Combined spatial and temporal deep learning for image noise reduction of fluoroscopic x-ray sequences

摘要

Radiation dose is of an important consideration for x-ray fluoroscopy imaging of interventional C-arm systems. Low-dose imaging is always expected, but it also results in noisy images. Noise reduction is one of the important topics for fluoroscopic images. Recently, the advances in deep learning have achieved outstanding denoising results for x-ray images. However, most existing methods in the field focus only on 2D image denoising from frame-by-frame independently, and removing temporal noise in image sequence remains a challenging problem. Our goal is simultaneously to reduce both spatial and temporal noises for fluoroscopic image sequences in a unified framework. In this paper, we propose a deep learning algorithm that extensively utilizes temporal information to maximize the efficiency of noise reduction. The proposed convolutional neural network (CNN) is based on DenseNet and DnCNN but with improved multi-channel input layers for image sequence. That network architecture not only enables spatial domain deep learning from the input of every individual frame, but also is able to make full use of temporally correlative information among adjacent frames for temporal domain learning. In order to further suppress temporal noise resulting in visual flickers of image sequence, an additional term is introduced to the network loss function. Besides two conventional terms of L2 and perceptual losses, the new proposed loss calculates the statistical variance of the network performance caused by random influence of temporal imaging. The developed algorithm is evaluated with fluoroscopic phantom images and clinical patient data, showing superior performance for spatio-temporal denoising.