ESTRO 2024 - Abstract Book

S3038

Physics - Autosegmentation

ESTRO 2024

Training was performed individually for the nnUNet and transformer model, based on the same dataset consisting of T1-weighted MRIs (with contrast) of 49 adult brain tumor patients with complementary annotations for eighth OARs. The training process involved a five-fold cross-validation approach. Test was performed on an independent test-set consisting of six patients with brain tumors. Predictions from the fivefold and physician ground truth were compared using three metrics: Dice similarity coefficient (DSC), 95 percentile Hausdorff distance (HD95) and normalized surface DSC 1mm (NSD1mm). The variance was calculated for each fold and across folds to assess the reliability and robustness of the final segmentations.

Results:

Accuracy