Populational and individual information based PET image denoising using conditional unsupervised learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Positron Emission Tomography (PET) imaging is crucial for disease diagnosis and monitoring.
• Improving the signal-to-noise ratio (SNR) in PET scans is essential for accurate image interpretation.
• Existing denoising methods often require specific training datasets or struggle with preserving fine details.

Purpose of the Study:

• To develop a novel conditional unsupervised learning method for enhancing PET image quality.
• To improve the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of PET images.
• To create a versatile method applicable to various PET/CT and PET/MR datasets without paired training data.

Main Methods:

• A two-step approach involving populational training and individual fine-tuning of a neural network.
• The network utilizes anatomical prior information from CT or MR images to condition the denoising process.
• Leverages noisy PET images as training labels, enabling application to existing datasets.

Main Results:

• Achieved significant contrast-to-noise ratio (CNR) improvements: 71.85% on PET/CT and 58.07% on PET/MR datasets.
• Outperformed traditional methods like Gaussian filtering, Non-Local Means (NLM), and Conditional Deep Image Prior (CDIP) in CNR enhancement.
• Demonstrated accurate restoration of tumor structures while effectively reducing noise in denoised PET images.

Conclusions:

• The proposed conditional unsupervised learning method effectively enhances PET image quality.
• This approach offers a flexible and powerful tool for improving diagnostic accuracy in PET imaging.
• The method's ability to leverage anatomical priors and avoid paired training data makes it widely applicable.