Generalizable CT-Free PET Attenuation and Scatter Correction for Pediatric Patients

A team of researchers from multiple Chinese institutions has developed the Generalizable PET Correction Network (GPCN), a deep learning model that eliminates the need for CT scans in PET attenuation and scatter correction for pediatric patients. Published on arXiv, the method addresses a critical clinical challenge: CT-based correction adds an average 10.8 mSv of radiation per scan—particularly harmful to children. Existing CT-free approaches degrade under scanner or radiotracer shifts, limiting real-world use. GPCN tackles this with a dual-domain architecture combining a multi-band contextual refinement module and a frequency-aware spectral decoupling module. The first module models pediatric anatomical variability via wavelet-based multiscale decomposition and long-range spatial context. The second performs coordinate-conditioned amplitude/phase refinement in the Fourier domain, explicitly separating invariant topological structures from domain-specific noise.

GPCN was trained and evaluated on 1085 whole-body pediatric PET scans acquired with two different scanners and five radiotracers. In both joint training and zero-shot cross-domain evaluations, GPCN outperformed representative baselines and maintained stable quantitative accuracy on unseen scanner-tracer combinations. The method was further validated through ablation studies, region-wise quantitative analysis, and downstream segmentation experiments. The authors have released source code, marking a step toward safer pediatric imaging by reducing radiation exposure without compromising diagnostic quality. For pediatric patients who may require multiple scans over time, this could significantly lower cumulative radiation risk.