Click here to


Are you sure ?

Yes, do it No, cancel

Automated Plan Check Software Using a Multi-Layered Rules and AI Based Approach

Samuel M.H. Luk*, Landon S. Wootton, Alan M. Kalet, University of Washington Medical Center, Seattle, WA


(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 3

Purpose: The initial plan checking process is currently performed manually and may be improved through automation as highlighted in the recent AAPM TG275. This study develops a novel multi-layered automated plan checking procedure utilizing a hybrid rules-based and Bayesian network (BN) algorithm to improve plan checks.

Methods: The multi-layered automated plan checking procedure consists of three components. The first component uses a RayStation script that runs multiple rules-based checks before exporting plan to the R&V system (Mosaiq). The second component is a rules-based algorithm consists of multiple rules that check post-transfer plan parameters and procedural processes (e.g. documents and images) in Mosaiq against pre-defined criteria. The third component uses a recently published error-detection BN that checks the “appropriateness” of treatment plans using clinical data in Mosaiq.
An in-house web-based tool was developed to process the second and third layers of automated plan checking. It is developed using R programming language with Shiny package, and the BN is constructed and inferred using Bayes Server (Bayes Server Limited, UK) API in R. The probabilistic knowledge of the BN is machine learned from three years of previously validated clinical data from our institutional Mosaiq relational database.

Results: Radiotherapy plans completed by dosimetrists are first checked by the RayStation script before export to Mosaiq, followed by the web-based tool using data from the Mosaiq database to perform rules-based and finally BN-based checks. Violations to rules and/or plan parameters with low probability are flagged for the physicist's attention and correction.

Conclusion: The multi-layered automated plan checking procedure combines the strength of rules-based algorithm in efficiently identifying simple errors with the strength of BN in highlighting deviations from clinical practice norms. Combining these methods mimics human reasoning in a way that assists the physicist to improve their efficiency and accuracy in initial plan checking.


Quality Assurance, Bayesian Statistics


IM/TH- Informatics: Informatics in Therapy (general)

Contact Email