Repeatable target
If the same anatomy or lesion appears across many cases, corrected examples can become a project-specific model.
Custom segmentation models
Train custom nnU-Net models from your corrected masks
MedSeg lets you move from annotation to model training inside the same project. Mark masks as training sources, choose 2D or 3D nnU-Net, optionally transfer from another project model, and queue hosted GPU training.
2D or 3D
Transfer learning
Batch inference
Why train instead of prompting every case?
Promptable tools are excellent for first masks. A trained model is better when the same target must be segmented repeatedly.
Site-specific definition
Train on your label definition instead of hoping a public model matches your cohort and annotation style.
2D and 3D datasets
Use 3D training for volumetric CT/MRI and 2D training for single-slice or high-resolution 2D workflows.
Training workflow
- Create or import masks.
Use manual drawing, TotalSegmentator, nnInteractive, VoxTell, or reference labels from public datasets. - Correct the labels.
Training quality follows label quality. Review mask boundaries and remove obvious failed labels before training. - Mark cases for training.
Use the T badge or row context menu to mark the active mask as a training source. - Choose training settings.
Select CT/MRI, 2D or 3D nnU-Net, validation split, duration, and optional transfer learning. - Run the trained model.
After training finishes, run the model from the Models menu on new or held-out cases.
Good first datasets
Open datasets are useful for practicing the workflow before using your own private cohort.
TotalSegmentator
Large labelled CT dataset for whole-body anatomy experiments.
AMOS22
Abdominal CT/MR organs for multi-organ model practice.
Pathology datasets
LIDC-IDRI, BHSD, ReMIND, HCC-TACE-Seg, and other catalog entries for lesion-oriented work.
Training caveats
Do not judge a model by training loss alone. Keep validation or held-out cases, inspect failures visually, and document the cohort. Custom models trained in MedSeg are research tools unless separately validated and regulated for clinical use.