Group-Aware Registration for Lesion-Level Quantitative Motion Correction in Respiratory-Gated PET/CT Biomedical
Hui Zhou1,2, Longxi He1, Yangsheng Hu1
1Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, China.
None:
Respiratory-gated PET/CT generates phase-resolved biomedical imaging data, but respiratory motion can cause spatial mismatch and lesion-level quantitative instability, especially for small thoracoabdominal lesions. This study proposes SCAR-Net, a Similarity-Constrained Adaptive Respiratory Registration Network, for retrospective phase-to-reference correction of respiratory-gated 18F-FDG PET/CT. SCAR-Net treats motion correction as a quantitative stability problem in phase-resolved PET/CT rather than only a generic registration task. It combines sampled group-aware feature encoding with adaptive group-attentive modulation to represent structured respiratory deformation and enhance motion-sensitive correspondence. The method was evaluated using controlled respiratory simulations and a retrospective two-center clinical cohort of 100 patients, jointly assessing lesion-level SUV repeatability, spatial correspondence, image similarity, deformation plausibility, center-stratified performance with B-spline FFD as a conventional reference, cross-dataset testing without target-domain fine-tuning, ablation behavior, and computational efficiency. In the independent clinical test set, lesion-level evaluation included 20 patients, 43 independent PET-avid lesions, and 245 evaluable lesion-phase pairs. In small-lesion phase pairs, SCAR-Net reduced median phase-to-reference variability to 6.45% for |ΔSUVmax| and 4.73% for |ΔSUVmean|, and increased median lesion Dice from 0.55 to 0.72. In large-lesion phase pairs, SCAR-Net achieved a post-correction Dice of 0.89 and remained competitive for SUV repeatability. Descriptive center-stratified analysis showed a consistent lesion-level performance pattern across the two clinical acquisition settings. These findings suggest that SCAR-Net can improve phase consistency and quantitative stability in respiratory-gated PET/CT, with the clearest benefit observed in small-lesion assessment. Downstream clinical endpoints, such as diagnostic accuracy, tumor staging, PERCIST-based response assessment, and patient outcomes, were not evaluated and require future prospective validation.
