Patient‐specific PTV margins for liver stereotactic body radiation therapy determined using support vector classification with an early warning system for margin adaptation

Purpose An adaptive planning target volume (PTV) margin strategy incorporating a volumetric tracking error assessment after each fraction is proposed for robotic stereotactic body radiation therapy (SBRT) liver treatments. Methods and materials A supervised machine learning algorithm employing retrospective data, which emulates a dry‐run session prior to planning, is used to investigate if motion tracking errors are <2 mm, and consequently, planning target volume (PTV) margins can be reduced. A fraction of data collected during the beginning of a treatment course emulates a dry‐run session (mock) before planning. Twenty features are calculated using mock data and used for support vector classification (SVC). A treatment course is labeled as Class 1 if the maximum root‐mean‐square radial tracking error for all remaining fractions is below 2 mm, or Class 2 otherwise. We evaluate the classification using fivefold cross‐validation, leave‐one‐out cross‐validation, 500 repeated random subsampling cross‐validation, and the receiver operating characteristic (ROC) metric. The classification is independently cross‐validated on a cohort of 48 treatment plans for other anatomical sites. A per fraction assessment of volumetric tracking errors is performed for the standard 5 mm PTV margin (PTVstd) for courses predicted as Class 2; or for a margin reduced by 2 mm (PTVstd‐2mm) for those predicted as Class 1. We perturb the gross tumor volume (GTV) by the tracking errors for each x‐ray image acquisition and calculate the fractional GTV voxel occupancy probability (Pi) inside the PTV for each treatment fraction i. For treatment courses classified as Class 1, an early warning system flags treatment courses having any Pi < 0.99, and the subsequent treatments are proposed to be replanned using PTVstd. Results The classification accuracies are 0.84 ± 0.06 using fivefold cross‐validation, and 0.77 when validated using an independent testing set (other anatomical sites). Eighty percent of treatment courses are correctly classified using leave‐one‐out cross‐validation. The sensitivity, precision, specificity, F1 score, and accuracy are 0.81 ± 0.09, 0.85 ± 0.08, 0.80 ± 0.11, 0.83 ± 0.06, and 0.80 ± 0.07, respectively, using 500 repeated random subsampling cross‐validation. The area under the curve for the ROC metric is 0.87 ± 0.05. The four most important features for classification are related to standard deviations of motion tracking errors, the linearity between the target location and external LED marker positions, and marker radial motion amplitudes. Eleven of 64 cases predicted to be of Class 1 have 0.96 < Pi < 0.99 for each treatment fraction, and require replanning using PTVstd. In comparison, the PTVstd always covers the perturbed GTVs with Pi > 0.99 for all patients. Conclusions Support vector classification is proposed for the classification of different motion tracking errors for patient courses based on a mock session before planning for SBRT liver treatments. It is feasible to implement patient‐specific PTV margins in the clinic, assisted with an early warning system to flag treatment courses that require replanning using larger PTV margins in an adaptive treatment strategy.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.