Penalized PET Reconstruction Using Deep Learning Prior and Local Linear Fitting.

Area of Science:

• Medical Imaging
• Deep Learning
• Positron Emission Tomography (PET) Reconstruction

Background:

• Deep learning shows great potential in medical imaging, particularly for positron emission tomography (PET) reconstruction.
• Existing deep learning methods often struggle with varying noise levels during iterative reconstruction, leading to performance degradation.
• The denoising convolutional neural network (DnCNN) is a potential prior but is sensitive to noise level disparities.

Purpose of the Study:

• To propose an iterative PET reconstruction framework utilizing a deep learning-based prior.
• To address the challenge of noise level disparity in deep learning priors for iterative reconstruction.
• To improve the quantitative and qualitative performance of PET image reconstruction.

Main Methods:

• Utilized a denoising convolutional neural network (DnCNN) trained on full-dose and low-dose images.
• Investigated the performance degradation of DnCNN at various noise conditions through simulations.
• Proposed a novel method incorporating a local linear fitting function with the DnCNN prior to mitigate bias.

Main Results:

• DnCNN alone produced additional bias due to noise level disparities between training and testing data.
• The proposed local linear fitting function effectively prevented unwanted bias, demonstrating robustness against noise variations.
• The new method outperformed conventional total variation and non-local means methods in bias and standard deviation studies.

Conclusions:

• The proposed iterative PET reconstruction framework with a bias-corrected deep learning prior significantly improves image quality.
• The method is robust to noise level disparities, making it more applicable in clinical settings.
• This approach offers a superior alternative to traditional reconstruction methods for PET imaging.