Clinical Report: Assessment of a Transparent Deep Learning Framework for Automated Review of IMRT Plans
Overview
This study evaluates an unsupervised deep-learning model for identifying anomalies in intensity-modulated radiation therapy (IMRT) plans. The model achieved an area under the curve (AUC) of 0.98.
Background
Intensity-modulated radiation therapy (IMRT) is a complex treatment modality that requires thorough quality assurance to ensure patient safety and treatment efficacy. Manual review of IMRT plans is time-consuming and relies heavily on the expertise of experienced planners. The integration of automated methods, particularly deep learning, aims to enhance the efficiency and reliability of the plan review process.
Data Highlights
Metric
Value
AUC
0.98
Average Accuracy
0.91
Precision
0.61
F1 Score
0.74
Key Findings
The autoencoder (AE) model outperformed other classic detection models in identifying anomalies in IMRT plans.
The area under the curve (AUC) for the AE model was 0.98.
Average accuracy, precision, and F1 score for the AE were 0.91, 0.61, and 0.74, respectively.
The feature perturbation interpretation (FPI) method effectively ranked the impacts of features on detected anomalies.
The top five features impacting anomalies were consistent across FPI, local-DIFFI, and SHAP methods.
Clinical Implications
The findings indicate that the AE model may assist in the automated review of IMRT plans.
Conclusion
The study demonstrates that the AE model, combined with the FPI method, offers a solution for anomaly detection in IMRT plans.
