Click here to

Session: Therapy General ePoster Viewing [Return to Session]

How AI Can Support Personalized Risk Mitigation During Automated Treatment Planning

M Bach1*, G Spira2, (1) ,Overath, NW, DE, (2) ,Cologne, ,DE


PO-GePV-T-158 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: To automatically determine individual risk priority numbers for patients during the radiation therapy process due to potential human and technical errors and to demonstrate how to mitigate some of these risks through artificial intelligence (AI).

Methods: APIs (Application Programming Interfaces) for treatment planning systems are used to expand their scope of functionality and be used to interface AI algorithms. In this work, a C# environment was used to implement an Expert System (own activity), supervised Machine Learning (ML.NET, Microsoft), and Eclipse Scripting (ESAPI, Varian Medical Systems).Ratings for the risk factors severity S, occurrence O, and detectability D were created for several error scenarios and grouped into three phases of an automated planning workflow (load patient with associated demographic information; checking prescription and consistency of PTV/OAR structures; calculating plan setup, dose, and quality metrics). The program calculates risk priority numbers (RPN=S*O*D) during each planning step and compares these with values mitigated by the AI algorithms.

Results: Several AI mitigations were successfully integrated, among them binary classification, anomaly detection, and recommendations. In more than 80% of automated planning workflows with initially critical (RPN between 30 and 125) or unacceptable risk levels (RPN above 125), the AI-supported treatment planning found mitigations to reduce values to an acceptable level (RPN below 30). Where algorithms could perform no mitigations, root cause analysis identified either missing branches in the Expert System or missing data.

Conclusion: This work demonstrates the potential for significant risk reduction by embedded AI support during treatment planning. With prior trained machine learning models and expert systems, the gain in safety comes basically without additional costs; and the modular structure of ESAPI also allows the implementation of other mitigations programmed by third-party institutions.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email