A Synthetic-Geometry, Mlc-Aware Machine Learning Framework for Collimator Angle Selection In Single-Isocenter Multi-Target VMAT SRS
Abstract
Purpose
Collimator orientation in single-isocenter multi-target (SIMT) SRS significantly impacts multileaf collimator (MLC) modulation and overall plan quality. We developed a machine learning (ML) framework to identify optimal collimator angles prior to Treatment Planning System (TPS) optimization, effectively minimizing iterative planning and improving deliverability.
Methods
Five hundred synthetic SIMT cases (3–10 targets) with spherical and ellipsoidal targets were generated with varying spatial distributions within a 75mm radius. Candidate collimator angles (7°–173°, 187°–353°) were evaluated using a clinical four-arc template (couch: 0°, 315°, 45°, 90°). For each configuration, targets were projected into the beam’s-eye-view (BEV) across 68 sampled gantry angles. A "source" complexity score was calculated by a weighted aggregation of geometry-based metrics: connected-component island count, small-island pixel fraction, normalized BEV aperture area, perimeter-to-area ratio, and leaf-span statistics. BEV masks were discretized into 60 leaf-pair rows to model clinical MLC geometry and estimate multi-opening frequency. These features trained a gradient-boosted decision tree model to predict the composite score. Using a GroupShuffleSplit, data were partitioned into a 75/25% training/testing split by case ID to ensure an independent test set and prevent data leakage.
Results
On the independent test set, the model achieved a mean absolute error (MAE) of 0.381 and R2 of 0.998. Ranking performance demonstrated high clinical relevance: mean score difference between predicted and true optimal was 0.022. The optimal collimator angle was identified within the model’s top-1 prediction in 86.8% of cases and within the top-5 in 99.4%. This confirms the model effectively bypasses suboptimal geometries and narrows the search space for the TPS optimizer.
Conclusion
This MLC-aware framework enables accurate collimator selection prior to planning. By integrating high-resolution MLC geometry into a pre-optimization ML model, planners can standardize workflows, minimize planning variability, and improve delivery efficiency in SIMT VMAT SRS.