ESTRO 2024 - Abstract Book

S3049

Physics - Autosegmentation

ESTRO 2024

Results:

We sampled 20 potential masks from the prediction distribution learned by the model. We then fused them into a tumor probability map, assessing the relative voxel-wise contribution of each prediction (Figure 2). Our model achieved a dice coefficient of 74% and a Hausdorff distance of 7.0mm between the ground-truth histology-based contours and the binary fused prediction (thresholded for probabilities >0.5). Depending on the initial masks, the dice coefficient ranges from 71% to 76%, highlighting the variability yet overall good quality of predictions. It also outperforms traditional methods like UNet (dice = 61%). More importantly, the overlap is higher than with the manual GTV (dice = 59% with histology labels), which overestimates tumor extent, emphasizing our method’s potential for more accurate tumor representations in RT planning.