Clinically Oriented Contour Evaluation Using Dosimetric Indices Generated From Automated Knowledge-Based Planning.

<bold>Purpose: </bold>Geometric indicators of contouring accuracy suffer from lack of clinical context in radiation therapy. To provide clinical relevance, treatment plans should be generated from the candidate contours, but manual planning could introduce confounding variations. Therefore, our objectives in this study were as follows: (1) determine the feasibility of using automated knowledge-based planning as an objective tool to generate dosimetric parameters for contour evaluation, (2) evaluate the correlation between geometric indices and dosimetric endpoints, and (3) report the dosimetric impact of multiple observations of head and neck target and organ-at-risk (OAR) volumes contoured by resident physicians.<bold>Methods and Materials: </bold>Twenty-two resident physicians contoured the clinical target volumes, parotids, and cochleae for a nasopharyngeal cancer case, and expert-generated contours were defined as the gold standard for this study. A validated knowledge-based planning routine generated 67 treatment plans with various resident/gold-standard and target/OAR combinations. Dosimetric indices (dose to hottest 98% volume of planning target volume, and mean dose of OAR) were calculated on gold-standard contours. Commonly used geometric indices (Dice coefficients, Hausdorff maximum/mean/median distances, volume differences, and centroid distances) were also calculated. R2 quantified the correlation between geometric and dosimetric indices.<bold>Results: </bold>The correlation between geometric and dosimetric indices was weak (R2 < 0.2 for 61% of the correlations studied-77 of 126) and inconsistent (no single geometric index consistently exhibited superior/inferior correlation with dosimetric endpoints). The lack of consistent correlations between geometric and dosimetric indices resulted in the inability to define any geometric index thresholds for clinical acceptability. Geometric indices also exhibited a high propensity for false positives and false negatives as a classifier of dosimetric impact. Finally, we found substantial interresident contour variation, whether quantified using geometric or dosimetric indices, with significant negative dosimetric impact should these contours be used clinically.<bold>Conclusions: </bold>Contour variation among resident physicians significantly affected dosimetric endpoints, highlighting the importance of resident education in head and neck anatomy delineation. Whenever available, dosimetric indices generated from automated planning should be used alongside geometric indices in radiation therapy contouring studies. [ABSTRACT FROM AUTHOR]