ESTRO 2024 - Abstract Book

S3031

Physics - Autosegmentation

ESTRO 2024

Purpose/Objective:

The proximal bronchial tree (PBT) is considered an OAR during thoracic radiotherapy 1 . However, the knowledge of the dose-response relationships for conventionally fractionated RT is limited 2 . One obstacle to perform dose response studies might be that it is time consuming to manually contour an organ that has not been routinely delineated in the past. Automatic segmentation of the PBT could make such studies easier to conduct. The purpose of this study was to evaluate the performance of four different methods for automatic segmentation of the PBT.

Material/Methods:

We selected 20 atlas patients and 20 test patients from patients previously treated with RT for NSCLC. The treatments were conventionally fractioned with a prescribed dose of 60-70 Gy. The atlas patients were treated in 2017 and were selected to have an even distribution of gender and lung volumes. The test patients were treated 2002-2016 and were selected to have an even distribution of gender, lung volumes, treatment year and slice thickness in the CT image series used for treatment planning. The PBT was manually contoured in ARIA (Version 16.0, Varian Medical Systems, Palo Alto, CA, USA) by an oncologist on the atlas and test patients according to RTOG 0618 guidelines. In MICE Toolkit (Version 2020.2.1 [BETA], NONPI Medical AB, Umeå, Sweden), deformable registrations using elastix were performed to register each test patient’s CT image series to all 20 atlases and the PBT contours in the atlases were transferred to the test patient. The 20 generated PBT segmentations for each test patient were fused using two different methods, with simultaneous truth and performance level estimation (STAPLE) and a probability threshold of 50% as well as with majority voting. In RayStation (Version 9B, RaySearch Laboratories, Stockholm, Sweden), a custom atlas set was created with the 20 atlas patients. Deformable registrations between each test patient and the 20 atlases were performed using the inherent function for atlas-based segmentation in RayStation. The function fuses the 20 generated PBT segmentations for each test patient with a method including majority voting. In Velocity (Version 4.1, Varian Medical Systems), a custom atlas set was created as well. Using the inherent function for atlas-based segmentation in Velocity, one of the 20 atlases considered the best match was selected for each test patient. The selection was based on rigid registrations between the considered test patient and the atlases. For each test patient, a deformable registration between the test patient and the selected atlas patient was performed and followed by a local deformable registration around the PBT before transferring the structure from the atlas patient to the test patient. The four methods were evaluated in Eclipse (Version 16.01.10, Varian Medical Systems). The geometrical indices Dice similarity coefficient (DSC), Hausdorff distance (HD) as well as absolute and relative difference in volume were used to compare the test patients’ automatic segmentations of the PBT with the manual contours. Furthermore, the difference in mean dose and near-maximum dose (D 2% ) was evaluated.

Results:

The geometrical agreement between automatic segmentations of the PBT and manual contours varied among the evaluated methods while the dosimetrical agreement was similar (all within 5% of PD) (Figure 1). RayStation performed best considering the mean and SD for all indices except D 2% . MICE with STAPLE and Voting systematically overestimated and underestimated the volume of the PBT, respectively. The mean HD were >16 mm for all methods