TL;DR
This paper introduces a new metric, LAAM, to quantify the complexity of stereotactic VMAT treatment plans. They demonstrate that this metric, along with others, can be used with machine learning models to accurately predict the success of patient-specific quality assurance (QA) tests before they are performed, potentially streamlining treatment planning and improving patient safety.
What Problem Does This Paper Address?
Delivering radiation therapy, especially stereotactic volumetric modulated arc therapy (VMAT), requires highly precise treatment plans. Patient-specific quality assurance (QA) is crucial to verify that the planned dose is accurately delivered. A common QA metric is the gamma passing rate (GPR), which compares the delivered dose to the planned dose. However, complex treatment plans can lead to higher uncertainties and a greater chance of QA failure, potentially delaying treatment. Predicting these QA failures before measurement is challenging because it requires understanding the inherent complexity of the treatment plan itself. This paper aims to address this by developing quantitative measures of plan complexity and using them to predict GPRs.
Methods
The researchers developed a new metric called Leaf Travel Modulation Complexity Score (LAAM). This metric combines several aspects of a VMAT plan:
- Leaf Travelling (LT) distance: How much the multi-leaf collimator (MLC) leaves move between control points.
- Ratio of average aperture area over jaw area (AAJA): How much of the area defined by the jaws is actually used by the MLC.
- Aperture Perimeter Variability (APV): How much the shape of the MLC aperture changes.
- Total monitor units (MU) per beam: The total dose delivered by the beam.
They calculated 27 different complexity metrics in total, including LAAM and established ones like the Index of Modulation (M) and Modulation Complexity Score (MCS). These metrics were then used to train two machine learning models: Random Forest and Gradient Boosting Regressor. The goal was to predict the GPRs based on these complexity features. They also performed ROC analysis to determine optimal thresholds for predicting pass/fail scenarios.
Datasets
The study utilized a dataset of 301 clinical stereotactic VMAT plans, comprising 594 beams. These plans were generated using a Varian Edge LINAC and covered various treatment sites, including brain, lung, spine, and liver. The patient-specific QA measurements were performed using Varian's portal dosimetry system with the EPID. The GPRs were calculated using a 2%/2 mm criteria with a 5% dose threshold.
Key Insights and Ideas
- Complexity is quantifiable and predictive: The study strongly suggests that plan complexity isn't just an abstract concept but can be quantified by specific metrics, and these metrics are indeed correlated with QA outcomes.
- LAAM shows promise: The newly proposed LAAM metric demonstrated the highest correlation with GPRs and performed well in ROC analysis, indicating its potential as a valuable complexity indicator.
- Machine learning is a powerful tool for QA prediction: Random Forest and Gradient Boosting models achieved very high prediction accuracy for GPRs, demonstrating that these methods can effectively learn the relationship between plan complexity and QA success.
- Early warning system for QA failures: By predicting GPRs, these methods can act as an early warning system, allowing physicists to identify potentially problematic plans before actual measurements, saving time and resources.
- Site-specific complexity variations: The study highlighted that plan complexity metrics can vary significantly between different treatment sites, suggesting that future models might benefit from site-specific adjustments.
Key Results
- High prediction accuracy: The Random Forest model achieved 98.74% prediction accuracy with a mean absolute error of 1.23%, and the Gradient Boosting model achieved 98.71% accuracy with a mean absolute error of 1.25% using five-fold cross-validation.
- LAAM's strong correlation: The LAAM metric showed the highest Spearman's rank correlation coefficient with GPRs at 0.508 (p < 0.001).
- Top complexity features: LAAM, leaf travelling distance (LT), AAJA, LT modulation complexity score (LTMCS), and index of modulation were identified as the top five most important complexity features for predicting GPRs.
- LAAM's ROC performance: LAAM achieved the best performance in ROC analysis with an AUC value of 0.801 and a threshold value of 0.365, offering a good balance between true positive and false positive rates.
Limitations and Future Work
- Limited dataset size and scope: The study acknowledges that the models were trained on data from a single institution. The accuracy and applicability of these models in other institutions and for different treatment sites need further testing and evaluation.
- Site-specific analysis: While the study noted variations in complexity metrics across treatment sites, a detailed correlation analysis for each site was not performed. The authors suggest exploring this in future work.
- Generalizability to non-stereotactic plans: The current models are specific to stereotactic VMAT plans. Future work could involve extending these methods to predict GPRs for non-stereotactic plans, such as those for nasopharyngeal cancer.
- Incorporating complexity into optimization: The authors suggest that plan optimization algorithms should incorporate plan complexity to minimize uncertainties and improve plan control, which could lead to better overall plan quality.
Summary updated