KEY POINTS
- The main Methods section reports a retrospective dataset of 576 plans, including 558 normal and 18 anomalous plans, although the abstract inconsistently states that 600 plans were collected.
- Each plan contained two tangential conformal fields and two intensity-modulated fields. The model analysed 30 parameters, including segment numbers, source-to-skin distances, collimator positions, gantry angles and monitor units.
- The autoencoder was trained only on normal plans and used reconstruction error to identify plans that deviated substantially from institutional patterns.
- The decision threshold was selected to preserve 100% sensitivity, accepting additional false-positive alerts to avoid missing potentially unsafe anomalous plans.
- The autoencoder achieved an area under the curve of 0.98, with average accuracy of 0.91, precision of 0.61 and F1 score of 0.74.
- Principal-component analysis was the strongest comparator but remained inferior, with an area under the curve of 0.96, accuracy of 0.81, precision of 0.42 and F1 score of 0.58.
- Feature perturbation, SHAP and Local-DIFFI identified the same five leading contributors to anomalous classifications: three monitor-unit parameters and segment counts from both intensity-modulated fields.
CLINICAL TAKEAWAY
An interpretable anomaly detector could complement rule-based checking by flagging unusual plan configurations and showing which parameters generated the warning. It cannot replace physics review because only 18 anomalous cases, one institution and one restricted planning technique were evaluated.