Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Multi-Site Evaluation of the AI-Assisted Auto Segmentation Quality for a CBCT Based Online Adaptive Radiotherapy System

B Meng*, M Dohopolski, S Jiang, M Lin, B Cai, University of Texas Southwestern Medical Center, Dallas, TX

Presentations

PO-GePV-M-298 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Novel clinical online adaptive radiotherapy (ART) system allows modification of the treatment plan to account for daily setup variations, internal organ motion, and deformation. The adaptive workflow and plan quality depends on efficient and accurate automated segmentations of targets and organs at risk (OARs). Herein, using clinical treatment data, we quantitatively examined the capability of a novel online ART system for delineation of various tumor sites’ anatomy using AI and automated deformable segmentation algorithms

Methods: Online daily ART or ART on-demand treatments were delivered using Varian Ethos system based on daily on-board cone-beam CT (CBCT) images. Twelve patients were selected providing 43 adaptive treatment fractions for 3 different tumor sites including pelvis, thorax, and head and neck (HN). Segmentation analyses comparing initial automated segmentations to physician modified treatment contours using volumetric dice similarity coefficient (DSC) were performed and compared using one-way ANOVA with a post hoc Tukey test for multiple comparisons.

Results: Target segmentation quality showed the lowest DSC compared to other OAR contours, with mean DSC 0.87, 0.59, and 0.88 for HN, pelvis and thoracic, respectively. In Ethos system, pelvis OAR segmentations using convolutional neural network (CNN) algorithm (mean DSC = 0.92) showed significantly lower DSC than OAR segmentations for other sites using deformable image registration (DIR) (mean DSC = 0.99), with p-value <0.0003. Additionally, no correlation was observed for segmentation quality between target and other OARs (r = 0.34).

Conclusion: These results suggest that unsupervised automated target segmentation does not provide accurate target delineation, but automated OARs segmentation provides clinical acceptable OAR contours. Although the algorithm performance requires optimization, AI segmentation algorithm could assist in complex segmentations using daily CBCT images. This study provides important insight for clinical adoption of online ART treatment system, automated segmentation algorithm development and clinical decision-making.

Funding Support, Disclosures, and Conflict of Interest: SJ received grants from Varian

Keywords

Cone-beam CT, Image Guidance, Segmentation

Taxonomy

TH- RT Interfraction Motion Management: X-ray projection/CBCT-based

Contact Email

Share: