ESTRO 2021 Abstract Book

S1536

ESTRO 2021

Fig. 1 : Automated workflow of QA for auto-segmentation.

Results In the framework of scenario 1, a significant distinction could only be obtained for the bladder structure between class ‘no adaptations’ and ‘adaptations necessary’ with f1-score of 0.91 and 0.96 respectively (Table 1). In the context of scenario 2, all online segmentations of the femoral heads were classified in the correct time-adjustment class (f1-score of 1.00 for all classes). In case of the prostate, only the most severe online segmentations (‘from scratch’) could be correctly detected (f1-score of 1.00). For the anorectum, no statistical distinction could be observed between any of the classes.

Table 1 : Summary of recall, precision and f1-score of the ML classification for every organ for scenario 1 and 2. Note that prostate is not mentioned in scenario 1, because there were no samples that were categorized in class ‘no adaptation’. In scenario 2, no online segmentation was categorised as ‘from scratch’ for anorectum . Conclusion The proposed method provided a proof of concept in that it has the potential to predict the segmentation quality automatically and assist physicians during the verification and revision process. However, further research will be needed to define appropriate classifiers for the different structures. PO-1808 Machine learning to predict best clinical plan for left-sided whole breast radiotherapy C. Fiandra 1 , F. Cattani 2 , C. Leonardi 3 , S. Comi 3 , S. Zara 4 , L. Rossi 5 , B.A. Jereczek-Fossa 3 , U. Ricardi 1 , B.J.M. Heijmen 6 1 University of Turin, Oncology Department, Turin, Italy; 2 IEO European Institute of Oncology, Radiotherapy Division, Milan, Italy; 3 IEO European Institute of Oncology, Division of Radiotherapy, Milan, Italy; 4 Tecnologie Avanzate, Research and Development, Turin, Italy; 5 Erasmus University Medical Center, Radiation Oncology, Rotterdam, The Netherlands; 6 Erasmus University Medical Center, Radiation Oncology, Turin, Italy Purpose or Objective Develop a Support Vector Machine (SVM) model that uses dosimetric and patient-related data to predict the clinician’s choice between two available treatment plans. Materials and Methods