Deep Kernel Representation for Image Reconstruction in PET.

Area of Science:

• Medical Imaging
• Computational Imaging
• Nuclear Medicine

Background:

• Positron emission tomography (PET) image reconstruction is complex due to ill-posed problems and low photon counts.
• Current kernel methods rely on empirical kernel construction, often yielding suboptimal performance.

Purpose of the Study:

• To introduce a deep kernel method for dynamic PET image reconstruction.
• To enable automated learning of improved kernel models using deep neural networks.

Main Methods:

• Established an equivalence between kernel representation and trainable neural network models.
• Developed a deep kernel method leveraging deep neural networks for automated kernel learning.
• Trained kernels using prior image data for optimized, non-empirical kernel generation.

Main Results:

• The deep kernel method demonstrated superior performance compared to traditional kernel and standard neural network methods.
• Validation was performed using both computer simulations and a real patient dataset.
• The proposed method is directly applicable to single-subject dynamic PET imaging.

Conclusions:

• The deep kernel method offers a significant advancement in dynamic PET image reconstruction.
• Automated kernel learning via deep neural networks provides a more robust and effective approach.
• This method holds promise for improving diagnostic accuracy in PET imaging.