Uncommon anatomy or pathology
Point at the target and let the model propose a 3D mask instead of drawing every slice from scratch.
Interactive 3D segmentation
Use nnInteractive online inside the MedSeg editor
nnInteractive turns sparse prompts into 3D medical masks. In MedSeg, you can guide it with foreground/background clicks, scribbles, freehand outlines, or a bounding box, then accept, edit, save, and export the mask.
Clicks
Scribbles
Freehand
Bounding boxes
When nnInteractive is the right tool
Use it when you know the structure in the image, but no fixed-class model covers it reliably.
Fast mask correction
Start from a rough automated mask, then add positive and negative prompts to nudge the boundary.
Training data creation
Create first-pass labels, correct them, mark them for training, and then move to a custom nnU-Net model.
How it works in MedSeg
nnInteractive is an editor tool, not a one-shot project-view model. You open a series first, then interact with the image.
- Open a CT or MRI series in the editor.
Load the scan and choose or create the segmentation label you want to fill. - Pick the prompt style.
Use points for compact targets, scribbles for elongated targets, freehand for a quick 2D outline, or a box when the target is well bounded. - Refine before accepting.
Add foreground and background prompts until the proposed 3D mask is good enough to lock into the active label. - Save, measure, export, or train.
The accepted result becomes a normal MedSeg mask, so it can be edited manually or used as training data.
nnInteractive vs VoxTell vs nnU-Net
| Tool | Input | Best use |
|---|---|---|
| nnInteractive | Clicks, scribbles, freehand, boxes | Interactive mask creation and correction inside the editor |
| VoxTell | Free-text prompts | Exploratory text-prompted candidate masks when the target is hard to name as a fixed class |
| Custom nnU-Net | Corrected masks from multiple cases | Repeatable batch segmentation once you have examples |
Limitations
Review every mask. nnInteractive proposes candidate segmentations; it does not replace expert review. GPU sessions are limited so the shared worker remains available to other researchers.