Plug-and-Play gradient-based denoisers applied to CT image enhancement

Highlights：

• The paper proposes three algorithms for deblurring and denoising of Computed Tomography (CT) images, which are usually characterized by a sparse gradient domain.

• The proposed methods include gradient-based regularization inducing sparsity in the gradient image domain.

• The proposed methods exploit in the Plug-and-Play (PnP) framework, where one or more denoisers are plugged in as priors.

• In our methods we propose as denoisers in the PnP framework: 1. A Convolutional Neural Network (CNN) trained on the image gradients (GCNN algorithm) 2. A combination of a CNN trained on the image gradients and a Total Variation (TV) gradient-based (GCNN-TV algorithm) 3. A combination of a CNN trained on the images and a Total Variation (TV) gradientbased (ICNN-TV algorithm).

• The proposed GCNN, GCNN-TV and ICNN-TV methods are tested on synthetic and real CT images and compared with state-of-art algorithms. The results obtained show that the proposed GCNN gradient-based outperforms all the other denoisers in recovering edges of low-contrasted objects, whereas the combination of a Neural Network and of Total Variation in the ICNN-TV and GCNN-TV methods suppresses the residual noise, while preserving the details and contours.

摘要

•The paper proposes three algorithms for deblurring and denoising of Computed Tomography (CT) images, which are usually characterized by a sparse gradient domain.•The proposed methods include gradient-based regularization inducing sparsity in the gradient image domain.•The proposed methods exploit in the Plug-and-Play (PnP) framework, where one or more denoisers are plugged in as priors.•In our methods we propose as denoisers in the PnP framework: 1. A Convolutional Neural Network (CNN) trained on the image gradients (GCNN algorithm) 2. A combination of a CNN trained on the image gradients and a Total Variation (TV) gradient-based (GCNN-TV algorithm) 3. A combination of a CNN trained on the images and a Total Variation (TV) gradientbased (ICNN-TV algorithm).•The proposed GCNN, GCNN-TV and ICNN-TV methods are tested on synthetic and real CT images and compared with state-of-art algorithms. The results obtained show that the proposed GCNN gradient-based outperforms all the other denoisers in recovering edges of low-contrasted objects, whereas the combination of a Neural Network and of Total Variation in the ICNN-TV and GCNN-TV methods suppresses the residual noise, while preserving the details and contours.