Depth resolved pencil beam radiography using AI - a proof of principle study

Area of Science:

• Medical Imaging
• Computational Physics
• Artificial Intelligence

Background:

• Clinical radiography relies on X-ray transmission, with scatter often treated as noise.
• Compton scatter, a type of photon scatter, contains characterizable information.
• Advanced imaging requires novel methods to extract more data from X-ray interactions.

Purpose of the Study:

• To investigate if deep learning can utilize scattered photon information for improved planar X-ray imaging.
• To resolve superimposed attenuators in 2D radiography by analyzing scattered X-rays.
• To demonstrate constructive use of Compton scatter data for material depth inference.

Main Methods:

• Simulated a monoenergetic X-ray imaging system with a high-resolution detector array.
• Measured off-axis scatter location and energy to maximize information capture.
• Employed a convolutional neural network trained on Monte Carlo simulations of stacked materials (air/bone/water) to classify material depth.

Main Results:

• Achieved high accuracy (0.91±0.01) in resolving material depth information from simulations.
• Demonstrated high average sensitivity (0.91) and specificity (0.95) in material identification.
• Material identification accuracy was slightly higher at the object's entrance and exit surfaces.

Conclusions:

• Proof of principle established for using deep learning to analyze Compton scatter patterns in radiography.
• Scattered photon information can constructively enhance 2D planar imaging for depth inference.
• Further research is needed to address limitations of simple test scenarios and clinical scalability.