ScatterFusionNet: physics-informed deep scatter correction for dual-detector CT using Klein-Nishina prior

Area of Science:

• Medical Imaging
• Computational Physics
• Artificial Intelligence

Background:

• Cone-beam CT (CBCT) images suffer from scatter artifacts, degrading diagnostic quality.
• Current deep learning methods struggle with generalization across diverse anatomical regions.
• Acquiring scatter-free data in clinical settings is often impractical due to time constraints.

Purpose of the Study:

• To develop a physics-informed deep learning framework for robust scatter correction in CBCT.
• To achieve cross-anatomy generalization without requiring extensive site-specific training data.
• To improve the quality of CBCT images by mitigating scatter artifacts.

Main Methods:

• Proposed ScatterFusionNet, a physics-informed neural network incorporating Klein-Nishina scattering priors.
• Utilized dual-detector CT side-detector measurements fused with a multi-scale backbone via Feature-wise Linear Modulation (FiLM).
• Trained the model on Monte Carlo simulations and fine-tuned with a single dataset.

Main Results:

• ScatterFusionNet achieved significant Contrast-to-Noise Ratio (CNR) improvements (5.7% on right-teeth, 3.6% on left-teeth datasets).
• Performance closely matched ground truth from beam stop array measurements.
• A classical SE UNet baseline showed substantially weaker generalization (0.8%-1.0% CNR gains).

Conclusions:

• Embedding physics-informed priors is crucial for robust scatter correction in deep learning.
• ScatterFusionNet demonstrates superior cross-anatomy generalization compared to purely data-driven methods.
• The framework reduces reliance on extensive anatomy-specific training data for effective scatter correction.