title: Pretrained 3D and medical imaging models in Python: what pardesline ships and what is coming description: Reference implementations and expert guides for DINOv3 features, 3D JEPA pretraining, PointNet segmentation, VAE anomaly detection, nnU-Net preparation and 3D metrics; one-line pretrained model access is on the roadmap.
Models¶
Deep learning in pardesline today is reference implementations and expert guides that read any pardesline dataset: no weights are shipped and there is no inference API yet. Each page below documents the scripts, the example that ran them on a bundled dataset, and the recipe at full scale.
Coming: one-line model access
Models will get the same treatment as datasets: a model manifest (source, weights URI, licence, input type, task, memory and GPU needs, provenance) and a single loader / runner API that looks the same whatever the source (Hugging Face Hub, torch.hub, vendor repositories): pl.load_model("dinov3") then pl.run(model, dataset), then visualise. The URLs under /models/<name>/ are reserved so these pages fill in without breaking links. Vote and comment on the model registry tracking issue; model manifests can be written now from templates/new_model/ and listed on Models wanted.
| Model page | Task | What ships today | Example (executed) |
|---|---|---|---|
| DINOv3 and DINOv2 features for 3D and medical imaging | dinov3 dense features | dinov3_faster_rcnn.py, dinov3_features.py |
dinov2_features_kitti |
| JEPA self-supervised pretraining on 3D medical volumes | jepa pretraining | jepa3d_pretrain.py |
jepa_hippocampus |
| PointNet segmentation of medical and industrial point clouds | pointnet segmentation | make_point_dataset.py, pointnet_seg.py |
pointnet_hippocampus |
| VAE anomaly detection on medical images | vae anomaly detection | vae_anomaly.py |
vae_anomaly_hippocampus |
| Pretrained 3D medical image segmentation models: nnU-Net and friends | medical segmentation models | prepare_nnunet_dataset.py |
skill script on MSD spleen |
| Dice, HD95, Chamfer and the other 3D evaluation metrics | evaluation metrics 3d | metrics3d.py |
metrics_liver_interobserver |
The PyTorch scripts need pip install torch (CPU is enough for the examples); DINOv3 weights need the licence accepted on the Hub, DINOv2 does not.