Quantifying AI Assisted Auto-Segmentations Performance for a Clinical Online Adaptive Radiotherapy System in Multiple Treatment Sites

Abstract

<h3>Purpose/Objective(s)</h3> The growing translation of online adaptive radiotherapy (ART) into clinical applications has sparked the need for a comprehensive assessment of artificial intelligence (AI) and automated segmentation performance employed by these novel treatment systems. Several groups have reported their initial experiences and evaluations using retroactive simulated data. Herein, incorporating clinical treatment data, we quantitatively examined the capability of a novel CBCT based online ART system for anatomy delineation of various tumor sites using AI and automated deformable segmentation algorithms. <h3>Materials/Methods</h3> Twelve patients treated with online ART were selected for 3 different tumor sites including pelvis, thorax, and head and neck (HN). The novel CBCT based ART system generates auto-segmentations for targets, influencer organs at risk (OARs) and other OARs. Influencers are site specific critical OARs which guide target generations. Convolutional neural networks (CNN) were used for pelvis influencer segmentations. Influencers for thorax and HN sites, initial target contours, and other non-influencer OARs were segmented using elastic deformable image registration (DIR). The initial auto-segmentations were extracted for targets, influencers, and high impact OARs. Contour qualities were quantified using volumetric dice similarity coefficient (DSC) against physician modified treatment contours, and compared using one-way ANOVA with a post hoc Tukey test for multiple comparisons. <h3>Results</h3> A total of 43 ART fractional data was analyzed. Physicians spent the most effort in modifying auto segmentation of target structures, with mean DSC of 0.59, 0.88 and 0.87 for pelvis, thoracic and HN sites, respectively. For influencer OARs, CNN generated pelvis influencers (mean DSC = 0.92) was less accurate than DIR generated influencers for thoracic and HN (mean DSC = 0.98 and 0.99) sites, with p-value < 0.005 and 0.0005. A table showing auto segmentation quality for influencers is included. All other high impact OAR auto segmentations required minimal revision with mean DSC of 0.99 for pelvis, thorax and HN sites. <h3>Conclusion</h3> Overall, CBCT based ART demonstrated high quality of auto segmentation for most OARs. However, manual revisions are still needed for targets and OARs with dosimetric importance, especially for treatment sites with inferior soft tissue contrast in CBCT. This study provides important insight for clinical adoption of online ART treatment system, automated segmentation algorithm development and clinical decision-making.