Abstract
Commercial iterative reconstruction techniques on modern CT scanners targetradiation dose reduction but there are lingering concerns over their impact onimage appearance and low contrast detectability. Recently, machine learning,especially deep learning, has been actively investigated for CT. Here we designa novel neural network architecture for low-dose CT (LDCT) and compare it withcommercial iterative reconstruction methods used for standard of care CT. Whilepopular neural networks are trained for end-to-end mapping, driven by big data,our novel neural network is intended for end-to-process mapping so thatintermediate image targets are obtained with the associated search gradientsalong which the final image targets are gradually reached. This learned dynamicprocess allows to include radiologists in the training loop to optimize theLDCT denoising workflow in a task-specific fashion with the denoising depth asa key parameter. Our progressive denoising network was trained with the MayoLDCT Challenge Dataset, and tested on images of the chest and abdominal regionsscanned on the CT scanners made by three leading CT vendors. The best deeplearning based reconstructions are systematically compared to the bestiterative reconstructions in a double-blinded reader study. It is found thatour deep learning approach performs either comparably or favorably in terms ofnoise suppression and structural fidelity, and runs orders of magnitude fasterthan the commercial iterative CT reconstruction algorithms.