ESTRO 2024 - Abstract Book

S4537

Physics - Machine learning models and clinical applications

ESTRO 2024

For VMAT the classifier performed similarly in TB0 (used for training) and in TB1 but performed worse in TB3 and in Clinac, specially misclassifying the failed beams (Figure 1a). We observed that in TB0 and TB1, significant metric differences existed between passed and failed beams (p-values<0.001) in 8 of 10 metrics, while in TB3, only 4 metrics showed significant differences, and in Clinac, none were significant, possibly due to a small test set (Figure 2a). For IMRT the classifier works well in the test set of TB0 and TB1 (used to train the classifier) but doesn’t transfer well to either TB3 or Clinac (Figure 1b). The reason is the same as with the VMAT classifier, for TB0 and TB1, significant metric differences existed between passed and failed beams (p-values<0.001) in 5 of the 6 metrics while for TB3 only 2 were significant (Figure 2b). In Clinac, differences were significant, but values differed from TrueBeams (Figure 2b). While the behaviour of the classifier for Clinac was expected, the results for TB3 were not. The three true beams are matched machines and portal imagers model are the same. Further analysis on the data sets showed a higher percentage of SBRT treatments in TB0 than in TB3 (37% vs 9%) which could justify, together with different distributions of the complexity metrics for TB3 and TB0/TB1 beams (Figure 2), the results of the classifier.