AI-Based Psqa Prediction for Single-Isocenter Multiple-Target Radiosurgery
Abstract
Purpose
Single-isocenter multiple-target (SIMT) stereotactic radiosurgery (SRS) presents significant dosimetric challenges for patient-specific quality assurance (PSQA). Standard complexity metrics often fail to predict deliverability for these complex plans. This study develops and evaluates AI-based models to pre-emptively identify potential PSQA failures.
Methods
A total of 89 VMAT arcs from 25 clinical SIMT SRS plans with lesion numbers ranging from 1 to 21, delivered on a C-arm LINAC equipped with a high-definition multi-leaf collimator (MLC), were selected: 66 for model development and 23 for validation. PSQA was performed for each arc using Electronic Portal Imaging Device (EPID) dosimetry. The PSQA failure classification threshold was defined as a gamma passing rate below 95% using a 3%/2mm criterion with a 10% threshold. We investigated 15 distinct complexity features, including two newly developed metrics Single Leaf Opening Fraction (SLF) and Small Aperture Area (SAR), which target the unique geometric challenges of SIMT SRS arcs. A stepwise feature selection strategy was implemented, and the model performance was evaluated across 11 machine learning models.
Results
Evaluation of the 11 AI models revealed that Subspace Discriminant and Medium KNN algorithms demonstrated superior predictive performance. The Subspace Discriminant model achieved the highest overall performance, yielding an Area Under the Curve (AUC) of 0.88 and an accuracy of 0.87. The Medium KNN model also performed robustly with an AUC of 0.84 and an accuracy of 0.87. Feature importance analysis identified that SAR features exhibited the highest contribution across each model, while SLF demonstrated significant performance contribution in five distinct models.
Conclusion
AI-based models, particularly Subspace Discriminant and KNN architectures, reliably forecast PSQA outcomes for SIMT SRS plans. Noval plan complexity enhance the prediction power. Incorporating these models into the clinical workflow enables early detection of dosimetric discrepancies, improving planning efficiency and quality control in radiosurgery.