Hello, I am trying to use the DentalSegmentator model from Zenodo to do batch processing through nnunet directly. When running inference with this model, segmentation quality is poor — specifically, label mixing between the mandible and upper skull/maxilla classes, with large regions assigned to the wrong structure.
From what I understand the nnU-Net inference should ensemble predictions across all 5 folds - is this correct? Only fold_0 is present in the Zenodo deposit. Running with -f 0 explicitly produces the segmentation, but with significant label confusion between mandible and skull classes (see photo).
Are the remaining folds (fold_1 through fold_4) available? If so, would it be possible to upload them to Zenodo or share them another way? If only fold 0 is necessary, I assume this is a bug on my end - apologies!
Thank you for making this tool available — it's a genuinely useful resource for the community!
Hello, I am trying to use the DentalSegmentator model from Zenodo to do batch processing through nnunet directly. When running inference with this model, segmentation quality is poor — specifically, label mixing between the mandible and upper skull/maxilla classes, with large regions assigned to the wrong structure.
From what I understand the nnU-Net inference should ensemble predictions across all 5 folds - is this correct? Only
fold_0is present in the Zenodo deposit. Running with-f 0explicitly produces the segmentation, but with significant label confusion between mandible and skull classes (see photo).Are the remaining folds (fold_1 through fold_4) available? If so, would it be possible to upload them to Zenodo or share them another way? If only fold 0 is necessary, I assume this is a bug on my end - apologies!
Thank you for making this tool available — it's a genuinely useful resource for the community!