AutoPET: Automated Lesion Segmentation in Whole-Body PET/CT

Le résumé fourni par la source

Positron emission tomography (PET) combined with computed tomography (CT) plays a central role in oncologic staging, response assessment, and disease monitoring. Currently, diagnostic assessment is performed by radiologists (i.e. human observers) on PET/CT scans through the detection of changes in tumour size and distribution using standardised criteria. Despite the highly time-consuming nature of this manual task, only unidimensional (diameter) evaluations of a subset of tumour lesions are used to assess tumour dynamics. Additional quantitative evaluation of PET information would potentially allow for more precise and individualized diagnostic decisions. Besides the risk of inter-observer variability, the manual approach only extracts a diminutive proportion of the morphologic tumour data derived from images, thereby neglecting valuable, significant, and prognostic information. Automation of tumour detection and segmentation may enable faster and more comprehensive information and data extraction. While recent editions of the autoPET challenge series have demonstrated substantial progress in automated whole-body lesion segmentation across tracers and centers, persistent challenges remain. These include sensitivity to domain shifts, physiological tracer uptake mimicking disease, complex lesion phenotypes, and the resulting limited alignment between fully automated segmentations and clinical expectations. In clinical practice, lesion segmentation will therefore not be a static, fully automated task, but a dynamic interpretive process in which human expertise remains essential. autoPET V addresses this gap by introducing a unified benchmark for interactive, clinician-in-the-loop lesion segmentation in whole-body PET/CT. Participants are tasked with developing algorithms that iteratively refine lesion segmentations in response to sparse corrective input, represented as scribbles targeting false positives and false negatives. The challenge evaluates a single task under two complementary interaction regimes using scribbles: a standardized simulated-interaction setting that enables scalable and reproducible benchmarking, and a clinician-driven interaction setting based on real expert input that captures realistic human-AI correction behavior. All submissions are evaluated under both regimes and are eligible for corresponding award categories. A second pillar of autoPET V is a fully new, harmonized, multi-center test cohort spanning diverse tracers, scanners, disease burdens, and a substantial fraction of lesion-absent studies. Building on lessons learned from previous autoPET editions and related challenges, including DeepPSMA, PET quantification is harmonized following QIBA-aligned SUV normalization procedures to improve cross-site comparability. Evaluation metrics are revised to better disentangle semantic tumor burden estimation from instance-level lesion delineation and to assess not only final segmentation quality but also adaptation efficiency over interaction steps. To support reproducibility and broad participation, autoPET V is accompanied by structured onboarding materials, including Jupyter notebook tutorials, example data loaders, and detailed task descriptions illustrating interactive segmentation workflows and the representation of corrective scribbles. In addition, a moderated question-and-answer forum facilitates clarification of best practices established in previous autoPET editions. These resources are designed to lower the technical entry barrier while promoting transparent and comparable algorithm development. Overall, autoPET V aims to advance PET/CT AI beyond static automation toward robust, adaptive, and clinically meaningful human-AI collaboration.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.