Time varying flat field correction of X-ray microtomography with an improved deep-learning method

Area of Science:

• Materials Science
• Imaging Science
• Artificial Intelligence

Background:

• X-ray microtomography (XRM) requires flat field images for accurate sample projection normalization.
• Dynamic in-situ/in-operando XRM is crucial for studying microstructure evolution but faces challenges in acquiring accurate flat field data due to fast acquisition and bulky equipment.
• Conventional flat field correction methods struggle with artifacts in dynamic XRM.

Purpose of the Study:

• To develop a deep-learning-based method for accurate flat field correction in X-ray microtomography.
• To address the limitations of conventional methods in dynamic CT imaging scenarios.
• To improve the quality of CT reconstructions by reducing systematic errors in intensity normalization.

Main Methods:

• An improved pix2pixHD generative adversarial network model was employed to generate flat field images from individual CT projections.
• The deep learning model was trained and validated using experimental data from X-ray microtomography.
• The method's performance was compared against conventional techniques and existing deep learning approaches.

Main Results:

• The proposed deep learning method significantly outperformed conventional and other deep learning-based flat field correction techniques.
• The method effectively reduced systematic errors during intensity normalization, leading to substantial improvements in CT reconstruction quality.
• The model demonstrated generalization capabilities, successfully correcting flat field images for various low Z material samples not included in the training set.

Conclusions:

• The developed deep learning method is a practical solution for flat field correction in dynamic in-situ CT imaging.
• The approach is applicable to new samples after effective neural network training, showcasing its versatility.
• This method enhances the reliability and accuracy of microstructure evolution studies using XRM.