(Saturday, 4/4/2020) [Mountain Time (GMT-6)]
Purpose: Artificial intelligence (AI) agent has shown promising performance in treatment planning application. Yet, AI agent sometimes struggles in prospective testing which is often attributed to unseen scenarios. The purpose of this study is to first evaluate the performance of the AI agent against human planner in prospective clinical setting, and second to collect human planner’s feedback to improve the AI agent.
Methods: Seventy-two whole breast (55)/chest wall (17) patients have been progressive included between May and December, 2019. A previously developed in-house AI breast auto-planning agent was released for dosimetrists to use for electronic compensation-based planning technique. The dosimetrist reviewed the auto-plan and made edits necessary before presenting the final-plan to physician for review. Planning time for deploying the agent (auto-planning) and manual editing (fine-tuning) was recorded respectively. Dosimetric endpoints for breast/chest wall, internal mammary node (IMN), tumor bed, heart and ipsilateral lung were recorded. Wilcoxon Signed-Rank test was performed to test the null hypothesis that both plan groups have no plan quality difference.
Results: A total of 74 plans were recorded (2 patients with bi-lateral treatment). Mean (standard deviation, SD) time for agent operation and manual editing was 15.4 (10.2) and 17.2 (14.1) min, respectively. Breast/chest wall V95% was comparable between auto-plan (95.9%) and final-plan (95.9%, p=0.617). Tumor bed V95% was 99.7% for both groups (p=0.241). No significant difference was observed for heart or ipsilateral lung. Improvement was made on IMN coverage: V95% improved from 71.2% to 77.1% (p=0.007). Max point dose was reduced from 112.6% to 110.6% (p<0.001).
Conclusion: AI breast auto-planning agent was successfully deployed in daily clinical setting. It provides decent breast/chest wall coverage, tumor bed coverage and substantial overall planning time reduction (from hours to ~30 mins). This study demonstrates our experience of how AI agent should be integrated in nowadays clinic.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by NIH R01CA201212 research grant.
ePosters
Contact Email
