Deep Learning Network with Object Shape Information for Efficient Continuous-Wave THz Computed Tomography
2
Issued Date
2025-01-01
Resource Type
ISSN
2156342X
eISSN
21563446
Scopus ID
2-s2.0-105017614199
Journal Title
IEEE Transactions on Terahertz Science and Technology
Rights Holder(s)
SCOPUS
Bibliographic Citation
IEEE Transactions on Terahertz Science and Technology (2025)
Suggested Citation
Jintamethasawat R., Inpuak S., Kaewhanam K., Nilruang M., Choobanna P., Bunyinkgool K., Boonyasiriwat C. Deep Learning Network with Object Shape Information for Efficient Continuous-Wave THz Computed Tomography. IEEE Transactions on Terahertz Science and Technology (2025). doi:10.1109/TTHZ.2025.3615105 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/112548
Title
Deep Learning Network with Object Shape Information for Efficient Continuous-Wave THz Computed Tomography
Corresponding Author(s)
Other Contributor(s)
Abstract
This paper presents an efficient continuous-wave THz computed tomography, which can be achieved based on the utilization of deep learning approach together with object shape information. A modified U-Net architecture that can incorporate object shape information was proposed to mitigate effects due to several wave phenomena arising during measurements, such as reflection, refraction, diffraction, or large beam geometry. This is beneficial for most conventional tomographic reconstruction techniques, which, while computationally efficient, struggle to achieve high image quality due to their assumption of straight and narrow ray wave propagation that overlooks such effects and results in strong artifacts in the reconstructed images. The proposed method was evaluated on various sample geometries and materials, and its performance was also compared with other similar deep learning architectures. Results show significant improvements both in terms of image appearance and structural similarity index measure (SSIM). The SSIM values range from 0.681 to 0.953, and from 0.599 to 0.890, for image reconstructions in simple and complex cases, respectively. The proposed method effectively mitigates boundary artifacts and enhances overall image appearance compared with cases where measurements are transformed by other similar deep learning architectures. This suggests its potential as an efficient THz computed tomography technique for non-destructive testing applications.