Author: Elizabeth L. Covington, Robert T. Dess, Charles S. Mayo, Michelle L. Mierzwa, Dan Polan, Jennifer Shah, Claire Zhang 👨🔬
Affiliation: University of Michigan, Department of Radiation Oncology, University of Michigan 🌍
Purpose: Auto-segmentation improves contour consistency and standardization in radiation therapy but may introduce variations from current practices, potentially impacting treatment outcomes and toxicity. Successful clinical integration requires collaboration across stakeholders. This study introduces a phase-based approach to efficiently implement auto-segmentation into clinical practice.
Methods: A three-phase workflow was designed: 1) Catalytic Phase: Initial dataset to assess infrastructure, develop workflows and automation tools, and establish contour quality metrics; 2) Feasibility Testing Phase: Limited dataset to refine workflows, validate tools, and gather feedback; 3) Final Evaluation Phase: Larger dataset to finalize workflows, assess clinical performance, identify limitations, and establish safety/QA processes. Auto-contours were generated using a vendor-provided model for all organs-at-risk (OARs). Contour quality was assessed using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and clinical dose metrics.
Results: Auto-contours demonstrated geometric differences compared to manual contours, with median DSCs [25th, 75th percentiles] for OARs across disease sites reported as follows: Head&Neck (0.69 [0.50, 0.82]), Prostate (0.53 [0.49, 0.92]), Lung (0.75 [0.68, 0.93]), Liver (0.81 [0.63, 0.94]), Breast (0.72 [0.68, 0.95]), and Brain (0.47 [0.41, 0.74]). The correlation between geometric discrepancies and dosimetric impacts varied by site and metric. In prostate cases (62–68 Gy in 20 fractions), the mean bowel D0.1cc difference was 14 Gy, strongly correlating with DSC (correlation coefficient: -0.98). In Head&Neck cases (70 Gy in 35 fractions), the differences of submandibular gland mean dose ranged from 0–8 Gy, showing weak correlation with DSC (correlation coefficient: 0.22) but strong correlation with HD (correlation coefficient: -0.92).
Conclusion: A phase-based approach utilizing a standardized dataset facilitates the evaluation of geometric accuracy variations in delineated structures during auto-segmentation. These variations can result in different dosimetric impacts across anatomical structures and disease sites, with varying correlations depending on the geometric metrics applied. Clinical impacts and associated risks should be carefully assessed.