Boosting positron emission tomography reconstruction with positional encoding-based deep image prior.

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence

Background:

• Positron Emission Tomography (PET) imaging faces challenges in image reconstruction.
• Conventional supervised learning methods have limitations for PET reconstruction.
• Spectral bias is a key issue affecting PET image quality.

Purpose of the Study:

• To enhance PET image reconstruction using deep learning.
• To propose an unsupervised deep learning approach for PET.
• To address the spectral bias problem in PET reconstruction.

Main Methods:

• Leveraging deep image prior for PET reconstruction.
• Employing deep neural networks in an unsupervised manner.
• Incorporating Gaussian Fourier features and Uniform Positional encoding to mitigate spectral bias.

Main Results:

• Demonstrated noticeable improvement in PET image reconstruction performance.
• Validated the framework on Brainweb and naive rat datasets.
• Showcased the efficacy of the proposed approach in overcoming spectral bias.

Conclusions:

• The developed unsupervised deep learning framework effectively enhances PET image reconstruction.
• The integration of Gaussian Fourier features and Uniform Positional encoding successfully tackles spectral bias.
• This work advances PET imaging methodologies through innovative deep learning applications.