Proof of Concept of Anatomical Independence in Total-Body PET Using AI-Based CT-Free Multi-Organ Segmentation

This study investigates the potential anatomical independence of PET imaging by assessing the feasibility of PET-only multi-organ segmentation using a deep learning-based approach. In this retrospective study, the model was trained on a dataset of non-corrected PET images acquired from two commercially available total-body PET/CT scanners across two different centers. Ground-truth labels were generated using Computed Tomography (CT) images as input to TotalSegmentator and subsequently validated by two board-certified physicians, each with over a decade of experience. A U-Net-like model was trained on the non-corrected PET images (n = 938) and validated on two internal (n = 68) and nine external (n = 382) test datasets from diverse scanners, sites, and tracers. The model was designed to segment 17 anatomical structures; we report performance for nine clinically relevant organs: brain, heart, kidneys, liver, lungs, pancreas, spleen, thyroid gland, and urinary bladder. Across the two internal test sets, the mean Dice score was 0.828 (95% CI, 0.816–0.839) for scans acquired on the Biograph Vision Quadra (Siemens Healthineers) in Bern (n = 34) and the uExplorer (United Imaging) in Shanghai (n = 34). On external datasets, the model achieved a mean Dice of 0.787 (95% CI, 0.782–0.791) across five cross-scanner cohorts and 0.726 (95% CI, 0.698–0.755) across four cross-tracer cohorts. Despite its potentially controversial nature, this methodology may broaden the application of PET imaging by avoiding redundant CT scans, which is particularly beneficial in multi-tracer studies or longitudinal monitoring.