Complexity-based unsupervised machine learning for patient-specific VMAT quality assurance.
Savino Cilla1, Carmela Romano1, Pietro Viola1
1Medical Physics Unit, Responsible Research Hospital, Campobasso, Italy.
Medical Physics
|August 24, 2025
Summary
Unsupervised machine learning identified three distinct clusters in radiotherapy plan complexity, aiding in predicting patient-specific quality assurance outcomes and improving treatment efficiency. This approach helps optimize radiotherapy planning and verification processes.
Area of Science:
- Medical Physics
- Radiotherapy Technology
- Machine Learning in Healthcare
Background:
- Patient-specific quality assurance (PSQA) is critical for safe, accurate radiotherapy.
- Increasing use of intensity-modulated techniques necessitates greater operational efficiency in PSQA.
- Machine learning (ML) shows promise for predicting PSQA outcomes, but unsupervised methods for analyzing plan complexity are underexplored.
Purpose of the Study:
- To explore unsupervised ML methods for identifying patterns and groupings in PSQA data.
- To analyze radiotherapy plan complexity using clustering algorithms.
- To investigate the relationship between plan complexity metrics and PSQA results.
Main Methods:
- Analysis of 1329 VMAT pretreatment verification datasets from 660 patients.
- Utilized Modulation Complexity Score (MCS) and dynamic log-files.
- Applied three unsupervised clustering algorithms: agglomerative hierarchical clustering (AHC), K-means (KM), and Gaussian mixture models (GMM).
- Validated cluster assignments using supervised models on an external cohort.
Main Results:
- Clustering analysis consistently identified three optimal clusters based on silhouette scores and dendrograms.
- Cluster 1 (overmodulated plans) showed lower gamma pass rates (76.7%) and higher MCS (0.112).
- Cluster 3 (optimal plans) demonstrated higher gamma pass rates (92.9%) and lower MCS (0.359), suggesting potential for skipping pretreatment verification.
- Head-and-neck cases exhibited higher complexity and classification rates in the overmodulated cluster.
Conclusions:
- Unsupervised clustering effectively reveals hidden patterns in VMAT plan complexity for dosimetric QA.
- A three-cluster classification scheme based on plan complexity is supported by the data.
- The Modulation Complexity Score (MCS) metric is a strong predictor of PSQA outcomes.
