Automatic target volume segmentation for offline adaptive head-and-neck radiotherapy

To assess the clinical utility of Smartfuse (Therapanacea, France), a deformable image registration (DIR) algorithm for automatic propagation of target volumes in the context of offline adaptive head-and-neck radiotherapy.

Ten patients underwent offline re-planning during head-and-neck radiotherapy. Target volumes (GTV and CTV) were manually delineated by radiation oncologists (ROs) on both the initial CT (CT_i) and one re-planning CT (CT_R). These manual contours were compared to those propagated by Smartfuse from CT_i to CT_R. The geometric agreement between DIR-propagated and RO-delineated contours was assessed using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD₉₅), and surface Dice similarity coefficient (sDSC) with 0 and 2 mm thresholds. Dosimetric evaluation was conducted by comparing dose distributions from generated plans using automatically propagated target volumes (PTV_DIR) with reference plans based on RO-delineated targets (PTV_RO). Coverage of RO-delineated targets (GTV_RO + CTV_RO and PTV_RO) was assessed using D95%, D50%, Dmax, and V95% ≥ 95%. Spatial dose differences were analyzed using dose difference (DD) metrics at 5% and 2% thresholds.

Median DSC, HD₉₅, sDSC_{0 mm} and sDSC_{2 mm} were 0.86, 4.0 mm, 0.29 and 0.73, respectively. For D_95%, median relative differences between DIR and RO plans were -0.6% for GTV_RO + CTV_RO and -2.1% for PTV_RO for D_95%. All GTV_RO + CTV_RO reached V95% ≥ 95% with DIR plans, but only 61% of PTV_RO did. Spatial DD analysis showed median pass rates of 99.2% (DD_5%) and 74.5% (DD_2%) for GTV_RO + CTV_RO, and 85.5% (DD_5%) and 54.9% (DD_2%) for PTV_RO.

Smartfuse may facilitate efficient propagation of target volumes in this study. However, medical review of auto-propagated volumes remains essential, as dosimetric discrepancies may arise when relying solely on automatically generated PTV.