Terahertz spatio-temporal deep learning computed tomography.

Area of Science:

• Physics
• Computer Science
• Imaging Science

Background:

• Terahertz computed tomography (THz CT) offers unique imaging capabilities but faces limitations in data efficiency and model extensibility.
• Current physics-model-based THz CT methods struggle with time-resolved signals and integrating diverse data, hindering practical applications.

Purpose of the Study:

• To introduce a supervised deep learning framework for THz CT (THz DL-CT) utilizing time-domain information.
• To enhance the reconstruction of 3D object tomographic images from spatio-temporal THz signals without requiring prior material data.

Main Methods:

• A supervised deep learning computed tomography (DL-CT) framework was developed for Terahertz (THz) imaging.
• The THz DL-CT model extracts features directly from spatio-temporal THz signals, bypassing the need for material information.
• Performance was evaluated against conventional and machine learning-based methods using RMSE and SSIM metrics.

Main Results:

• THz DL-CT demonstrated superior performance, achieving at least 50.2% better root mean square error (RMSE) and 52.6% better structural similarity index (SSIM) compared to existing methods.
• The framework successfully reconstructed high-quality tomographic images of 3D objects.
• Experimental validation confirmed the model's generalizability to multi-material systems without prior information.

Conclusions:

• The proposed THz DL-CT framework significantly improves image reconstruction quality and data efficiency in Terahertz computed tomography.
• This deep learning approach offers a novel pathway for non-invasive functional imaging and object investigation using THz CT.
• The ability to reconstruct images without prior material information broadens the applicability of THz CT in various scientific and industrial fields.