Physics-Constrained Deep Learning for Effective Atomic Number and Density Calculation of Biological Tissues in...

Area of Science:

• Medical Imaging
• Computational Physics
• Artificial Intelligence

Background:

• Photon-Counting Detector Computed Tomography (PCD-CT) offers superior spectral utilization for material decomposition.
• Traditional methods for calculating effective atomic number ($Z_{eff}$) and density ($\rho$) are limited by model approximations and noise amplification.
• High-precision material decomposition is crucial for advanced quantitative imaging.

Purpose of the Study:

• To develop a physics-constrained deep learning network for high-precision joint estimation of $Z_{eff}$ and $\rho$ in PCD-CT.
• To dynamically model nonlinear X-ray interactions for improved decomposition accuracy.
• To overcome limitations of traditional material decomposition techniques.

Main Methods:

• A deep learning network using Swin-Unet as a backbone was developed.
• A Multilayer Perceptron (MLP) dynamically modeled X-ray interactions, including energy-dependent photoelectric effect and Compton scattering.
• A hybrid loss function (L1, SSIM, and physics-informed loss) was employed to integrate data-driven learning with physical principles.

Main Results:

• The proposed method achieved Mean Absolute Percentage Error (MAPE) below 5% for $Z_{eff}$ and $\rho$ decomposition in standard materials, outperforming existing methods.
• Superior Noise Power Spectrum (NPS) performance was observed for the proposed method.
• For biological samples (crayfish, mouse), the method generated $Z_{eff}$ images with high Multi-Exposure Fusion Structural Similarity Index (MEF-SSIM) (0.9559 for crayfish, 0.8950 for mouse) and enhanced detail recovery.

Conclusions:

• The physics-constrained deep learning network effectively learns nonlinear decomposition processes within a data-driven framework.
• The method significantly improves decomposition accuracy, reduces noise, and enhances image quality in PCD-CT.
• This approach offers a promising solution for precise quantitative material analysis in medical imaging.