Models · 3D volume segmentation

nnInteractive

nnInteractive (DKFZ): a promptable 3D segmenter: click a point, draw a box or a scribble on any structure in a CT, MRI, PET or microscopy volume and it returns that structure's 3D mask; further prompts (positive or negative) refine it. Non-commercial weights (CC-BY-NC-SA-4.0).

experimental Fork of MIC-DKFZ/nnInteractive@6bf57ad · backends: portable, kernels

Checkpoint nninteractive

Paper 2025 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 201 ms 200x200, fp16
vs the faster original 2.3× original fp16: 460 ms
Peak memory 2,643 MiB original: 4,076 MiB
Accuracy vs original fp32 – original not run on the test set

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import volume

model = volume.load("nninteractive")  # kernels on a supported GPU, else portable PyTorch
result = model(volume, spacing)
result.mask
result.backend  # what ran: mode, device, dtype, kernels

The weights are non-commercial: fastforks downloads them only with FASTFORKS_NONCOMMERCIAL=1 set.

fastforks bench nninteractive

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 7050.65×7,156 MiB/8,408 MiB1.91.05
Original fp16fp16 460–4,076 MiB/4,636 MiB3.21.04
Portable fp32fp32not measured ––––
Portablefp16 2152.1×3,075 MiB/3,492 MiB1.10.48
Kernelsfp16 2012.3×2,643 MiB/3,072 MiB0.70.40
Portable + torch.compile 1622.8×3,643 MiB/3,832 MiB18 (build)–
TensorRTnot measured here ––––

Measured on RTX 5060 Ti 16 GB, torch 2.10.0+cu130, 2.14.1+cu130, 2026-11-14 to 2026-11-15. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-15Portablefp323525 11.2 GiB processmatches200x200
Kernelsfp16not implemented–200x200

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Non-commercial only

Code
Apache-2.0 LICENSE
Weights
CC-BY-NC-SA-4.0 source
Training data
120+ volumetric 3D datasets (CT, MRI, PET, 3D microscopy; readme) with simulated prompts.
  • weights: CC-BY-NC-SA-4.0

Our reading of the licences, not legal advice.

Weights

nnInteractive_v1.0/plans.json 6.3 KB· hosted on Hugging Face
nnInteractive_v1.0/fold_0/checkpoint_final.pth 411 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download nninteractive). Non-commercial weights need FASTFORKS_NONCOMMERCIAL=1.

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

TotalSegmentator test subset (8 CT cases x 12 prompts)

mean Dice vs the ground-truth mask

Rundtypedice ↑agree ↑portable_agree ↑prompted runsms / prompted run
Original fp32not measured
Original fp16fp160.9688–– 96 –
Portable fp32not measured
Portablefp160.96880.9999– 96 –
Kernelsfp160.96880.9999– 96 –

The original code was not run on this dataset, so the table cannot show the distance to it.

A small subset of real CT cases of the TotalSegmentator dataset (not the full test split), each run at its card's spacing (1.5 or 3 mm) in the same pipeline; mean over the cases of the Dice (0 to 1) averaged over the classes present, against the ground-truth segmentations (vertebrae_L6 of the subset has no counterpart and is skipped). Original = TotalSegmentator 2.18.0, default fp16 autocast. nnInteractive: per case 12 simulated prompts (point, box, scribble, point + box on 3 organs), one interaction each; Dice of the predicted mask against the ground-truth organ mask, mean over the runs. Agreement is Dice against the original FP16 or portable FP16 mask as labelled. Original = the authors’ code, same checkpoint and protocol. Measured on original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-15; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 715d7cb5, 2026-11-15.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.0 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.003 and p99 <= 0.5 (absolute error of the output mask).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0 not recorded
cpu_label_mismatch≤ 0.001 not recorded
GPU default dtypes, mean |diff|≤ 0.003 0.00006
GPU default dtypes, p99 |diff|≤ 0.5 0

Recorded for nninteractive on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-14, with these kernels: vol_norm

Details

nnInteractive (Isensee et al. 2025): DKFZ's promptable 3D segmentation for CT, MRI, PET and 3D microscopy. Point, box (2D or 3D) and scribble prompts, positive or negative, refine one object mask in a session; a ResEnc-L U-Net on 192³ patches with AutoZoom (zooms out up to 4x while the mask grows past the patch, then refines the changed region at full resolution). Lasso prompts and initial masks are not ported.

