Anomaly detection in radiotherapy plans using deep autoencoder networks.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Radiotherapy treatment plans require rigorous safety and quality checks by human experts.
• Manual review of these plans can be time-consuming and may miss subtle flaws.

Purpose of the Study:

• To develop an automated method for identifying flawed radiotherapy treatment plans.
• To utilize an unsupervised learning approach, specifically an autoencoder, for this quality assurance task.

Main Methods:

• Extracted features from treatment plans and used them to train an autoencoder model.
• Identified questionable plans based on high reconstruction error, indicating deviation from normal plan distribution.
• Compared the autoencoder's performance against Local Outlier Factor (LOF), HDBSCAN, OC-SVM, and PCA using 576 breast cancer treatment plans.

Main Results:

• The autoencoder achieved superior performance with an Area Under the Curve (AUC) of 0.9985, significantly outperforming the next best algorithm (LOF at 0.9535).
• The autoencoder maintained 100% recall, with average accuracy and precision of 0.9658 and 0.5143, respectively.
• Demonstrated effectiveness in distinguishing flawed plans from a large set of normal plans.

Conclusions:

• An autoencoder effectively automates the identification of questionable radiotherapy treatment plans.
• This unsupervised learning method eliminates the need for labeled data, simplifying the quality assurance process.
• The autoencoder presents a promising solution for efficient and automatic plan checking in radiotherapy settings.