Detection of the failed-tolerance causes of electronic-portal-imaging-device-based in vivo dosimetry using machine learning for volumetric-modulated arc therapy: A feasibility study

TL;DR

This study explores using machine learning to identify the causes of dosimetry errors in radiation therapy, showing promising results for some error types.

Contribution

A novel machine learning approach is proposed to classify causes of dosimetry failure in volumetric-modulated arc therapy.

Findings

ML models achieved over 90% accuracy for multileaf collimator position and monitor unit variation errors.

Lower accuracy (around 66%) was observed for lateral position, pitch, and roll rotation errors.

All models showed AUC values over 0.7, indicating acceptable classification performance.

Abstract

When electronic-portal-imaging-device (EPID)-based in vivo dosimetry (IVD) identifies dose tolerance failures, the cause of the failures should be evaluated. This study aimed to develop a machine-learning (ML) model to classify the cause of EPID-based IVD failures in volumetric-modulated arc therapy (VMAT) treatment.

Twenty-three prostate VMAT plans were used to recalculate the dose distribution in homogeneous phantom images as no-error (NE) plans. Errors in the randomized multileaf collimator (RMLC) position, monitor unit (MU) variation, lateral position, pitch rotation, and roll rotation were simulated. The IVD results of the NE plans and introduced errors were obtained using EPIgray software. Support vector machines (SVMs) were used to develop ML models for each error. The accuracy percentage, F1-score, and area under the receiver operating characteristic (ROC) curve (AUC) were used to evaluate models’ performances. The models were verified using five additional plans with an Alderson Rando phantom.

The models obtained accuracies of over 90% and F1-scores of 0.9 for the RMLC position and MU variation. For lateral position, pitch rotation, and roll rotation errors, the accuracies were 66.1%, 65.2%, and 66.8%, and the F1-scores were 0.66, 0.65, and 0.67, respectively. The AUCs for all the errors were over 0.7. Additionally, the model verification results consistently classified EPIgray data for all the error types.

The developed ML models classified the causes of the failed tolerance of the EPID-based IVD.

Figures