title: How to run a pretrained 3D segmentation model on your CT scans (nnU-Net, TotalSegmentator) with pardesline description: Run a pretrained 3D medical segmentation model on your CT scans in Python today: prepare the data in the nnU-Net layout from any pardesline dataset, run nnU-Net or TotalSegmentator, load the predictions back as label maps, measure Dice against the reference; and what the coming model registry changes.

How to run a pretrained 3D segmentation model on your CT scans¶

Status

pardesline does not ship weights or an inference API yet. This page shows the working path today (prepare → run the model's own tool → load the result back) and what one-line model access will replace it with.

1. Prepare the data in the model's layout¶

pip install "pardesline[medical]"
python skills/medical-segmentation-models/scripts/prepare_nnunet_dataset.py msd_task09_spleen --out ./nnUNet_raw/Dataset509_Spleen --take 20
python skills/medical-segmentation-models/scripts/prepare_nnunet_dataset.py /data/my_cts --image ct --mask seg --out ./nnUNet_raw/Dataset900_Mine

The script writes imagesTr/<case>_0000.nii.gz, labelsTr/<case>.nii.gz and dataset.json (channel names, labels, number of cases) from any pardesline dataset or scanned folder: NIfTI, MetaImage, DICOM series and npz grids are converted with SimpleITK; several binary mask roles are merged into one label map in the order given. Split by patient before exporting (ds.split(by="patient")).

2. Run the model with its own tool¶

nnU-Net v2: nnUNetv2_predict -i imagesTs -o preds -d 509 -c 3d_fullres with the trained or downloaded weights.
TotalSegmentator (117 structures on CT): TotalSegmentator -i case_0000.nii.gz -o seg_case --ml writes one multi-label NIfTI.

Both read NIfTI in physical coordinates, which is what the preparation step wrote.

3. Load the predictions back and measure¶

import pardesline as pl

pred = pl.load("preds/spleen_10.nii.gz").as_labels({1: "spleen"})
case = pl.load("msd_task09_spleen")["spleen_10"]
surface = pred.to_mesh(label=1).smooth()
surface.save("spleen_10_pred.stl")

python skills/evaluation-metrics-3d/scripts/metrics3d.py --help        # dice, hd95, surface_dice with spacing

The metrics guide explains which numbers mean something; the threshold baseline is the comparison a model report should include.

Why models fail on new scanners¶

Spacing and field of view outside the training distribution, contrast phase, cropping, and label-definition mismatches (is the splenic vein spleen?) account for most failures. The medical segmentation models guide lists the checks: resample to the model's spacing, compare intensity histograms, run on a few cases with reference labels first.

What changes with the model registry¶

A model manifest (source, weights URI, licence, input type, task, memory needs, provenance) and pl.load_model("totalsegmentator") / pl.run(model, dataset) will replace steps 1–3 with one call that returns LabelMap objects with the model's label names and provenance. Follow the tracking issue; model manifests can be written now from templates/new_model/.