Mixed scale dense convolutional networks for x-ray phase contrast imaging.

Area of Science:

• Medical Imaging
• Computational Physics
• Machine Learning

Background:

• X-ray in-line phase contrast imaging uses Fresnel diffraction patterns to visualize phase shifts and attenuation.
• Reconstructing phase and attenuation is a complex, nonlinear inverse problem.
• Existing methods often struggle with accuracy and resolution.

Purpose of the Study:

• To develop and evaluate supervised learning methods for simultaneous phase and attenuation retrieval.
• To utilize mixed scale dense (MS-D) convolutional neural networks for improved X-ray imaging.
• To enhance the quantitative accuracy and resolution of reconstructed images.

Main Methods:

• Proposed supervised learning using MS-D convolutional neural networks.
• Employed dilated convolutions for multi-scale feature extraction and dense connections.
• Trained networks with simulated homogeneous/heterogeneous objects and experimental data.
• Compared MS-D networks against U-Net and contrast transfer function (CTF) methods.

Main Results:

• MS-D networks demonstrated significant improvements in reconstruction accuracy on simulated noisy data compared to the CTF method.
• Quantitative improvements in low-frequency behavior and resolution were observed on experimental data.
• The MS-D architecture efficiently handles the Fresnel operator for accurate phase and attenuation recovery.

Conclusions:

• Supervised learning with MS-D convolutional neural networks offers a powerful approach for X-ray phase contrast imaging.
• This method surpasses traditional linear methods in both simulated and real-world imaging scenarios.
• The proposed network architecture effectively addresses the challenges of phase and attenuation retrieval.