Exploiting network optimization stability for enhanced PET image denoising using deep image prior

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Positron Emission Tomography (PET) imaging quality is limited by statistical noise from dose and scan duration constraints.
• Deep learning denoising methods can improve PET images but may cause over-smoothing, obscuring details and affecting quantitative accuracy.
• Conditional Deep Image Prior (DIP) is a deep learning technique used for PET image reconstruction.

Purpose of the Study:

• To develop a more reliable deep learning-based method for Positron Emission Tomography (PET) denoising.
• To address the over-smoothing issue in deep learning PET denoising that can compromise image quality and quantitative accuracy.
• To enhance the diagnostic performance and quantitative accuracy of PET imaging, particularly in low-dose scenarios.

Main Methods:

• Introduced 'stability information' into the conditional DIP optimization process to identify unstable network regions.
• Developed a stability map derived from intermediate network outputs at various optimization steps.
• Combined the DIP output with the original PET image using a stability map-weighted linear combination.

Main Results:

• The proposed method effectively reduced background noise in brain [18F]FDG PET images while preserving fine structural details.
• Outperformed existing methods in peak-to-valley ratio and noise suppression across different low-dose PET data.
• Maintained quantitative accuracy, avoiding under- or over-estimation in region-of-interest analyses.
• Significantly reduced noise in full-dose PET images, comparable to unfiltered images in terms of peak-to-valley ratio.

Conclusions:

• The novel method offers a robust approach to deep learning-based PET denoising, improving reliability and quantitative accuracy.
• This technique has the potential to enhance performance in high-sensitivity PET scanners.
• The method can overcome inherent image quality limitations of current PET scanners.