Implicitly Defined Material Decomposition Estimator and Learned Physics-Informed Neural Proxy for Photon Counting CT.

Area of Science:

• Medical Imaging
• Computational Imaging
• Data Science

Background:

• Photon counting detector-based CT (PCCT) systems enable quantitative imaging through material decomposition (MD).
• Iterative maximum-likelihood estimation (MLE) for MD is accurate but computationally intensive.
• Empirical methods offer speed but can introduce bias and noise.

Purpose of the Study:

• To develop a computationally efficient method for material decomposition in PCCT.
• To leverage the implicit function defined by iterative MLE for knowledge distillation.
• To enable high-quality, real-time quantitative spectral imaging.

Main Methods:

• Applied the Implicit Function Theorem to approximate MLE's implicit mapping.
• Utilized neural networks and Sobolev Training to create an explicit proxy model (Proxy MD).
• Performed theoretical Jacobian analysis for differentiable MD and end-to-end training.

Main Results:

• Proxy MD achieved >200 times speedup compared to iterative MLE.
• The proposed method closely approached the performance of iterative MLE.
• Proxy MD outperformed conventional empirical methods in accuracy and noise handling.
• Demonstrated differentiable PCCT quantitative imaging capabilities.

Conclusions:

• Proxy MD offers a computationally efficient and accurate alternative for quantitative spectral imaging in PCCT.
• The approach enables real-time applications and opens avenues for differentiable imaging pipelines.
• Theoretical insights facilitate further advancements in iterative MD optimization.