Scan-wise generalized PET denoising with contrastive adversarial learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Deep learning (DL) excels at low-count Positron Emission Tomography (PET) denoising.
• Distribution shifts from anatomical/physiological variations cause biased outputs and poor generalization in existing DL models.
• Current methods fail to address scan-wise variations effectively.

Purpose of the Study:

• Formulate PET denoising as a scan-wise Domain Generalization (DG) problem to mitigate variations.
• Achieve robust and unbiased denoising for unseen PET scans.
• Develop a DL framework that generalizes well across different scan distributions.

Main Methods:

• Propose a contrastive adversarial domain generalization framework to learn scan-invariant features.
• Utilize multiple noise realizations from single raw list-mode PET scans to create scan-wise domains.
• Introduce an ordered and memory-queued contrastive adversarial framework for longitudinal data, employing a novel noisy-robust multipositive ordinal contrastive loss.

Main Results:

• The contrastive adversarial DG approach outperformed cross-entropy-based adversarial methods and standard baselines.
• Ordered contrastive loss enhanced peak signal-to-noise ratio and structural similarity index.
• Bias and standard deviation were reduced in Alzheimer's-related regions and the whole brain.

Conclusions:

• This study is the first to address cross-scan denoising degradation using domain generalization.
• Utilizing longitudinal scans as pseudo-positives within an ordered contrastive learning scheme is a pioneering approach.
• The proposed methods offer a pathway to robust clinical PET imaging applications.