ESTRO 2021 Abstract Book

S1413

ESTRO 2021

that uses the optimal contouring for each OAR will be proposed. Materials and Methods

Fifteen head and neck (HN) cancer patients were enrolled. OAR segmentation was performed on planning CT followed by IMRT or VMAT treatment planning. For each patient, the rCT with the largest geometric change was selected. For re-segmentation of OARs on selected rCT, 3 methods were used, including human segmentation (HS) by 3 observers, deformable image registration (DIR) based contour propagation and deep learning contouring (DLC). The clinical treatment plan was re-calculated on rCT to obtain D means of relevant OARs. A comprehensive set of 135 NTCP-models were employed to translate the dose distributions of the rCT’s into NTCP-values. Average D mean and NTCP-predictions of 3 HSs were considered gold standard. Absolute comparisons of all methods to gold standard were performed to calculate variability of D mean and NTCP prediction (the maximum of 3 observers’ was considered variability of HS). A variability of more than 3% in NTCP was considered major variability. The SAS preferentially adopted OARs segmentation with lower variability of D mean from DIR or DLC to obtain near variability of NTCP-prediction to HS. Results For parotid glands, the mean variability of D mean after HS was 1.40 Gy, significantly lower than using DIR and DLC (3.64 Gy;p<0.001 and 3.72 Gy; p<0.001) (Tab.1). DLC showed the highest variability of D mean in the medial pharyngeal constrictor muscle (Mean variability: 5.13 Gy;p=0.01). DIR showed higher variability of D mean in most OARs than DLC, especially in the cricopharyngeal inlet (Mean variability:2.85Gy/ p =0.01). The frequencies of each OAR presenting in NTCP models with major variability were counted, showing the highest frequency for parotid glands in both DIR (53) and DLC (54). The SAS comprised segmentation of parotid glands from HS, pharyngeal constrictor muscle from DIR and the other OARs from DLC. The variability of the NTCP- predictions in DIR and DLC were both higher than HS and SAS (Median: 0.49%, 0.47%, 0.30%,0.33%; 90 th percentile: 2.19%, 2.24%, 1.10%,1.50%; Maximum: 11.89%, 13.86%, 9.94%,7.01%) (Fig.1). Tab.1

Fig. 1

Conclusion DLC performs better than DIR for most HN OARs. Human segmentation of the parotid glands remains necessary