To evaluate an unsupervised deep-learning model with an interpretable tool for anomaly detection in IMRT treatment plans.
Approach:
Data Collection: Six hundred IMRT treatment plans were collected from our institute, and relevant features were extracted.
Model Development: A standard autoencoder (AE) was used to build an anomaly detection model from normal plans.
Anomaly Detection: Anomalies were identified based on exceptionally high reconstruction errors.
Interpretation Method: A feature perturbation interpretation (FPI) method was introduced to rank the impacts of features on detected anomalies.
Comparison of Methods: The effectiveness of the FPI method was compared with local-DIFFI and SHAP methods, as well as four classic detection models: local discrete factor (LOF), density-based spatial clustering of applications with noise (HDBSCAN), one class of support vector machines (OC-SVM), and principal component analysis (PCA).
Key Findings:
The AE achieved the best detection performance with an AUC of 0.98.
Average accuracy, precision, and F1 score for the AE were 0.91, 0.61, and 0.74, respectively.
The top five features impacting detected anomalies were consistent across FPI, local-DIFFI, and SHAP methods.
Interpretation:
The AE model is effective for identifying anomalous plans, and the FPI method provides insights into feature impact.
Limitations:
The study relies on a specific clinical database, which may limit generalizability to other settings.
The performance of the model may vary with different datasets or clinical settings.
Conclusion:
The combination of the AE model and FPI method provides a framework for automatic plan review in radiotherapy.
