Deep Learning Driven Noise Reduction for Reduced Flux Computed Tomography.

Area of Science:

• Geoscience
• Materials Science
• Computer Science

Background:

• Computed tomography (CT) is crucial for geomaterial analysis but often requires high radiation doses and long acquisition times.
• High-density materials in geomaterials increase X-ray attenuation, necessitating higher doses for acceptable image quality.
• Prolonged scanning in micro-CT hinders the study of fast-evolving phenomena.

Purpose of the Study:

• To apply deep convolutional neural networks (DCNNs) to enhance the quality of rock CT images.
• To significantly reduce CT scan exposure times for geomaterials.
• To investigate the effectiveness of transfer learning and different loss functions for DCNNs in this context.

Main Methods:

• Utilized DCNNs to reconstruct high-quality CT images from low-quality, low-dose data.
• Applied transfer learning to optimize DCNN performance without extending training duration.
• Compared DCNN performance using mean squared error and structural similarity index loss functions.

Main Results:

• Achieved simultaneous improvement in rock CT image quality and reduction in exposure time by over 60%.
• Demonstrated the efficacy of transfer learning in enhancing DCNN results for geomaterial CT.
• Showcased the applicability of the DCNN approach across various computed tomography technologies.

Conclusions:

• DCNNs offer a powerful solution for overcoming limitations in geomaterial CT scanning.
• The developed method enables faster, higher-quality imaging, facilitating the study of dynamic processes.
• The approach is adaptable and provides a valuable tool for diverse CT applications.