Given the potential X-ray radiation risk to the patient, low-dose CT hasattracted a considerable interest in the medical imaging field. The currentmain stream low-dose CT methods include vendor-specific sinogram domainfiltration and iterative reconstruction, but they need to access original rawdata whose formats are not transparent to most users. Due to the difficulty ofmodeling the statistical characteristics in the image domain, the existingmethods for directly processing reconstructed images cannot eliminate imagenoise very well while keeping structural details. Inspired by the idea of deeplearning, here we combine the autoencoder, the deconvolution network, andshortcut connections into the residual encoder-decoder convolutional neuralnetwork (RED-CNN) for low-dose CT imaging. After patch-based training, theproposed RED-CNN achieves a competitive performance relative tothe-state-of-art methods in both simulated and clinical cases. Especially, ourmethod has been favorably evaluated in terms of noise suppression, structuralpreservation and lesion detection.
展开▼
