ESTRO 2025 - Abstract Book

S3429

Physics - Machine learning models and clinical applications

ESTRO 2025

high dimensional space of parsed log sequences using DRAIN parser. Then, a Long Short-Term Memory Variational Auto Encoder (LSTM-VAE), trained in semi-supervised mode, learnt the normal behavior in the data and detected abnormal sequences by observing abnormally high reconstruction errors. Conformal Prediction was used to determine the reconstruction error threshold while ensuring statistical confidence in anomaly detection. This approach was tested on a log dataset extracted from 14 different TomoTherapy® machines, focusing on the Multileaf Collimator (MLC) subsystem and on 20 bumper-pack-related failures. Results: This method generates a high-dimensional representative space where normal log sequences form identifiable clusters (blue in Figure 1). The LSTM-VAE model detects sequences distant from these normal clusters as anomalies (in orange, red, brown and black in Figure 1). This study shows that detected anomalies are linked to an increasing number of known MLC-problem messages (e.g., leaf bounces and overtravel) when statistical confidence increases. The model also predicts shutdown-related sequences, characterized by greater lengths, as anomalies with confidence. Temporal analysis highlights increasing proportions of abnormal sequences when approaching an observed MLC failure.

Conclusion: This study developed a comprehensive anomaly detection approach for TomoTherapy® log data, generalizable to any log type and radiotherapy machine subsystem. Further research is required to refine detection, incorporating complementary data for machine context, and develop alerting tools to warn of impending failures. Investigating the model's robustness and explainability could enhance its reliability and foster user confidence.

Keywords: TomoTherapy®, Artificial Intelligence, Log Data