ESTRO 2024 - Abstract Book

S1820

Clinical - Lung

ESTRO 2024

Open-source AI-based networks for automatic segmentation of NSCLC on 3D and 4D PET/CT images

Gianluca Radicioni 1,2 , Dejan Kuhn 1,3 , Tobias Fechter 1,3 , Dimos Baltas 1,3 , Michael Mix 4 , Ursula Nestle 5 , Anca-Ligia Grosu 1,2 , Luis Martì-Bonmati 6 , Eleni Gkika 1,2 , Montserrat Carles 6 1 German Cancer Research Center (DKFZ), Partner Site Freiburg, German Cancer Consortium (DKTK), Freiburg, Germany. 2 University Medical Center Freiburg, Department of Radiation Oncology, Freiburg, Germany. 3 University Medical Center Freiburg, Department of Radiation Oncology, Division of Medical Physics, Freiburg, Germany. 4 University Medical Center Freiburg, Department of Nuclear Medicine, Freiburg, Germany. 5 Kliniken Maria Hilf GmbH Moenchengladbach, Department of Radiation Oncology, Moechengladbach, Germany. 6 La Fe Health Research Institute, Biomedical Imaging Research Group (GIBI230-PREBI) and Imaging La Fe node at Distributed Network for Biomedical Imaging (ReDIB) Unique Scientific and Technical Infrastructures (ICTS), Valencia, Spain

Purpose/Objective:

Positron Emission Tomography (PET) has demonstrated its benefit for tumor detection, staging and treatment of lung cancer. Respiratory motion compensation (4D) protocols showed improvement in the reproducibility and reliability of PET metrics. We developed open-source algorithms for automatic tumor segmentation on PET and Computed Tomography (CT), with (4D) and without (3D) respiratory motion compensation, with the aim of introducing them in the clinical practice.

Material/Methods:

We trained the nnU-Net with segmentations manually defined by two radiation oncologists (ground truth, GT) for PET and CT scans of lung cancer patients already treated in prospective protocols and patients enrolled in retrospective analyses (8%). One additional expert contoured the 3D- and 4D-CT sets for the 8 patients for the internal validation of the CT algorithm. 3 additional experts contoured each of the time-bins for the retrospective cohort employed in the internal evaluation of 4D-PET algorithm. The development of the PET-algorithm involved 810 4D-images (70% for training with 5-fold cross-validation, 30% for testing) and an external validation with 19 3D images from 9 centers. The same training method with 100 3D-images was applied for the development of the CT algorithm with the test involving 80 4D-images and 27 3D-images.

Results:

For PET segmentation, the Dice Similarity Coefficient (DSC) was 0.74 ± 0.06 for 4D-PET, an improvement of 19% compared to the DSC between GT and the additional experts. For external validation, the DSC (3D-PET) was 0.82 ± 0.11. For CT segmentation, the DSC (3D-CT) was 0.63 ± 0.34, an improvement of 15% relative to experts, and DSC(4D-CT) was 0.61 ± 0.28, an improvement of 4%. For external validation, DSC (3D-CT) was 0.59 ± 0.27 with a positive predictive value of 0.77 ± 0.27.

Conclusion: