Cross-Round Aggregation in Interactive Whole-Body PET/CT Lesion Segmentation: a Training-Free Study Alongside a Baseline Submission to autoPET V

Abstract

autoPET V scores interactive whole-body PET/CT lesion segmentationby the area under the Dice-versus-interaction-round curve, so thetrajectory across rounds matters as much as the final state. Wesubmit the organisers’ nnU-Net interactive baseline withoutmodification, and report alongside it an offline study of training-freestrategies that reuse predictions from earlier interaction rounds. Onthe single worked example released with the challenge repository, arunning union of predictions across rounds raises the AUC of Dicefrom 3.5141 to 3.5484 and converts a non-monotone Dice trajectoryinto a monotone one, which is precisely what an area-under-curvemetric rewards. Every tightening or smoothing aggregate we testedwas worse than the union, and one apparently reasonable variant —deleting whole connected components touched by a backgroundscribble — was catastrophic (AUC 2.7795), because underaccumulation the lesion region merges into a single largecomponent. Decomposing the union’s residual error shows that 78.6 % of it consists of voxels the model never predicted in any round,which bounds what any inference-time aggregation can achieve andis our reason for not carrying this into the submission. All numberscome from one case, one scribble strategy and one tracer; they arereported as a hypothesis-generating observation, not as a validatedresult.