By introducing prior information, traditional reconstruction methods for sparse-view CT are based on convex optimization. However, how to select the prior information and accelerate convergence remains a great challenge. Fortunately deep learning has a great promise in image processing such as de-noise, de-blur, or feature automatic selection etc.

A research group led by Prof. XIE Yaoqin, from Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences (SIAT), has made new progress in sparse-view CT reconstruction method for the rapid and low-dose CT.

In this work, a new deep learning (hereinafter referred to as DL) based reconstruction method for sparse-view CT is proposed. It can outperform the other methods in terms of computational efficiency and image quality. The denseNet and deconvolution (hereinafter referred to as DD-Net) reconstruction method takes advantage of the classical FBP reconstruction technique, which is an analytical process and can be calculated efficiently, and the DL based image optimization approach, which is known to be capable of efficiently learning low-level and high-level common features from dataset through a multi-layer network. Compared to the other state-of-the-art reconstruction methods, the DD-Net method can increase the structure similarity (SSIM) by up to 18%, and reduce the root mean square error (RMSE) by up to 42%. These results indicate that DD-Net has great potential for sparse-view CT image reconstruction.

Fig. The corresponding network architecture of the DD-Net.(Image by Prof. XIE Yaoqin)

The paper entitled“A Sparse-View CT Reconstruction Method Based on Combination of DenseNet and Deconvolution” has been published in IEEE Transaction on Medical Imaging.