Session Item

Artificial intelligence (AI)-based applications have potential to assist physicists and technicians in routine clinical tasks such as quality assurance (QA) of radiotherapy treatment planning. For instance, the Bayesian network (BN) alert system developed by Luk et al.(1) has shown to be a promising tool for the detection of potential treatment planning errors. The goals of this study were to 1) assess the effectiveness of an evolved version of the BN with new variables and links in an international multi-centric setting, and 2) establish an interoperable framework that works across different technologies, clinical guidelines, and patient characteristics.

Treatment planning, diagnostic and dose prescription data from three different radiotherapy centres were collected: Maastro (Netherlands) using ARIA, University of Washington (UW) and University of Vermont Medical Center (UVMMC) using Mosaiq as “record and verify systems”. Data from the three centres were harmonised and discretised, in order to predict radiotherapy errors, related to dose prescription, treatment plan parameters and setup equipment. Semi-structured interviews were conducted with radiotherapy experts to identify relevant variables and additional links to construct the BN topology. We simulated errors based on the failure modes reported in TG 275 of the American Association of Physicists in Medicine(2) and the reported errors in the three centres. The BN’s ability to identify errors at each centre was tested, after training it on data from another centre (cross-site validation).

The structure of the BN network is presented in figure 1. The highest performance was observed for the BN when it was trained at the UW centre and validated at the UVMMC with an area under the receiver operating characteristic curve (AUC) value of 0.80. The different AUC values of each validation setting are summarised in table 1. High performance was observed for the gantry angle  errors (AUC=0.84) while on the contrary, the BN did not perform adequately in the detection of the number of fractions errors when trained at UVMMC and validated at MAASTRO (AUC=0.61).

Figure 1: Structure of thE BN 

We successfully investigated the efficacy of the BN in an international multi-centric setting. The BN has shown the efficacy to detect possible radiotherapy planning errors only from the same institution that was trained on. This result shows that the model is generalizable across different practices with data harmonization. However, the performance on specific components, such as prescription, could be affected due to the significant differences in treatment protocols across institutes. Future work is required in terms of the error simulation method and the data standardisation methodology for the performance improvement of the BN with the involvement of additional institutions.

References

1.    PMID:30927253

2.    PMID:31967655