Virtual patient-specific QA with DVH-based metrics

Abstract

We demonstrate a virtual pretreatment patient-specific QA (PSQA) procedure that is capable of quantifying dosimetric effect on patient anatomy for both intensity modulated radiotherapy (IMRT) and volumetric modulated arc therapy (VMAT). A machine learning prediction model was developed to use linear accelerator parameters derived from the DICOM-RT plan to predict delivery discrepancies at treatment delivery (defined as the difference between trajectory log file and DICOM-RT) and was coupled with an independent Monte Carlo dose calculation algorithm for dosimetric analysis. Machine learning models for IMRT and VMAT were trained and validated using 120 IMRT and 206 VMAT fields of prior patients, with 80% assigned for iterative training and testing, and 20% for post-training validation. Various prediction models were trained and validated, with the final models selected for clinical implementation being a boosted tree and bagged tree for IMRT and VMAT, respectively. After validation, these models were then applied clinically to predict the machine parameters at treatment delivery for 7 IMRT plans from various sites (61 fields) and 10 VMAT multi-target intracranial radiosurgery plans (35 arcs) and compared to the dosimetric effect calculated directly from trajectory log files. Dose indices tracked for targets and organs at risk included dose received by 99%, 95%, and 1% of the volume, mean dose, percent of volume receiving 25%-100% of the prescription dose. The average coefficient of determination (r² ) when comparing intra-field predicted and actual delivery error was 0.987 ± 0.012 for IMRT and 0.895 ± 0.095 for VMAT, whereas r² when comparing inter-field predicted versus actual delivery error was 0.982 for IMRT and 0.989 for VMAT. Regarding dosimetric analysis, r² when comparing predicted versus actual dosimetric changes for all dose indices was 0.966 for IMRT and 0.907 for VMAT. Prediction models can be used to anticipate the dosimetric effect calculated from trajectory files and have potential as a "delivery-free" pretreatment analysis to enhance PSQA.

Keywords: AI; IMRT QA; artificial intelligence.

FIGURE 1

Workflow for virtual PSQA process in which a machine learning prediction of linear accelerator parameters is used to calculate dose on the patient anatomy prior to treatment delivery. Accuracy of the prediction model is then assessed after the first (and subsequent) fraction(s) using trajectory files. PSQA, patient‐specific QA

FIGURE 2

Data and methods for training and validation of machine learning model to predict machine parameters at treatment delivery from a new DICOM‐RT treatment plan

FIGURE 3

Prediction model performance for IMRT and VMAT for validation dataset. Each point represents a single field (n = 24 for IMRT, 41 for VMAT). RMS, root mean square

FIGURE 4

Change in dosimetric indices after accounting for machine parameters at treatment delivery for the 7 IMRT and 10 VMAT cases used for the dosimetric analysis. Dosimetric effects predicted pretreatment by the prediction model are highly correlated (r ² > 0.9) with the effect derived using the posttreatment trajectory file

FIGURE 5

Change per dosimetric index after accounting for machine parameters at treatment delivery for 7 IMRT and 10 VMAT cases used for dosimetric analysis