Low-Dose X-ray CT Image Reconstruction Based on a Shearlet Transform and Denoising Autoencoder

Computed tomography (CT) delivers a dose of radiation to the patient with significant risk; however, reducing the radiation dose can introduce noise into CT images, which brings uncertainty to clinical diagnosis. To reduce noise in low-dose CT images, the present paper proposes a deep convolutional neural network (CNN) combined with a shearlet transform and denoising autoencoder. Shearlets can provide more information regarding noisy low-dose CT images than the traditional wavelets for denoising. The residual learning is used to avoid building a complicated regression model for mapping low-dose images to normal-dose images due to the inherently rich details in CT images. The merge connections pass shearlets coefficient details from encoder for better reconstruction while upsampling in the decoder. Experimental results show that the proposed method effectively suppresses noise, thereby preserving the edges and structures in low-dose CT images.