AutoPET Challenge on Fully Automated Lesion Segmentation in Oncologic PET/CT Imaging, Part 2: Domain Generalization.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Automated lesion segmentation in positron emission tomography/computed tomography (PET/CT) is crucial for cancer diagnosis and treatment monitoring.
• The autoPET challenge aims to advance machine learning (ML) models for this task, focusing on real-world deployment challenges.

Purpose of the Study:

• To evaluate the domain generalization capabilities of ML-based segmentation models trained on single-source PET/CT data.
• To assess model performance across diverse clinical variations, including different institutions, pathologies, populations, and tracers.

Main Methods:

• The second autoPET challenge involved training ML models on 1,014 whole-body 18F-FDG PET/CT scans.
• Models were tested on 200 samples from 5 distinct clinical domains.
• Performance was quantified using Dice Similarity Coefficient, false-positive volume, and false-negative volume.

Main Results:

• Generalization from a single data source remains a significant challenge, with out-of-domain performance substantially deteriorating.
• The best model achieved a Dice score of 0.5038, but performance dropped on pediatric and PSMA data.
• Error analysis indicated issues with physiologic uptake and detection of small or low-uptake lesions.

Conclusions:

• The autoPET challenge highlights the limitations of current automated PET/CT segmentation models in handling data variability.
• There is a critical need for diverse, multi-domain public datasets to improve the robustness and clinical applicability of these algorithms.