Commissioning of an Artificial Intelligence (AI) Tool for Automated Head and Neck Intensity Modulated Radiation Therapy (IMRT) Treatment Planning

Abstract

<h3>Purpose/Objective(s)</h3> Radiation treatment planning for head and neck (HN) cancers is challenging due to complex anatomy and labor-intensive inverse optimization. Recently we developed an Artificial Intelligence (AI) tool to generate IMRT plans for bilateral planning target volumes (PTVs) with promising results in both plan quality and efficiency. In this study, we describe the clinical commissioning process of this tool. <h3>Materials/Methods</h3> The AI tool was interfaced to the commercial treatment planning system (TPS) in the clinical environment. Essential information about patient anatomy was collected through vendor-supported scripts and transmitted to a designated workstation. This workstation was configured to perform AI core functions to avoid modification and burden on TPS workstations. The AI-generated optimal fluence maps were then imported into TPS for leaf sequencing, dose calculation, and optional fine-tuning. Planners could examine the plan dose and make further adjustments as clinically needed. The workflow and performance were evaluated and validated by experienced physicians and physicists specializing in HN treatment. Training, validation, and testing datasets of the AI tool consisted of 200, 16, and 15 retrospective cases. The commissioning dataset of the AI tool used 36 recent cases of various disease sites in HN. The commissioning AI plans were evaluated based on the corresponding clinical plans regarding critical dose-volume endpoints and 3D dose distributions. All plans were normalized to 44Gy covering 95% of PTV. Wilcoxon signed-rank tests were performed with a significance level of 0.05. <h3>Results</h3> The average AI planning time was 10 to 15min per case with default settings from start to finish. The dose distributions were comparable to clinical plans with slightly better dose conformity and slightly worse but acceptable dose heterogeneity and maximum dose. OAR sparing was comparable. <h3>Conclusion</h3> The AI planning tool for HN was commissioned in our clinic. The commissioning process demonstrates outstanding performance and robustness of the AI tool and provides sufficient validation and confidence for clinical use. Table. Statistics of dosimetric endpoints (median ± std) of the 36 commissioning cases.