Meta-PET: A graininess-modulated test-time adaptive network for low-count PET denoising

Fast low-count PET imaging shortens acquisition time, enhances patient comfort, and improves clinical workflow efficiency. These benefits have attracted increasing attention in both clinical practice and research. However, reducing the count level to accelerate acquisition degrades image quality, motivating extensive studies on improving low-count PET denoising. Most existing methods are developed for a single count level and lack the flexibility to handle multiple or arbitrary count conditions within one unified model. Joint multi-count training can leverage richer information across different count levels but often leads to domain shifts and performance compromises. Although condition-aware strategies improve generalization by incorporating noise-level information during training, their inference process typically remains parameter-fixed, limiting their ability to adapt explicitly to the noise statistics of each individual unseen input. Toward this challenge, we propose Meta-PET, an adaptive PET denoising framework that combines the strengths of multi-count learning and test-time adaptation. Inspired by model-agnostic meta-learning, Meta-PET is trained to acquire parameters that can be rapidly adapted to new noise conditions through limited self-supervised updates. Rather than beginning from random weights, the model utilizes a joint-learned prior obtained from multi-count training, providing a more structured and generalizable starting point for meta-training scheme. During inference, Meta-PET performs inner-loop self-supervised optimization to align its parameters with the noise statistics of each individual input, enabling flexible and data-specific adaptation without retraining. Moreover, the proposed graininess-aware module directs the model toward noise-sensitive regions while constraining adaptation to fewer tunable parameters, making Meta-PET more efficient and stable during adaptation. Comprehensive experiments demonstrate that Meta-PET achieves quantitative accuracy and visual fidelity comparable to individually trained single-count models, while maintaining superior consistency across multiple count levels and representative existing methods.