ESTRO 2024 - Abstract Book

S3063

Physics - Autosegmentation

ESTRO 2024

7. Mylonas, Adam. Contour Alignment Tool. (2023): https://github.com/ACRF-Image-X-Institute/contour-alignment tool

1538

Digital Poster

AI tumor delineation for all breathing phases in early NSCLC

Luis R de la O Arevalo, Nanna M Sijtsema, Johannes A Langendijk, Lisanne V van Dijk, Robin Wijsman, Peter M.A. van Ooijen

University Medical Center Groningen, Radiotherapy, Groningen, Netherlands

Purpose/Objective:

Delineating GTV (Gross Tumor Volume) and ITV (Internal Target Volume) for early-stage lung tumors is essential in Stereotactic Body Radiation Therapy (SBRT). The ITV, necessary to make the treatment plan robust for breathing motion, is normally generated from manually contoured GTVs in all breathing phases [1], which is very time consuming. Deep learning algorithms expedite the process with automated contouring, ensuring consistency and efficiency. The purpose of this research is to streamline the delineation process, by developing a deep learning algorithm that delineates automatically the GTV in all breathing phases of the 4DCT to assemble an ITV automatically for early-stage Non-Small Cell Lung Cancer (NSCLC) patients.

Material/Methods:

The dataset included 222 early-stage NSCLC patient records, all who received Stereotactic Body Radiotherapy (SBRT) between 2013-2021. For each patient a 4D-CT scan was obtained, containing ten reconstructed breathing phases including a 3D-average scan of the full 4D-CT scan. The Gross Tumor Volume (GTV) contours were manually delineated using the 50% breathing phase by a lung radiation oncologist. The Internal Target Volume (ITV) was created by manually expanding the 50% phase GTV contour such that the ITV contour encompasses the tumor on all breathing phases. The dataset was partitioned as: 124 patients (56%) and 31 patients (14%) were designated for the training and validation cohort, respectively, all of which received treatment from 2013 to 2018. The hold-out test set contained 67 patients (30%), all of which were treated from 2019 to 2021. For the deep learning approach, we utilized two models, SwinUNetR and Dynamic UNet (DynUnet), both from the Monai library. These models were trained and optimized separately and in combination (SwinUNetR+DynUnet). The selection of the best-performing network for 50% GTV delineation was based on metrics such as the Dice Score (DSC), 3 mm Surface Dice Score (SDSC), and Hausdorff distance 95th percentile (HD95). The selected best-performing network was then used to delineate GTVs in all breathing phases of the 4DCT scans within the test set. The predicted GTV contours across all phases were used to create ITVs using two different methods: 1) incorporating all breathing phases and 2) focusing solely on the maximum inspiration and expiration phases (peak-to-peak). These two predicted ITVs were subsequently compared with the manually delineated ITV as ground truth. A visual representation of the methodology can be found in Figure 1