Feature-enhanced X-ray imaging using fused neural network strategy with designable metasurface

Area of Science:

• Physics
• Materials Science
• Computer Vision

Background:

• Scintillation-based X-ray imaging is crucial for absorption contrast visualization in science and engineering.
• Current methods rely on separate improvements in scintillators and neural network postprocessing, leading to information loss and higher training costs.
• Existing techniques often require extensive data precollection for specific sample types.

Purpose of the Study:

• To develop an integrated imaging and postprocessing framework for X-ray imaging.
• To overcome the limitations of separated optical transmission and electronic calculation processes.
• To enhance feature-specific image quality and reduce computational overhead.

Main Methods:

• Integration of a designable optical convolutional metasurface with a scintillator.
• Development of a fused optical-electronic convolutional autoencoder network.
• Direct connection of the autoencoder to the X-ray down-conversion process.

Main Results:

• Demonstrated feature-specific enhancement of incoherent X-ray images.
• Achieved applicability to multi-class samples without additional data precollection.
• Experimental validation showed enhanced textural and regional features with up to 11.2 dB signal-to-noise ratio improvement.

Conclusions:

• The proposed fused optical-electronic framework effectively integrates X-ray imaging and neural network postprocessing.
• This approach minimizes optical information loss and reduces training costs.
• The method shows significant potential for advancing X-ray imaging, particularly in number recognition and bioimaging.