Neural KEM: A Kernel Method With Deep Coefficient Prior for PET Image Reconstruction.

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence

Background:

• Positron emission tomography (PET) image reconstruction from low-count data is challenging.
• Kernel methods, like kernelized expectation-maximization (KEM), incorporate image priors but can be complex to optimize.
• Existing methods often require explicit regularization, increasing computational complexity.

Purpose of the Study:

• To propose an implicit regularization method for kernel-based PET image reconstruction using deep learning.
• To develop a novel algorithm, neural KEM, for improved reconstruction of low-count PET data.
• To demonstrate the effectiveness of neural KEM compared to existing methods.

Main Methods:

• A deep coefficient prior was used to represent kernel coefficients within the PET forward model via a convolutional neural network.
• The principle of optimization transfer was applied to derive the neural KEM algorithm.
• The algorithm iteratively updates the image using a KEM step and the kernel coefficients using a deep learning step.

Main Results:

• The neural KEM algorithm guarantees monotonic increase in data likelihood.
• Computer simulations and real patient data showed superior performance of neural KEM.
• The proposed method outperformed conventional KEM and deep image prior techniques.

Conclusions:

• Neural KEM offers an effective approach for implicit regularization in PET image reconstruction.
• This method enhances image quality in low-count PET data scenarios.
• The integration of deep learning with KEM provides a promising direction for advanced medical imaging reconstruction.