Automated Lesion Segmentation in Whole-Body PET/CT - Multitracer Multicenter Generalization¶

🎬 Introduction¶

We invite you to participate in the third  autoPET Challenge. The focus of this year's challenge is to further refine the automated segmentation of tumor lesions in Positron Emission Tomography/Computed Tomography (PET/CT) scans in a multitracer multicenter setting.

Over the past decades, PET/CT has emerged as a pivotal tool in oncological diagnostics, management and treatment planning. In clinical routine, medical experts typically rely on a qualitative analysis of the PET/CT images, although quantitative analysis would enable more precise and individualized tumor characterization and  therapeutic decisions. A major barrier to clinical adoption  is  lesion segmentation, a necessary step for quantitative image analysis. Performed manually,  it's tedious, time-consuming and costly. Machine Learning offers the potential for fast and fully automated quantitative analysis of PET/CT images, as previously demonstrated in the first two autoPET challenges.

Building upon the insights gained in these challenges, autoPET III expands the scope to address the critical need for models to generalize across multiple tracers and centers. To this end, we provide participants access to a more diverse PET/CT dataset containing  images of two different tracers - Prostate-Specific Membrane Antigen (PSMA) and Fluorodeoxyglucose (FDG) acquired from two different clinical sites (Figure). In this challenge, we give participants the chance to let their models compete in two award categories. In award category one, participants are tasked with developing robust segmentation algorithms applicable to two different tracers. In award category two, the significance of data quality and preprocessing in algorithm performance is addressed. Here, participants are encouraged to enhance our baseline model using innovative data pipelines, fostering advancements in data-centric approaches to automated PET/CT lesion segmentation.

Join us in autoPET III to pave the way for robust deep-learning-based medical image analysis in PET/CT, optimizing diagnostics and personalizing therapy guidance in oncology. The autoPET III challenge is hosted at MICCAI 2024 and supported by the European Society for hybrid, molecular and translational imaging (ESHI). The challenge is the successor of autoPET and autoPET II.

Figure: The image depicts a comparison between two cohorts: the Fluorodeoxyglucose (FDG) cohort on the left and the Prostate-specific membrane antigen (PSMA) cohort on the right. The PSMA cohort exhibits lower image resolution and greater dataset heterogeneity compared to the FDG cohort.