Author: Kathryn J. Dess, Martha M. Matuszak, Dan Polan 👨🔬
Affiliation: University of Michigan 🌍
Purpose: A recent survey demonstrated 18 of 20 top academic institutions have implemented auto-segmentation. Studies to date have focused on geometric contour changes and dosimetric differences using the original optimized plan. With increasing reliance on auto-segmentation for adaptive workflows, we quantified the dosimetric impacts of re-optimization using differing structure representations.
Methods: 29 hepatocellular carcinoma (HCC) patients treated on a prospective adaptive SBRT trial were anonymized. Two commercial auto-segmentation platforms (P1,P2) generated heart, liver and stomach contours. Using structures from P1 and P2, plans were re-optimized to the new geometry, preserving the original planning-target-volume, beam arrangement, and optimization parameters. New plans were copied back to the original datasets for dosimetric analysis. Original contours and treated plans were considered the reference standard. Dice similarity coefficients (DSC) compared original contours/plans to auto-segmentation counterparts. DSCs were scored: High (>=0.9), Good (0.8-<0.9), Moderate (0.7-<0.8), and Unacceptable (<0.7).
Results: For P1, median DSCs for heart, liver, and stomach were 0.93, 0.94, and 0.84, respectively; for P2, 0.95, 0.95, and 0.86. Heart and liver were High or Good for 99% of P1/P2 structures. Stomach showed greater variability, with 16% and 14% of structures Moderate or Unacceptable scores for P1/P2. Minimal dose variability was observed for heart (average D0.5cc difference for P1: -0.03Gy (range -2.05-1.43Gy) and for P2: 0.21Gy (-2.06-1.99Gy)) and liver (average mean difference for P1: -0.06Gy (-0.59-0.53Gy) and for P2: -0.02Gy (-0.98-0.91Gy)). Conversely, stomach doses varied (average D0.5cc difference for P1: 0.64Gy (-3.82-11.72Gy); and for P2: 0.68Gy(-2.78-12.72Gy)).
Conclusion: For HCC patients treated with SBRT, two commercial auto-segmentation tools produced high DSCs with minimal dose impact for nearby structures including heart and liver. However, when evaluating the stomach, a structure with larger geometric variability, dosimetric variance is magnified. This demonstrates that evaluation of auto-segmentation should consider geometric differences as well as dosimetric changes associated with re-optimization.