3071 - Predicting Adaptive Need in Helical Pancreatic Radiotherapy Using Fraction-Level Dosimetry and Machine Learning

Purpose/Objective(s): To quantify fraction-to-fraction dose variability in helical pancreatic radiotherapy, establish objective dose-based thresholds for adaptive intervention, and evaluate the feasibility of predicting fractions that would benefit from adaptation using machine learning.

Materials/Methods: Five pancreatic cancer patients treated with helical delivery on Radixact systems, each prescribed 3600 cGy in 15 fractions (75 fractions total), were retrospectively analyzed. Using a vendor-provided retrospective adaptive evaluation platform (Adapt LTE, Accuray), the daily kVCT from each fraction was deformably registered to the planning CT, generating propagated contours and a daily merged CT, on which fractional dose was recalculated under two conditions: IGRT-only delivery and adaptive re-optimization. Both dose sets were evaluated against reference planning goals. To identify objective adaptive triggering criteria, a self-organizing neural network map was applied to fraction-level dosimetric features derived from PTV coverage and organs-at-risk (OARs). Based on the identified thresholds, a cross-validated quadratic support vector machine (SVM) model was developed to predict fractions that would benefit from adaptive re-optimization. Model performance was assessed using the cross-validated area under the receiver operating characteristic curve (CV-AUC).

Results: Substantial fraction-to-fraction dose variability was observed with IGRT-only helical delivery. Clustering using the self-organizing neural network map identified adaptive trigger thresholds based on deviations from planning goals: =4.7% reduction in PTV V(3600 cGy) relative to the 95% coverage goal, and =1.7%, 2.4% and 2.2% increases in D(0.03 cc) for stomach, small bowel, and duodenum, respectively. Fractions meeting at least two of these criteria were classified as candidates for adaptation based on dosimetric benefit. Using these thresholds, 53.9% of daily fractions were identified as likely to benefit from adaptive re-optimization. The quadratic SVM model accurately predicted these fractions, achieving a CV-AUC of 0.96, indicating excellent discriminative performance.

Conclusion: Fraction-level dose analysis showed substantial variability in helical pancreatic radiotherapy when relying on IGRT-only delivery. Objective, data-driven thresholds derived from unsupervised clustering effectively identified fractions likely to benefit from adaptive intervention. A quadratic SVM model demonstrated strong potential for predicting adaptive need on a per-fraction basis. Further validation using larger datasets, bias assessment, and mixed-effects modeling is warranted to establish robustness and clinical generalizability.