Clustering of dosimetric profiles reveals distinct local control probabilities after SABR in oligometastatic head and neck cancer: insights from the OMET phase II trial quality assurance Process.

Abstract

Stereotactic ablative radiotherapy (SABR) is increasingly used in the management of oligometastatic disease. However, variability in SABR plans raises questions about their impact on local control at SABR-treated lesions (LC). We aimed to explore whether quantitative dosimetric parameters could predict LC in head and neck squamous cell carcinoma (HNSCC) patients in the OMET (GORTEC 2014-04) trial. OMET is a multicentre randomized phase II trial comparing SABR-alone versus chemo-SABR in patients with ≤ 3 PET-confirmed oligometastases. A post-hoc analysis of all irradiated lesions (N = 98) from 69 patients was performed. Twenty spatial and dosimetric indices, together with conventional metrics including Dmin, Dmean, Dmax, total target volume and homogeneity/conformity indices, were extracted from the DICOM files. Hierarchical clustering was used to identify phenotypes of plan quality. Kaplan-Meier analyses evaluated associations with LC. Wide inter-patient variability in dosimetric parameters and three clusters was observed, despite SABR standardization per trial protocol. The cluster of lesions (N = 13) with high intra-tumoral dose heterogeneity and non-optimal conformity was associated with significantly improved LC. In contrast, a more homogeneous and conformal phenotype was linked to inferior LC (N = 14). The largest cluster (N = 69) showed no clearly distinctive pattern and had intermediate LC. In SABR for oligometastatic HNSCC, intra-tumoral dose heterogeneity may be more predictive of LC than strict conformity, particularly in high-dose per fraction regimens. A quantitative, phenotype-based machine learning approach using unsupervised clustering of composite dosimetric metrics may be explored further within SABR quality assurance frameworks beyond binary expert review alone.