EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Performance assessment of the winning autoPET III model on fluorodeoxyglucose positron emission tomography–computed tomography within and beyond its training domain across diverse cancer types

Maksym Fritsak, Hubert S. Gabryś, Maximilian Rokuss, Sebastian M. Christ, Nicolas Martz, Alina Paunoiu, Isabelle Opitz, Rolf Stahel, Martin Hüllner, Matthias Guckenberger, Stephanie Tanadini-Lang

Physics and Imaging in Radiation Oncology · 2026-11 · 101112

DOI: 10.1016/j.phro.2026.101112 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physics and Imaging in Radiation Oncology, 2026-11. Verfasst von Maksym Fritsak und 10 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background and purpose Reliable automated cancer lesion detection and segmentation remains a challenge, and existing models require extensive external evaluation. Training on metastatic disease may improve generalization beyond the training domain. Thus, we independently evaluated the top-performing autoPET III model within and beyond its training domain. Materials and methods The winning autoPET III model, trained on a heterogeneous metastatic dataset, was independently validated on FDG-PET/CT scans from patients with metastatic melanoma, oligo/non-metastatic lung cancer, head and neck cancer, soft-tissue sarcoma, and breast cancer, and patients in complete remission. Performance was evaluated at the scan and lesion levels within and beyond the training domain.

Results

The model demonstrated strong-to-moderate performance within its training domain for detection (F1 score 0.75 for melanoma and 0.66 for lung cancer) and segmentation (Dice score 0.61 for melanoma and 0.55 for lung cancer). Performance deteriorated for out-of-domain cancer types. False-positive findings occurred across all datasets, predominantly for lesions 1 cm 3 within (melanoma, lung cancer) and beyond (head and neck cancer) the training domain (ICC(3,1) ≥ 0.94). Total tumor burden estimation showed the highest agreement in melanoma (ICC(3,1) = 0.86).

Conclusions

The evaluated model shows potential as a semi-automated tool for lesion detection, segmentation, and lesion-level quantification in melanoma and lung cancer. However, due to the high rate of false-positive findings, including cancer-free scans, it is not considered suitable as a pan-cancer model for scan-level evaluation.

Zitieren

APA

Fritsak, M., Gabryś, H. S., Rokuss, M., Christ, S. M., Martz, N., Paunoiu, A., Opitz, I., Stahel, R., Hüllner, M., Guckenberger, M., & Tanadini-Lang, S. (2026). Performance assessment of the winning autoPET III model on fluorodeoxyglucose positron emission tomography–computed tomography within and beyond its training domain across diverse cancer types. Physics and Imaging in Radiation Oncology, 101112. https://doi.org/10.1016/j.phro.2026.101112

Vancouver

Fritsak M, Gabryś HS, Rokuss M, Christ SM, Martz N, Paunoiu A, et al. Performance assessment of the winning autoPET III model on fluorodeoxyglucose positron emission tomography–computed tomography within and beyond its training domain across diverse cancer types. Physics and Imaging in Radiation Oncology. 2026:101112. doi:10.1016/j.phro.2026.101112

BibTeX
@article{Fritsak2026,
  author = {Fritsak, Maksym and Gabryś, Hubert S. and Rokuss, Maximilian and Christ, Sebastian M. and Martz, Nicolas and Paunoiu, Alina and Opitz, Isabelle and Stahel, Rolf and Hüllner, Martin and Guckenberger, Matthias and Tanadini-Lang, Stephanie},
  title = {Performance assessment of the winning autoPET III model on fluorodeoxyglucose positron emission tomography–computed tomography within and beyond its training domain across diverse cancer types},
  journal = {Physics and Imaging in Radiation Oncology},
  year = {2026},
  pages = {101112},
  doi = {10.1016/j.phro.2026.101112}
}
Kostenlos starten →