Low-dose dual energy CT image reconstruction using non-local deep image prior

Abstract

For low-dose dual energy CT (DECT) scans, the difference image between the low and high energy images are noisy and always post-smoothed to achieve diagnosis value. Recently the deep image prior framework shows that convolutional neural networks (CNNs) can learn intrinsic structural information from the corrupted images, without pre-training or high-quality training labels. Inspired by this concept, we represented the low-energy and the difference images as the two-channel output of a CNN and embedded this representation into the DECT system model. Summation of low and high energy CT images reconstructed using FBP was treated as the prior image and supplied as the network input. A non-local layer calculated based on the prior image was integrated into the network structure as additional constraints. Through this CNN representation, the low and high energy images are reconstructed jointly and benefit from the features extracted from the prior image. We formulated the proposed DECT joint reconstruction framework as a constrained optimization problem and solved it using the alternating direction method of multipliers (ADMM) algorithm. Experimental evaluation based on a low-dose DECT dataset shows that the proposed method can outperform the reference denoising methods.