ESTRO 2024 - Abstract Book

S3064

Physics - Autosegmentation

ESTRO 2024

Figure 1: Flowchart of the methodology. The first row shows the datasets partitioning. The second row shows the networks in continuous line as training and validation and dot line as testing. The output of both networks is merged and results in a single breathing phase GTV prediction. The ITV is created from all GTVs of all breathing phases (last row in the flowchart).

Results:

Table 1 summarizes the results of the evaluations for GTV and ITV segmentation. For 50% GTV automatic segmentations, the SwinUNETR demonstrated superior metrics overall than DynUNET, but the combination of both networks yielded the highest performance, achieving a mean value SDSC of 0.80±0.11 in the test set. Therefore, only the Swin+Dyn Network was used for both ITV methods, which performed similarly in the test set. The SDSC that could be achieved for the ITV was 0.83±0.08 in the all-breathing phases approach, and 0.85±0.07 in the peak-to-peak