Terahertz deep learning computed tomography in extreme sparse view

We introduce a novel deep learning-based framework to decrease data acquisition time and enhance the 3D reconstruction resolution of sparse view THz-CT. Using asymmetric convolution blocks and channel attention mechanisms, our network ASE-UNet effectively suppresses terahertz image artifact in extreme sparse views. Furthermore, a preprocessing step involving integrating bandwidth power effectively converts time-domain data into detailed spatiotemporal data. The results show that our method delivered superior reconstruction quality and is more efficient in data usage than traditional iterative algorithms