Deep Learning-based Inaccuracy Compensation in Reconstruction of High Resolution XCT Data

Area of Science:

• Materials Science
• Imaging Technology
• Computational Science

Background:

• High-resolution X-ray computed tomography (XCT) faces challenges with sub-micron precision alignment and object motion.
• These limitations lead to artifacts, degrading the quality of 3D images, especially at micro- and nanoscale applications.

Purpose of the Study:

• To develop a novel, autonomous reconstruction methodology for XCT.
• To address unavoidable misalignment and object motion during data acquisition.
• To enable high-quality 3D image generation and data recovery from incomplete datasets.

Main Methods:

• Developed reconstruction software with correction modules using gradient descent and deep learning.
• Employed a computer vision and deep convolutional neural network (CNN) approach for motion estimation by tracking features in projections.
• Utilized a novel CNN for inferring high-quality reconstruction data from incomplete projection sets.

Main Results:

• Successfully suppressed artifacts from motion, misalignment, and detector inaccuracies.
• Achieved high-quality 3D images without additional computational cost.
• Demonstrated effective data recovery from incomplete projections, yielding corrected projection data.

Conclusions:

• The developed autonomous methodology significantly improves XCT image quality and data integrity.
• Proven effective for micro- and nano-XCT, applicable across all length scales.
• Offers a robust solution for artifact compensation and data recovery in XCT imaging.