Enhanced 3D reconstruction of extreme sparse view terahertz computed tomography by using ASE-UNet incorporating asymmetric convolution blocks and channel attention mechanisms

•A novel deep learning-based methodology to enhance the 3D reconstruction resolution of sparse view THz-CT is proposed.•The proposed neural networks used ASE-UNet incorporating asymmetric convolution blocks and channel attention mechanisms.•The proposed deep regression network is designed to enhance the quality of reconstructed images by effectively denoising artifact-ridden data. By capturing time-resolved electric field signals, terahertz computed tomography (THz-CT) provides valuable insights into the intrinsic geometrical features of the THz time-domain spectral system. However, the constraints of hardware-based scanning methods introduce challenges such as suboptimal acquisition efficiency for THz-CT imaging and substantial computational overhead during the reconstruction process. To address these challenges, we introduce a novel deep learning-based methodology to enhance the 3D reconstruction resolution of sparse view THz-CT by using asymmetric convolution blocks and channel attention mechanisms. Our proposed deep regression network, enhanced UNet with asymmetric convolution and squeeze-and-excitation block (ASE-UNet), is designed to improve the quality of reconstructed images by effectively denoising artifact-ridden data. Additionally, a preprocessing step that involves the integration of bandwidth power facilitates the transformation of time-domain data into high-resolution spatiotemporal data. According to experimental findings, the framework can suppress stripe artifacts in sparse views and is more efficient in data usage than traditional iterative algorithms.