Wavelet Residual Network for Low-Dose CT via Deep Convolutional Framelets

Model based iterative reconstruction (MBIR) algorithms for low-dose X-ray CT are computationally expensive. To address this problem, we recently proposed the world-first deep convolutional neural network (CNN) for low-dose X-ray CT and won the second place in 2016 AAPM Low-Dose CT Grand Challenge. However, some of the texture were not fully recovered. To cope with this problem, here we propose a deep residual learning approach in directional wavelet domain. The proposed method is motivated by an observation that a deep convolutional neural network can be interpreted as a multilayer convolutional framelets expansion using non-local basis convolved with data-driven local basis. We further extend the idea to derive a deep convolutional framelet expansion by combining global redundant transforms and signal boosting from multiple signal representations. Extensive experimental results confirm that the proposed network has significantly improved performance and preserves the detail texture of the original images.