A Comparison of Denoising Algorithms for Effective Edge Detection in X-Ray Fluoroscopy

Abstract

X-ray fluoroscopy provides various diagnosis and is widely used in interventional radiology. However, the low-dose involved in fluoroscopy generates an intense Poisson-distributed quantum noise. Object recognition and tracking help in many fluoroscopic applications. Edge-detection is essential, but common derivative operators require noise suppression to provide reliable results. Moreover, homoscedasticity of noise is generally assumed, but is not the case of fluoroscopic images. However, the Anscombe transform can stabilize the quantum noise variance. This study presents a comparison of two denoising algorithms to evaluate their performance in edge-detection for real fluoroscopic sequences. VBM4D is one of best video-processing method for Additive White Gaussian Noise (AWGN), while Noise Variance Conditioned Average (NVCA) is a recent, real-time, algorithm specifically tailored for fluoroscopy. Some real fluoroscopic sequences screening the motion of lumbar spine were processed. Noise parameters were estimated using image sequences of a static scene: the relationship between the luminance and the noise variance was obtained. Generalised Anscombe transform and its inverse were applied to use the VBM4D algorithm. Edge-detection was performed by means of the Sobel operator. The Anscombe transform resulted able to stabilise the noise variance and consequently allow the use of algorithms designed for AWGN. The results show that both approaches provide effective identification of object contours (i.e. vertebral bodies). Despite of its simplicity the NVCA algorithm shows better performances than VBM4D on delineation of boundaries of examined spine fluoroscopic scenes. Furthermore, the NVCA algorithm can be realized in hardware and can offer real-time fluoroscopic processing.