Fig. 3
From: Deep learning-based segmentation of multisite disease in ovarian cancer
Examples of ground truth, automated and trainee radiologist segmentations (pink, cyan and blue, respectively). The first two columns (a, b, d, e) show the cases with median and 90-percentile DSC from the pooled validation and test set. The visual comparison between the automatically generated and manual trainee radiologist segmentation demonstrates typical mistakes of the two instances. Both seem to struggle with the inclusion and exclusion of objects close to the segmentation boundary. The last column (c, f) shows examples of outliers at the extreme ends of the volumes. The segmentation model confused dense components of breast tissue with omental disease as both are embedded in fat