ESTRO 2025 - Abstract Book

S2525

Physics - Autosegmentation

ESTRO 2025

For validation, we selected 30 GBM cases with ground truth (GT) contours and a test-set of manually adjusted contours mimicking realistic geometric variation with a mean DSC of 0.57 and a mean HD of 8.10 mm compared to the GT. Clinical treatment plans were generated for both GT and the test-set. Their dosimetric differences determined which adjusted contours were critical for review. The test-set was taken as input for the QA approach to evaluate for critical segmentations (Figure 2). The results of the QA approach are compared with the validation data by sensitivity and specificity.

Results: Based on clinical plans, the validation data defined 171 out of 510 adjusted segmentations as critical. The QA approach flagged 164 of these. The overall sensitivity and specificity were 0.96 and 0.56 respectively. The dosimetric layers were more dominant in the decision with 225 unacceptable thresholds exceeded vs 78 and 68 for the DSC and HD respectively. Conclusion: The results of the QA approach on an OAR test-set in GBM cases, containing a high fraction of critical geometric variation, have high sensitivity. We anticipate that improvements in the dose prediction model lead to improved sensitivity while optimizing the decision criteria can improve specificity outcomes in true clinical data.

Keywords: QA model, dose prediction, contour review