Complexity-based unsupervised machine learning for patient-specific VMAT quality assurance

Background: Patient-specific quality assurance (PSQA) is essential to guarantee the requested accuracy and safety of high-precision radiotherapy treatments. With the widespread adoption of modulated-intensity techniques, there is a growing need for increased operational efficiency. The potential of machine learning (ML) to accurately predict PSQA results has been investigated in recent years. In particular, plan complexity metrics have been used as model input features to be related to the PSQA outcome results in a number of supervised ML models. However, an unsupervised cluster analysis, able to uncover hidden patterns or groupings in data, has not been yet performed. Purpose: The primary aim of this research was to investigate the potential of different unsupervised ML methods to unravel hidden patterns and groupings in PSQA data based on a clustering analysis of plan complexity. Methods and materials: A total of 1329 pretreatment verification data from 660 consecutive patients with different tumour sites treated using volumetric modulated arc therapy (VMAT) were analyzed using the modulation complexity score (MCS) and the dynamic log-files generated by the linac. Predicted and measured fluences were compared using γ-analysis in terms of mean γ-values (γmean) and γ-pass rate (γ%) at the 2%(local)/2 mm criterion. Three unsupervised clustering algorithms, including agglomerative hierarchical clustering (AHC), K-means (KM) and Gaussian mixture models (GMM), were implemented to investigate the existence of natural groupings or clusters based on plan complexity. In addition, we subsequently trained several supervised models to validate cluster assignments on an external cohort of 202 VMAT arcs. Results: For each clustering algorithms, the silhouette scores and the dendrogram analysis indicate the optimal number of clusters is three. The GMM clustered 65 arcs (4.9% of total arcs) into cluster 1 with mean values of γ%, γmean and MCS of 76.7%, 0.85 and 0.112, respectively. 916 arcs (68.9% of total arcs) were grouped into cluster 2 with mean values of γ%, γmean and MCS of 86.5%, 0.58 and 0.209, respectively. Lastly, 348 arcs (26.2% of total arcs) were grouped into cluster 3 with mean values of γ%, γmean and MCS of 92.9%, 0.40 and 0.359, respectively. Cluster 1 was associated with overmodulated plans, providing a warning MCS cutoff value of 0.145 for prompt replanning. Similarly, cluster 3 was associated with PSQA optimality, providing a MCS cutoff value of 0.278, beyond which plans have an a-priori very high QA pass results and can avoid the pretreatment dosimetric verification. Head-and-neck cases reported the higher (12.0%) and the lower (4.0%) classification rates in clusters 1 and 3, respectively, suggesting a major increase of the complexity score for these plans. Conclusion: This study demonstrated the potential of clustering analysis to unravel hidden patterns of plan complexity in dosimetric quality assurance of VMAT treatments. The results suggested that a three-clusters classification scheme has a true basis in plan complexity, supporting the hypothesis that the MCS metric strongly underlies PSQA results.