The rapid advancements in PET technology, coupled with the need for accurate and efficient imaging, necessitate the development of robust and generalizable methods for CT-free attenuation and scatter correction (ASC). Deep learning offers a promising solution, but exhibits limited performance when tested in diverse clinical settings and varying imaging conditions. We propose a few-shot fine-tuning paradigm that enables efficient adaptation of models from a source domain to a new target domain. Our backbone network incorporates statistical modulation to extract domain-specific distribution information and employs pixel-wise factor scaling modeling to disentangle ASC factor maps from input images. On a large and diverse dataset of 1539 subjects across multiple tracers, scanners, and centers, we evaluate model performance under single-tracer training, multi-tracer joint training, and few-shot adaptation strategies. Although joint training demonstrates strong performance on known tracers, the proposed few-shot adaptation approach, CrossPET-Adapt, excels at adapting to unseen domains with minimal data, outperforming joint training. This method significantly reduces radiation exposure and data requirements, offering a rapid and robust solution for CT-free PET ASC in varied clinical environments.
Wang et al. (Thu,) studied this question.