Low Dose CT Denoising Using Dilated Residual Learning with Perceptual Loss and Structural Dissimilarity

Low Dose CT Denoising is an open research problem which aims to reduce the risks of radiation exposure to patients. Recently researchers have used deep learning to denoise images with unknown noise distributions with promising results. However, approaches that use mean-squared-error (MSE) tend to over smooth the image resulting in loss of fine structural details in low contrast regions of the image. These regions are often crucial for diagnosis and must be preserved in order to maintain the diagnostic value of the image. In this work we show that using a new objective function which combines MSE, perceptual loss and structural dissimilarity (DSSIM) can effectively denoise low dose CT images while preserving fine structural details in low contrast regions. Further, we show that using a dilated residual network with fewer parameters outperforms a traditional deep convolutional neural network.