Physically-Based Inverse Rendering Framework for PET Image Reconstruction

Area of Science:

• Medical Imaging
• Computer Graphics
• Computational Physics

Background:

• Differentiable rendering is crucial for inverse problems in computer graphics.
• PET image reconstruction requires accurate modeling of photon transport.
• Existing methods may lack physical interpretability and optimization efficiency.

Purpose of the Study:

• To develop a physically-based inverse rendering (IR) framework for PET image reconstruction.
• To leverage differentiable rendering principles for enhanced PET image analysis.
• To improve signal-to-noise ratio and tissue contrast in PET imaging.

Main Methods:

• Integrated Monte Carlo sampling with an analytical projector for forward rendering.
• Employed automatic differentiation to obtain voxel-wise gradients for optimization.
• Utilized the Dr.Jit platform for efficient gradient computation.
• Implemented the Maximum Likelihood Expectation Maximization (MLEM) algorithm.

Main Results:

• The IR framework achieved higher signal-to-noise ratio (SNR) and improved image quality compared to CASToR.
• Clinical evaluation showed higher hippocampal standardized uptake value ratios (SUVR) and gray-to-white matter ratios (GWR).
• Demonstrated enhanced tissue contrast for potential improvements in Alzheimer's disease assessment.

Conclusions:

• The proposed IR framework provides a physically interpretable and extensible platform for high-fidelity PET image reconstruction.
• The method shows strong performance in both phantom and clinical brain PET data.
• Offers potential for more accurate disease staging and localization in neurodegenerative disorders.