Fig. 2
From: Deep learning-based segmentation of multisite disease in ovarian cancer
Evaluation of model performance on unseen datasets in terms of DSC (a, e) and volume (b–d, f–h). Trainee radiologist segmentations were only available on the evaluation set. The brackets indicate significant differences. All volumes are given in cm3. It can be observed in panels a and e that our method outperforms the nnU-Net baseline for both sites on the evaluation and test set and that our method does not perform significantly different from a trainee radiologist in segmenting pelvic/ovarian lesions in contrast to nnU-Net. Panels b–d and f–h suggest that the model in its current state can be used to determine disease volume for both sites. DSC Dice similarity coefficient, nnU-Net No-new-Net