by automatic tracking of vertebrae motion, independently by the specific methodology employed (e.g. manual marking, corner or edge automatic detection, etc.), is mainly due to noise: low-dose X-ray image sequences exhibit severe signal-dependent noise that should be reduced, while preserving anatomical edges and structures. Noise in low-dose X-ray images originates from various sources, however quantum noise is by far the more dominant noise in low-dose X-ray images and other sources can be neglected. Signal degraded by quantum noise is commonly modeled by a Poisson distribution, but it is possible to approximate it as additive zero-mean Gaussian noise with signal-dependent variance. In this work we propose a digital spatial filter for reducing noise in low-dose X-ray images. The proposed filter is based on averaging of only similar pixels (whose grey level is contained within ±3σ) instead of spatial averaging of all neighbouring pixels. The effectiveness of the filter performance was evaluated by fluoroscopic image sequence processing, comparing the results of the automatic vertebra tracking on filtered and unfiltered images.