Subject-aware PET Denoising with Contrastive Adversarial Domain Generalization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiochemistry

Background:

• Deep learning (DL) significantly enhances positron emission tomography (PET) denoising.
• Subject-specific variations in PET data limit DL model generalizability and clinical reliability.
• A need exists for robust DL models that perform consistently across diverse patient data.

Purpose of the Study:

• To develop a generalizable DL framework for subject-wise domain generalization (DG) in PET denoising.
• To mitigate performance variations caused by subject-specific count levels and spatial distributions in PET imaging.
• To improve the reliability and trustworthiness of DL-based PET denoising for clinical use.

Main Methods:

• Proposed a contrastive adversarial learning framework for subject-wise domain generalization (DG).
• Integrated a contrastive discriminator with a UNet-based denoising module to identify and remove subject-related information.
• Employed adversarial training to enforce the extraction of subject-invariant features using low-count PET data realizations.

Main Results:

• The contrastive adversarial DG framework demonstrated superior denoising performance compared to conventional UNet.
• Outperformed cross-entropy-based adversarial DG methods in subject-wise denoising.
• Evaluated on 97 ¹⁸F-MK6240 tau PET studies, showing improved generalization across subjects.

Conclusions:

• The proposed contrastive adversarial DG framework effectively addresses subject-wise variations in PET data.
• Achieved enhanced denoising performance and generalizability for clinical PET applications.
• Offers a more reliable and trustworthy solution for DL-based PET image analysis.