Automated Lesion Segmentation in Whole-Body PET/CT - The human-AI frontier¶

📰 News¶

April 1st:
Challenge opens

🎬 Introduction¶

We invite you to participate in the fifth edition of the autoPET Challenge, focusing on interactive, clinician-in-the-loop lesion segmentation in whole-body PET/CT.

Positron Emission Tomography combined with Computed Tomography (PET/CT) plays a central role in oncologic imaging, supporting diagnosis, staging, therapy response assessment, and disease monitoring. In current clinical practice, radiologists assess tumor burden by visually inspecting PET/CT scans and identifying changes in lesion size and distribution using standardized criteria. Despite the substantial effort required for this manual process, only a limited subset of lesions is typically evaluated using simplified one-dimensional measurements. As a result, a large fraction of the rich quantitative and spatial information contained in PET imaging remains underutilized. In addition, manual assessment is subject to inter-observer variability and may not fully capture complex disease patterns.

Automated lesion detection and segmentation have the potential to enable more comprehensive and reproducible analysis of tumor burden. While recent advances in deep learning have led to significant progress in automated whole-body PET/CT segmentation, important challenges remain. These include limited robustness to domain shifts across scanners, tracers, and centers, difficulties in distinguishing physiological uptake from pathological lesions, and reduced performance in complex or low-contrast cases. Most importantly, fully automated solutions often fail to align with clinical expectations, where expert interpretation and correction remain essential.

In clinical reality, lesion segmentation is therefore not a purely automatic task, but an interactive and iterative process in which clinicians refine and correct algorithmic predictions. autoPET V embraces this paradigm shift by introducing a benchmark for interactive lesion segmentation in whole-body PET/CT. Participants are asked to develop algorithms that generate an initial segmentation and iteratively improve it in response to sparse corrective input provided during inference in the form of scribbles targeting false-positive and false-negative regions.

A central feature of autoPET V is the evaluation under two complementary interaction regimes. First, a standardized simulated-interaction setting enables scalable and fully reproducible benchmarking by generating corrective scribbles in a controlled manner. Second, a clinician-driven interaction setting incorporates real expert annotations, capturing realistic human–AI correction behavior in clinical workflows. All submitted methods are evaluated consistently across both regimes in a newly harmonized multi-center test cohort, allowing the challenge to assess both algorithmic adaptability and clinical relevance.

Join us for autoPET V to move beyond static automation and establish a new benchmark for adaptive, robust, and clinically meaningful human–AI collaboration in PET/CT imaging.

The autoPET V challenge is hosted at MICCAI 2026: TODO
in collaboration with the European Society of Radiology (ESR) under "AI-based assessment of PET imaging for oncology" and supported by the European Society for hybrid, molecular and translational imaging (ESHI). The challenge is part of the autoPET series and the successor of autoPET, autoPET II, autoPET III, and autoPET/CT IV.