Program Information
Error Patterns of PET Based Manual Target Definition in the Lung
F Yang1*, P Johnson1 , L Young2 , (1) Department of Radiation Oncology, University of Miami School of Medicine, Miami, FL, (2) Department of Radiation Oncology, University of Washington, Seattle, WA
Presentations
SU-H3-GePD-J(A)-1 (Sunday, July 30, 2017) 4:00 PM - 4:30 PM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: Target volume delineation is one of the most crucial stages yet the weakest link in radiation therapy. Uncertainty and variability in target definition are well recognized; however the error patterns of target delineation were little known and rarely quantified. The present study is aimed to quantitatively assess the error patterns of target delineation for PET imaged lung cancer in a setting with complete ground truth.
Methods: Image data being used consisted of 26 synthetic PET datasets created by using the anthropomorphic Zubal phantom and the Monte Carlo based SimSET computational package. Each dataset included a different PET-positive lesion which varied in terms of shape, heterogeneity, and location inside the lung. Target contours were provided by 10 physicians and the contour accuracy was evaluated by using Dice, FPD (False Positive Dice), FND (False Negative Dice), AUC (Area Under ROC), KAPPA (Cohens Kappa), Rand Index, ADJRIND (Adjusted Rand Index), ICCORR (Interclass correlation), VOLSMTY (Volumetric Similarity), MAHLNBS (Mahalanobis Distance), VARINFO (Variation of Information), GCOERR (Global Consistency Error), Sensitivity, Specificity, and Precision.
Results: In terms of volume overlap, manual contouring showed an average agreement of 85% with the ground truth and tended to over-segment the lesions with mean FPD of 24%. Shape of the segmented volumes matched fairly well to the ground truth with average VOLSMTY of 89% while on voxel basis the positive predictive value of the segmented volumes is only 79%. Rand index with a mean of 0.99 while declining to 0.86 after corrected for chance shows there exists a random component in manual contouring.
Conclusion: Quantification of segmentation error in manual contouring of PET positive lesion in the lung reveals general patterns in what otherwise might be thought of as simple randomness. These findings allow for the formation of new hypotheses towards seeking to achieve less subjective target delineation.
Contact Email: