Calibration by differentiation - Self-supervised calibration for X-ray microscopy using a differentiable cone-beam...

Area of Science:

• Biomedical Imaging
• Materials Science
• Computational Science

Background:

• High-resolution X-ray microscopy (XRM) is crucial for studying micro-scale biological structures, including bone.
• Imaging living mice at micrometer resolution for osteoporosis research presents significant data processing challenges.
• Current XRM reconstruction software is often proprietary, limiting user control and deep learning integration.

Purpose of the Study:

• To develop an open-source, differentiable X-ray microscopy reconstruction pipeline for biological imaging.
• To enable deep learning integration for advanced image processing and artifact correction in XRM.
• To improve the resolution and quality of in vivo bone imaging for osteoporosis research.

Main Methods:

• Developed an analytically computed, differentiable XRM reconstruction pipeline.
• Integrated trainable modules for pre- and post-reconstruction processing compatible with deep learning frameworks.
• Implemented a self-supervised quality metric for calibrating a cupping correction module on raw projection data.

Main Results:

• The differentiable pipeline offers full user control and deep learning compatibility.
• Self-supervised calibration of the cupping correction module significantly reduced artifacts.
• Image quality improved, decreasing grey value differences between outer and inner bone by 68-94% without external calibration data.
• The reconstruction process became independent of the XRM manufacturer.

Conclusions:

• Differentiable reconstruction is a powerful tool for advancing XRM in biological imaging.
• This pipeline facilitates the exploration of deep learning methods for XRM and other computed tomography systems.
• It represents a significant step towards achieving single-micrometer resolution for in vivo bone imaging.