Researchers unveil GLOW-FDG AI for cancer lesion segmentation
New model automates cancer detection across 5 major cancer types with 1,563 training scans
A team led by Maksym Fritsak from ETH Zurich and the University of Zurich has released GLOW-FDG, an open-source AI model designed to automate cancer lesion segmentation in whole-body PET/CT scans. The model leverages 1,563 training scans spanning breast cancer, lung cancer, head and neck cancer, and melanoma, validated against 185 external scans from independent institutions. GLOW-FDG demonstrated superior lesion detection performance compared to publicly available benchmarks while significantly reducing false positives.
The model's ability to quantify total tumor burden and total lesion glycolysis showed remarkable robustness across diverse patient cohorts, with performance metrics aligning closely with inter-expert variability among radiation oncologists. By eliminating the manual, time-consuming process of lesion delineation, GLOW-FDG addresses scalability challenges in cancer care, enabling more consistent and reproducible quantitative imaging biomarker extraction.
- GLOW-FDG is an open-source AI model for automated cancer lesion segmentation in PET/CT scans, built by researchers from ETH Zurich and the University of Zurich
- Trained on 1,563 scans across 5 cancer types and validated on 185 external scans, outperforming existing models in lesion detection with fewer false positives
- Achieves tumor burden quantification accuracy comparable to expert radiation oncologists, enabling scalable and reproducible cancer care
Why It Matters
GLOW-FDG could transform oncology workflows by automating tumor detection and quantification, reducing manual effort and improving consistency in cancer diagnosis and treatment planning.