Analysis of Deep Neural Network Architectures and Similarity Metrics for Low-Dose CT Reconstruction

Abstract

Computed tomography (CT) is a widely used imaging technique in the medical field. During CT procedures patients are exposed to high amounts of radiation, posing a tangible threat to their health. Developed low-dose procedures lower exposure but produce noise and artifacts in images. To improve diagnostic accuracy, deep learning techniques are proposed to remove noises and artifacts from low-dose images. In this paper, the performance of several neural network architectures and similarity metrics as loss functions for low-dose CT image reconstruction are analyzed. Experimental results showed that selection of loss function can have significant impact on model performance, with the GSSIM metric outperforming other contemporary metrics SSIM, MSSSIM and MSE. Experiments were conducted using open-access and local cancer research institution data.