ComptoNet: a Compton-map guided deep learning framework for multi-scatter estimation in multi-source stationary CT.

Area of Science:

• Medical Imaging Physics
• Computational Imaging
• Deep Learning Applications

Background:

• Multi-source stationary computed tomography (MSS-CT) offers gantryless scanning and simultaneous multi-source emission advantages.
• Lack of anti-scatter grids in MSS-CT causes severe forward and cross scatter contamination, degrading image quality.
• Accurate and efficient scatter correction methods are crucial for MSS-CT applications.

Purpose of the Study:

• To develop an innovative deep learning framework, ComptoNet, for accurate scatter estimation and correction in MSS-CT.
• To integrate Compton-scattering physics with deep learning to address scatter contamination challenges.
• To validate the performance of ComptoNet against existing scatter correction methods.

Main Methods:

• Proposed ComptoNet, a decoupled deep learning framework utilizing Compton-map representation for scatter signals outside the field of view.
• Employed a dual-network architecture: a conditional encoder-decoder for cross scatter and a frequency U-Net with attention for forward scatter.
• Utilized Monte Carlo-simulated data for training and validation of the scatter estimation framework.

Main Results:

• ComptoNet achieved a mean absolute percentage error of 0.84% in scatter estimation.
• Post-correction CT images exhibited nearly artifact-free quality across diverse phantoms and photon counts.
• Demonstrated superior performance in mitigating scatter-induced artifacts compared to other methods.

Conclusions:

• ComptoNet effectively reduces scatter contamination in MSS-CT by leveraging Compton scattering physics and deep learning.
• The proposed framework significantly enhances CT image quality, offering artifact-free results.
• ComptoNet shows robustness and potential for widespread adoption in MSS-CT applications.