Deep residual unfolding: A novel sparse computed tomography reconstruction method leveraging iterative learning and neural networks

Abstract

Addressing the challenge of compromised imaging quality in sparse Computed Tomography (CT) reconstructions—a significant issue due to the inherent sparsity of CT scan data—we present a novel method grounded in both iterative unfolding and deep learning techniques. This sparse CT reconstruction method involves constructing and minimizing an energy model, which comprises both a data fidelity term and a regularization term. Unlike traditional methods that rely on sparse variation operators like Total Variation (TV), our approach employs a deep learning model instantiated as a Residual Convolutional Neural Network. Enhanced with Group Normalization and the Swish activation function, this network iteratively unfolds the energy model to improve reconstruction quality. Our method's efficacy is validated through metrics such as Peak Signal Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE) on a specific CT dataset. Overall, this work contributes to the advancement of sparse CT reconstructions, with significant potential implications for medical imaging applications.