A deep LSTM autoencoder-based framework for predictive maintenance of a proton radiotherapy delivery system

Area of Science:

• Medical Physics
• Machine Learning in Healthcare
• Proton Therapy

Background:

• Unscheduled downtime in proton Pencil Beam Scanning (PBS) systems disrupts patient treatment and negatively impacts outcomes.
• Conventional Quality Assurance (QA) programs identify operational performance but often fail to predict machine malfunctions.
• Early detection of potential machine failures is crucial for maintaining treatment continuity and efficacy.

Purpose of the Study:

• To propose and evaluate a Predictive Maintenance (PdM) approach for proton PBS systems using deep neural networks.
• To develop a framework for early detection of anomalous machine events that could lead to treatment interruptions.
• To improve the reliability and operational performance of proton therapy delivery systems.

Main Methods:

• Collected beam delivery log file data from daily QA at a proton therapy center.
• Developed a PdM framework using Long Short-Term Memory-based Stacked Autoencoder (LSTM-SAE) for unsupervised anomaly detection.
• Quantified anomalies using Mahalanobis distance (M-Score) and validated the model with clinical datasets using AUPRC and AUROC metrics.

Main Results:

• LSTM-SAE models demonstrated strong performance in predicting QA beam pauses, achieving high AUPRC and AUROC values.
• The model significantly enhanced the detection of treatment interruptions (up to 51.2-fold improvement) and clinical operational issues.
• The PdM approach showed substantial improvements in detecting rare anomalous machine events compared to baseline rates.

Conclusions:

• The developed deep LSTM-SAE-based framework effectively predicts anomalous machine events in proton PBS systems.
• This novel approach shows significant promise for enabling predictive maintenance in proton therapy.
• The findings suggest a pathway to reduce treatment interruptions and enhance patient safety and outcomes.