AI‐assisted VMAT planning incorporating deep learning‐based dose prediction for head and neck cancer: feasibility of quality standardization and human intervention for irregular cases

BACKGROUND: Volumetric-modulated arc therapy (VMAT) represents a standard of care for head and neck cancer (HNC). However, treatment plan quality depends heavily on the planner's expertise, thereby challenging the standardization of treatment quality. While the clinical application of artificial intelligence (AI) in radiotherapy planning is rapidly increasing, the robustness of these models-for both typical and irregular cases-remains insufficiently understood. PURPOSE: To evaluate the clinical utility and generalizability of an artificial intelligence (AI)-assisted volumetric-modulated arc therapy (VMAT) planning workflow for head and neck cancer (HNC). This study focused on the robustness of AI-assisted planning using dose distribution prediction for various cases including complex anatomical variations, such as bulky tumors and irregular tumor ruptures, and explored the feasibility of quality standardization through a "human-in-the-loop" intervention approach via re-optimization. METHODS AND MATERIALS: Twelve VMAT plans from 11 HNC patients, including complex anatomical scenarios with bulky masses and post-rupture anatomical changes, were analyzed. Clinical plans developed by an experienced planner were compared with AI-assisted plans generated using RatoGuide (AiRato Inc.). The AI-assisted planning workflow utilized deep learning-based dose distribution prediction to generate dose-derived ring contours for optimization. A human-in-the-loop intervention was implemented, wherein planners performed manual fine-tuning if initial AI-assisted plans violated the institutional dose constraint of Dmax <115%. Dosimetric metrics for the planning target volume (PTV) and organs at risk (OARs), along with Paddick conformity index and monitor units (MU), were evaluated. RESULTS: AI-assisted plans achieved PTV coverage comparable to clinical plans while improving OAR sparing, particularly for the spinal cord and brainstem. The mean spinal cord PRV Dmax was reduced from 27.50 Gy in clinical plans to 23.97 Gy in AI-assisted plans. Even in cases with bulky or ruptured tumors, AI-assisted planning maintained high conformity and the overall mean Paddick conformity index was comparable to clinical plans. Although initial AI plans exhibited hotspots (Dmax > 115%) in three cases near the body surface, manual fine-tuning of optimization successfully suppressed these hotspots, meeting clinical criteria. AI-assisted planning maintained comparable mean MU values, and reduced the total planning time from approximately 2 h to 30 min. CONCLUSION: The AI-assisted planning workflow demonstrated the potential to produce high-quality VMAT plans, even for complex HNC cases. By integrating physicist human-in-the-loop intervention with AI efficiency, this approach could facilitate the standardization of radiotherapy treatment planning quality while supporting patient safety.