Combining autoencoder with clustering analysis for anomaly detection in radiotherapy plans.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Radiotherapy treatment plans require rigorous quality assurance to prevent errors.
• Identifying error-prone treatment plans is crucial for patient safety.
• Automated methods can enhance the efficiency and accuracy of quality assurance.

Purpose of the Study:

• To develop an unsupervised anomaly detection method for identifying suspicious and potentially error-prone radiotherapy treatment plans.
• To compare the effectiveness of principal component analysis (PCA) and autoencoder (AE) based dimensionality reduction techniques within this anomaly detection framework.

Main Methods:

• Utilized k-means clustering for initial normal plan grouping in the training set (400 normal plans).
• Calculated distances between testing set samples (158 normal, 19 abnormal plans) and cluster centers.
• Compared PCA and autoencoder (AE) for dimensionality reduction to assess their impact on anomaly detection accuracy.

Main Results:

• The AE-based anomaly detection model achieved higher sensitivity (94.7%) and Area Under the ROC Curve (AUC) (0.90) compared to PCA (84.2% sensitivity, 0.81 AUC).
• AE demonstrated improved specificity (69.0%) over PCA (64.6%).
• The combination of AE and k-means clustering proved effective in distinguishing abnormal plans.

Conclusions:

• Unsupervised learning is a viable approach for detecting anomalies in radiotherapy treatment plan features.
• Autoencoder-based dimensionality reduction enhances the accuracy of anomaly detection.
• The AE and k-means clustering combination offers an automated solution for identifying abnormal clinical treatment plans.