Prospective Clinical Integration of AI Based Treatment Planning Tool for Whole Breast Radiation Therapy (WBRT): A Single Institution's Three-Year Experience

Abstract

<h3>Purpose/Objective(s)</h3> Artificial intelligence (AI) has shown promising results of treatment planning in well-controlled retrospective environment. However, its clinical acceptability and long-term robustness has not been widely investigated. We report the prospective clinical results of AI-based whole breast/chest wall radiation therapy (WBRT) planning tool. We evaluated the 3-year overall performance and the progression of AI tool's acceptance among human planners. <h3>Materials/Methods</h3> In-house AI based WBRT treatment planning tool using random forest was commissioned for clinic since 2019. A total of 1151 patients have been planned with AI tool by Dec 2021, and all were included in this study. The AI tool is seamlessly integrated with clinical planning workflow and treatment planning system (TPS) – AI automatically generates fluence maps, which are used for final dose calculation in TPS (AI plan). The planner evaluates the plan quality and attempts to improve by manually modifying fluence if deemed necessary before physician's approval (final plan). The manual modification value (MMV) of each beamlet is the difference between fluence maps in AI and final plans. The manually modified beamlets were located and correlated to the dosimetry quality enhancement of the final plan over AI plan. In the longitudinal dimension, MMV also tracks the adaptation of human planner towards AI tools over 3 years. A regression model captures planner's frequency of editing AI plans over time. If the planner becomes more acceptive of the AI over time, the regression trend line will be significantly different from zero. <h3>Results</h3> Table 1 shows the dosimetric endpoints, as well as mean MMV of the entire patient population. No statistical significance was observed between the AI and final plans in PTV coverage and OAR sparing (Mood's median test), except V105% (reduced in final plan by 18.4 cc, <i>p</i>< 0.001). Spatially, irradiated lung region received most frequent editing, while skin flash region received minimum editing. Among all 12 planners, the number of cases planned ranged from 13 to 150. The max mean absolute MMV of a single planner was 5.0% (planned 131 cases) and minimum was 0.8% (planned 38 cases). Ten out of 12 planners showed a significantly decreasing editing AI plans with yearly mean MMV of 4.0%, 2.3% and 0.8% for 2019, 2020 and 2021 respectively, indicating improving acceptability/trust over time (<i>p</i><0.001). <h3>Conclusion</h3> This 3-year study demonstrated robustness of AI planning and steady improvement in acceptability by planners over time. It also revealed key areas towards human planner's preference that can be adopted in future AI tools.