Evaluation of a quasi-automatic treatment planning workflow for head-and-neck cancer in radiotherapy

Automation of radiotherapy workflows enhances efficiency, consistency and would support adaptive online dosimetry on standard linacs. This study investigated the feasibility and quality of a quasi-automatic treatment planning workflow for head-and-neck cancer.

Forty head-and-neck cancer patients previously treated with VMAT were included. Clinical plans were optimized using manually delineated targets and corrected organs-at-risk (OARs) autocontours. Autoplans were generated using clinical and uncorrected autocontours of OARs and CTVn levels, with macroscopic tumors copied from clinical contours. Dose differences between auto- and clinical plans were quantified by projecting autoplans onto clinical contours post-optimization. For 10 patients, 5 medical experts (MEs) manually delineated CTVn levels. Inter‑observer variability (IOV) and mean autocontour deviation from experts (MAV) were quantified using Dice Similarity Coefficient (DSC), surface‑DSC, and 95th percentile-Hausdorff distance (HD_95%). IOV‑derived asymmetric safety margins around PTVn autocontours were computed from Euclidean distances.

Autoplans optimized on clinical contours achieved comparable or superior sparing of OARs with higher PTV coverage. Autoplans based on uncorrected autocontours also improved OARs sparing but showed significantly lower target coverage compared to clinical plans (meanV_95%: 91.9 ± 4.2% vs 98.0 ± 0.9%, p < 0.001). HD_95% and surface Dice scores demonstrated significant differences in CTVn contours for two MEs relative to other MEs (p < 0.05). Applying quantified IOV-based safety margins resulted in meanV_95% of 95.1 ± 5.4%, while no significant impact on OARs dose was observed (p > 0.4).

The quasi‑automatic workflow generated clinically acceptable, often superior plans. IOV‑based CTVn margins enhanced robustness and may hasten clinical adoption.