PET Image Reconstruction Using Deep Image Prior.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Deep neural networks (DNNs) show promise in medical imaging but require extensive training data, often unavailable in clinical settings.
• Medical image reconstruction, particularly for modalities like PET-CT, necessitates raw data and faces challenges with data scarcity.

Purpose of the Study:

• To develop a personalized network training method for medical image reconstruction that bypasses the need for large prior training datasets.
• To leverage patient-specific prior information within an iterative reconstruction framework.

Main Methods:

• Proposed a personalized deep network training approach inspired by the deep image prior framework.
• Formulated maximum-likelihood estimation as a constrained optimization problem.
• Solved the optimization using the alternating direction method of multipliers (ADMM) algorithm.
• Applied the framework to magnetic resonance imaging-guided positron emission tomography (PET) reconstruction.

Main Results:

• The proposed method demonstrated superior performance compared to traditional Gaussian post-smoothing.
• Outperformed anatomically guided reconstructions using kernel methods or neural network penalties.
• Quantification results from both simulated and real data validated the framework's effectiveness.

Conclusions:

• The personalized network training method offers a viable solution for medical image reconstruction without requiring extensive prior training data.
• This patient-specific approach enhances the quality and accuracy of PET-MR image reconstruction.