Uncertainty-Aware AI Model Boosts PET/CT Lesion Segmentation by 10%

A team of researchers from multiple institutions (including Dartmouth and Yale) has published a preprint on arXiv (2606.10115) tackling a critical challenge in oncology: accurately segmenting lesions from whole-body PET/CT scans. Their uncertainty-aware framework extends the nnU-Net architecture with Bayesian ensembling to reduce training stochasticity, voxel-wise uncertainty quantification (epistemic + aleatoric decomposition), and uncertainty-augmented training to improve lesion detection. The model was trained and evaluated on two public datasets: AutoPET-III (1,611 scans) and Deep-PSMA (200 scans), covering FDG and PSMA tracers across multiple cancer types.

Key results show that Bayesian ensembling improves robustness and performance over deterministic nnU-Net on the unseen AutoPET-III test set. Uncertainty maps correlate well with misclassifications, particularly false positives. Uncertainty-augmented training improves lesion recovery at the cost of increased false positive volume, reflecting a precision-recall trade-off. To manage this trade-off, the authors propose a case-adaptive routing strategy that selects between the base and augmented models per case, further improving Dice scores. This is the first systematic study of uncertainty quantification in multi-tracer, pan-cancer PET/CT segmentation, offering a pathway to more clinically reliable AI tools for cancer staging.