1062 - Interpretable Artificial Intelligence Decision Support Framework for Online Adaptive Radiotherapy Quality Assurance

03:15pm - 03:20pm ET

Room 156

Presenter(s)

Jeff Winter, PhD - Princess Margaret Cancer Centre, Toronto, Ontario

I. Mansour¹, T. G. Purdie², C. McIntosh³, T. Tadic⁴, and J. Winter¹; ¹Radiation Medicine Program, Princess Margaret Cancer Centre, Toronto, ON, Canada, ²Princess Margaret Cancer Centre, Toronto, ON, Canada, ³Toronto General Hospital Research Institute, University Health Network, Toronto, ON, Canada, ⁴Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Toronto, ON, Canada

Purpose/Objective(s):

Treatment plan evaluation is complex and multifactorial. In online adaptive treatments, time constraints make conventional review processes difficult. Online adaptive MR-Linac workflows include adapt-to-position (ATP; rigid shift) and adapt-to-shape (ATS; full adaptation). Although ATS addresses non-rigid anatomical changes, it adds to the QA workload, and previous research has shown that ATS does not always outperform ATP plans.

We hypothesize that an interpretable artificial intelligence (AI) classifier can reduce QA burden by automatically identifying deviations from clinical practice and support efficient online decision-making.

Materials/Methods:

An isolation forest model (Curait Medical, Toronto, ON) generated deviation scores quantifying departure from a distribution of 94 clinically accepted high-quality training cohort of prostate cancer patients treated with 42.7 Gy in 7 plans. The deviation scores are expressed in terms of learned features that characterize the PTV dosimetry, OAR dosimetry, and the overall plan.

Patients treated using 42.7 Gy in 7 (N = 15) or 30 Gy in 5 (N=20) fractions on a 1.5T MRL were analyzed. A binary-classification model was developed using the deviation score to predict whether ATS or ATP was preferred; trained using Leave-One-Patient-Out Cross-Validation (LOOCV) with bootstrapped (N=500) confidence intervals. Separately, fractions were classified as ATS preferred if ATS had a higher conformality index than ATP.

Results:

For 42.7 Gy in 7, LOOCV sensitivity was 96.6% (95% CI: 91.2–100.0%) with Negative Predictive Value (NPV) 92.2% (80.0–100.0%). Model efficiency was 52.4% indicating half of the plans identified as preferred did not require additional review. ROC AUC was 0.779.

Applied without retraining to the independent 30 Gy in 5 cohort, sensitivity remained 93.8% (87.3–98.6%) with NPV 76.6% (53.3–94.1%). Efficiency decreased to 37.5%, with ROC AUC 0.695. Although discrimination degraded under distributional shift, high error-detection sensitivity was preserved.

Conclusion:

An interpretable anomaly-detection AI framework quantified deviations from high-quality adaptive plans and determining ATS vs. ATP preferences. Trained on a 42.7 Gy in 7 workflows, the model demonstrated in excess of 96% sensitivity in determining ATS vs. ATP preference, and when applied to an independent 30 Gy in 5 cohort to examine performance under distributional shift the model maintained >93%. The interpretable rule structure enables clinically actionable thresholds and supports prospective integration as real-time decision support in MR-guided adaptive radiotherapy.

Metric	30 Gy in 5 (N=20, CI = 95%)	42.7 in 7 (N=15, CI = 95%)
Sensitivity	93.8% (87.3–98.6)	96.6% (91.2–100.0)
NPV	76.6% (53.3–94.1)	92.2% (80.0–100.0)
Efficiency	37.5% (22.6–54.0)	52.4% (37.5–66.7)
F1 Score	0.824 (0.750–0.885)	0.829 (0.761–0.895)
ROC AUC	0.695 (0.568–0.819)	0.779 (0.674–0.880)
C-Index (Concordance)	0.653 (0.585–0.719)	0.726 (0.651–0.792)