X-Ray Scatter Estimation Using Deep Splines.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Radiology

Background:

• X-ray scatter degrades image quality in flat-panel imaging and cone-beam CT.
• Current U-net based scatter removal methods show promise but lack physics constraints, risking misleading results.
• Ensuring the reliability of scatter compensation is critical for accurate medical diagnoses.

Purpose of the Study:

• To develop a novel scatter compensation technique for X-ray imaging by integrating B-splines into neural networks.
• To constrain neural network predictions with physical properties for more reliable scatter removal.
• To evaluate the performance and robustness of the proposed B-spline embedded neural network against existing methods.

Main Methods:

• Proposed embedding B-splines as a constrained operator within neural network architectures.
• Utilized synthetic head and thorax datasets, along with real thorax phantom data for evaluation.
• Compared the proposed method against state-of-the-art U-net based scatter removal techniques using quantitative metrics.

Main Results:

• The B-spline embedded neural network performed comparably to U-net on quantitative metrics.
• The proposed approach demonstrated reduced runtime and parameter complexity.
• The method exhibited superior robustness to varying noise levels, preserving signal frequency characteristics unlike U-net which produced artifacts.

Conclusions:

• Embedding B-splines into neural networks offers a physics-constrained approach for effective X-ray scatter compensation.
• This method provides a more robust and efficient alternative to current U-net based techniques.
• The approach enhances diagnostic reliability by preventing spurious results and preserving essential signal information.