In this work, a novel cascaded neural network architecture was developed to perform cone beam CT image reconstruction using the deep learning method. The proposed architecture consists four individual stages: a manifold learning stage to perform projection data pre-processing, a convolutional neural network (CNN) stage to perform data filtration, a fully connected layer with sparse regularization to perform single-view backprojection, and a final fully connected layer with linear activation to generate the target image volume. In manifold learning stage, a novel feature combining technique was proposed to improve noise properties of the final reconstructed images. These 13-layer deep neural network work trained using extensive numerical phantom with noise contaminated projection data and ground truth image in a stage-by-stage pretraining stage. After pretraining with numerical phantom data, the cascaded neural network model was fine tuned using physical phantom data from a diagnostic MDCT scanner. After training, the trained neural network model was used to reconstruct low dose CT images for human subjects from a prospective low dose CT protocol. In these studies, it was found that the proposed cascaded neural network based deep learning method can (1) enable low dose CT reconstruction without noise streaks and with reduced noise amplitude; (2) well maintain reconstruction accuracy at reduced dose levels; and (3) unlike the currently available statistical model based image reconstruction (MBIR) methods, the proposed deep learning reconstruction method can well maintain the similar dose-normalized noise power spectrum (NPS) with that of the FBP reconstructed images.
|