ESTRO 2024 - Abstract Book

S3006

Physics - Autosegmentation

ESTRO 2024

Lotte H Land 9 , Svetlana Borissova 6 , Irene D Larsen 10 , Azza A Khalil 1,2 , Lise S Mortensen 1 , Hjørdis H Schmidt 1 , Stine S Korreman 3,2 , Martin Kyndt 11 , Ditte S Møller 1,2 1 Aarhus University Hospital, Department of Oncology, Aarhus, Denmark. 2 Aarhus University, Department of Clinical Medicine, Aarhus, Denmark. 3 Aarhus University Hospital, Danish Center for Particle Therapy, Aarhus, Denmark. 4 Aalborg University Hospital, Department of Oncology & Clinical Cancer Research Center, Aalborg, Denmark. 5 Aalborg University, Department of Clinical Medicine, Aalborg, Denmark. 6 Copenhagen University Hospital Herlev-Gentofte, Department of Oncology, Herlev, Denmark. 7 Copenhagen University Hospital Rigshospitalet, Department of Oncology, Copenhagen, Denmark. 8 University Hospital of Southern Denmark, Vejle Hospital, Department of Oncology, Vejle, Denmark. 9 Odense University Hospital, Department of Oncology, Odense, Denmark. 10 Aalborg University Hospital, Department of Oncology, Aalborg, Denmark. 11 MIM Software Inc., Cleveland, OH, USA

Purpose/Objective:

Manual delineation of organs at risk (OARs) in radiation therapy (RT) planning is time-consuming and subject to intra- and inter-individual variation. Commercial algorithms for OAR delineation are available. However, if clinicians disagree with the delineations produced, additional time is spent on corrections or even re-delineations. Therefore, we aimed to develop an algorithm for automated delineation of thoracic OARs based on national consensus definitions.

Material/Methods:

Twelve members of the Danish Lung Cancer Oncology Group and the Danish Esophagogastric Cancer Group, including ten oncologists and two radiographers from seven different RT departments, reached national consensus on delineation of nine thoracic OARs (trachea, bronchi, heart, aorta, left/right lung, esophagus, spinal canal, and spinal cord). Ground truth OAR sets were manually delineated on contrast-enhanced 4D-CT scans from 10 lung cancer patients included in the NARLAL 2 study (NCT02354274) on heterogeneous FDG-guided dose-escalated RT. Five independent expert contour sets were delineated per patient, resulting in 50 sets of OARs. A STAPLE contour was derived from the five manual delineations for each of the ten ground truth patients. Subsequently, a training set of thoracic OARs was delineated on 100 patients with lung cancer. A bespoke U-Net AI algorithm (b-AI) was trained in collaboration with MIM software. This b-AI was applied on the initial 10 CT scans to create auto-delineated OARs (figure 1). Likewise, a commercially available Siemens algorithm (c-AI) for thoracic OARs was applied. For quantitative performance evaluation, 1 mm Surface Dice Similarity Coefficient (SDSC) was calculated to compare b-AI, manual expert contour and c-AI delineations to the STAPLE ground truth contour. We report median SDSC [inter quartile range (IQR)] for all observers and all 10 patients. For qualitative assessment, seven reviewers performed blinded, graded evaluations, with at least three independent visual reviews per set of OARs (table 1). Finally, resulting OAR doses from clinical dose-escalated treatment plans for b-AI and STAPLE were compared. Mean dose and D1cm3 (the 1cm3 receiving the highest dose) for each OAR were plotted as scatter plots (b-AI vs. STAPLE) and compared using Pearson correlation coefficient.

Results:

For the b-AI, the median SDSCs ranged from 0.63 to 0.96. For manual expert delineations, the SDSC ranged from 0.80 to 0.97. IQRs were largely overlapping for the two methods (table 1). The c-AI, if available, had more variable SDSCs,