Direct Dual Energy CT Material Decomposition Using Noise2Noise Prior

Abstract

Dual energy computed tomography (DECT) can provide material decomposition capability, which can be useful for many clinical diagnosis applications. But the decomposed images can be very noisy due to the dose limit in the scanning and the ill-condition of decomposition process. Recently Noise2Noise framework shows its potential on restoring images by using only noisy data. Inspired by this, we proposed an iterative DECT reconstruction algorithm with a Noise2Noise prior. The algorithm directly estimates material images from projection data and thus can significantly reduce possible bias which may occur in other post-smoothen methods. The Noise2Noise prior was built by a deep neural network, which did NOT need external data for training. The data fidelity term and the Noise2Noise network are alternatively optimized respectively using separable quadratic surrogate (SQS) and Adam algorithm. The method was validated both on simulation data and real clinical data. Quantitative analysis demonstrates the method's promising performance on denoising, bias avoiding and detail reservation.