Technical feasibility of AI-based longitudinal multi-organ volumetry on low-dose PET/CT over 13 years

This study aimed to (i) compare TotalSegmentator Fast and Normal inference modes on low-dose PET/CT, (ii) evaluate whether CT denoising improves segmentation longitudinal consistency, and (iii) examine whether the selected pipeline can capture long-term within-participant volumetric trends over approximately 13 years.

This retrospective study drew from a PET/CT screening registry of 15,597 examinees. In Phase 1, 10 participants (137 scans) were selected for pipeline optimization: Fast versus Normal modes and five denoising methods were compared using detrended within-participant CV. In Phase 2, 30 participants aged ≥60 years (337 scans) were analyzed. Volumetric changes in six structures were modeled using a linear mixed-effects model with false discovery rate correction.

Fast mode demonstrated similar detrended longitudinal variability to Normal mode across the six analyzed structures (detrended CV range: 2.26-9.32%), with approximately 10-fold faster processing. No tested denoising method meaningfully improved the selected serial-variability metrics over raw CT. In the primary random-intercept model, all six structures showed significant population-average volumetric trends (all p < 0.001, FDR q < 0.001): autochthon (paraspinal) muscles (- 0.72%/year), iliopsoas (- 0.74%/year), liver (- 0.55%/year), kidneys (- 0.49%/year), pancreas (- 0.97%/year), and aorta (+ 0.94%/year). In the random-slope sensitivity analysis, all six structures-including the iliopsoas and aorta-remained statistically significant (all RS p < 0.01) when individual trajectory slopes were accommodated.

Fast mode on raw low-dose PET/CT provided the most favorable speed-stability trade-off for TotalSegmentator-based multi-organ volumetry. Long-term within-participant volumetric trends were detectable over a 13-year follow-up, supporting technical feasibility but not yet establishing biological attribution or clinical utility.