PET Image Denoising Using a Deep Neural Network Through Fine Tuning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Positron emission tomography (PET) is a crucial functional imaging technique in clinical diagnosis.
• Improving PET image quality is essential for accurate diagnosis and treatment monitoring.
• Current methods for noise reduction in PET images have limitations.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN) for enhancing PET image quality.
• To investigate the use of perceptual loss for preserving image details during CNN training.
• To address the challenge of limited real patient data by utilizing a pre-training and fine-tuning strategy.

Main Methods:

• Training a deep convolutional neural network (CNN) using perceptual loss derived from a pre-trained VGG network.
• Employing a two-stage training approach: pre-training on simulation data and fine-tuning on real patient data.
• Comparing the proposed CNN method against traditional Gaussian filtering for noise reduction.

Main Results:

• The proposed CNN method demonstrated superior noise reduction capabilities compared to Gaussian filtering.
• Evaluation on simulation, real brain, and real lung PET datasets confirmed the method's effectiveness.
• Perceptual loss effectively preserved image details, unlike conventional mean squared error.

Conclusions:

• The developed CNN with perceptual loss offers a significant improvement in PET image quality.
• The pre-training and fine-tuning strategy effectively overcomes limitations of small real-world datasets.
• This approach holds promise for enhancing diagnostic accuracy in clinical PET imaging.