Learning-Based Material Decomposition in Dual Energy CT Using an Unrolled Estimator

Reconstructing multiple material-specific images in dual-energy CT is a challenging non-linear inverse problem. Traditionally, the reconstruction process consists of two steps: material decomposition and tomographic reconstruction. Model-based iterative reconstruction methods that combine material decomposition and tomographic reconstruction into a unified "one-step" framework provide improved estimates, but require longer reconstruction times. To address this, we propose a supervised machine learning technique to accelerate one-step iterative dual-energy CT reconstruction. Specifically, we train a deep neural network embedded in an "unrolling" of a model-based iterative algorithm. We demonstrate this approach on the problem of identifying three materials (adipose tissue, fibroglandular tissue, and calcifications) from simulated breast dual-energy CT data. Empirically, we find that the unrolling approach gives accurate material map estimates in few iterations, and outperforms a baseline image-domain learning approach.