Computed tomography image denoising by learning to separate morphological diversity

Computed Tomography (CT) image denoising is a challenging topic because of the difficulty in modeling noise. In this paper, we propose an image decomposition approach to remove noise from low-dose CT images. We model the image as y = X₁ + X₂ where the main structures and noise are two superimposed layers. Total Variation (TV) minimization is used to learn two dictionaries to represent structure and noise respectively and sparse coding is used to separate x₁ and x₂. Finally, an iterative post-processing stage is introduced that uses image-adapted curvelet dictionaries to recover edges. Our results demonstrate that image separation is a viable alternative to the classic K-SVD denoising method.