Input / output
Input: a CT volume [D, H, W] in Hounsfield units and its voxel spacing in mm (volume.read_nifti reads both from a NIfTI file). Output: a class index per voxel [D, H, W] on the input grid (0 = background).
What the numbers mean, and which models need the camera focal length

Encoder models (nnFoundation) take a CT or MRI volume and give its encoder feature maps. For a trained nnU-Net v2 folder or checkpoint, see the [nnU-Net v2 loader](../api/volume.md#nnu-net-v2).

Input size
any H x W >= 32, padded to a multiple of 32; tested up to 512x512 (5,874 MiB allocated)
Outputs
mask
CUDA kernels
  • vol_norminstance norm and LeakyReLU on 3D volumes
Apple (MPS)
fp32
Full reference, generated from the repository

nnInteractive (DKFZ): a promptable 3D segmenter: click a point, draw a box or a scribble on any structure in a CT, MRI, PET or microscopy volume and it returns that structure's 3D mask; further prompts (positive or negative) refine it. Non-commercial weights (CC-BY-NC-SA-4.0).

Commercial use: no (weights: CC-BY-NC-SA-4.0).

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
nninteractivesynthetic (200x200)original (fp32)
original
portable
kernels
portable + torch.compile
705
460
215
201
162
7156 / 8408
4076 / 4636
3075 / 3492
2643 / 3072
3643 / 3832
nninteractive192x192original
portable
kernels
TensorRT fp16
–
–
–
262 (U-Net on one prepared 8-channel, 192³ patch only; memory estimated)
–
–
–
2738 / 2740
Details
  • Original code: https://github.com/MIC-DKFZ/nnInteractive at 6bf57adf42ca
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: CC-BY-NC-SA-4.0
  • Licence evidence: https://huggingface.co/MIC-DKFZ/nnInteractive/blob/3f308d751c00644e4fde6f09c600264b393b21b5/nnInteractive_v1.0/LICENSE ; https://github.com/MIC-DKFZ/nnInteractive/blob/master/LICENSE
  • Training data: 120+ volumetric 3D datasets (CT, MRI, PET, 3D microscopy; readme) with simulated prompts.
  • Kernel libraries: vol_norm
  • Variants: nninteractive
  • Outputs: mask
  • Input size: any D x H x W volume and its voxel spacing (resampled to the model spacing, sliding window)
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): nninteractive 512x512x512 voxels (5.7 GiB peak allocated, 6.1 reserved)
  • nninteractive 200x200, first call: portable + torch.compile 18 s (cold Inductor cache).
  • nninteractive 192x192, first call: TensorRT fp16 58 s (ONNX + engine).
  • nninteractive load / first call: original 3.23 s / 1.038 s, portable 1.09 s / 0.482 s, kernels 0.65 s / 0.403 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

A small subset of real CT cases of the TotalSegmentator dataset (not the full test split), each run at its card's spacing (1.5 or 3 mm) in the same pipeline; mean over the cases of the Dice (0 to 1) averaged over the classes present, against the ground-truth segmentations (vertebrae_L6 of the subset has no counterpart and is skipped). Original = TotalSegmentator 2.18.0, default fp16 autocast. nnInteractive: per case 12 simulated prompts (point, box, scribble, point + box on 3 organs), one interaction each; Dice of the predicted mask against the ground-truth organ mask, mean over the runs. Agreement is Dice against the original FP16 or portable FP16 mask as labelled. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp16portablekernelsTensorRTnote
nninteractiveTotalSegmentator test subset (8 CT cases x 12 prompts), mean Dice vs the ground-truth mask / agreement with the original / agreement with portable FP160.96880.9688 / 0.99990.9688 / 0.99990.9688 / 0.9997

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 715d7cb5, 2026-11-15; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 715d7cb5, 2026-11-15; tensorrt: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks dc38c37c, 2026-11-17.

Parity

CPU, float32: max |diff| <= 0.0 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.003 and p99 <= 0.5 (absolute error of the output mask).

variantkernels error vs original, mean / p99GPU, torchkernels used
nninteractive6e-05 / 0.0RTX 5060 Ti, torch 2.14.1+cu130vol_norm@44036d4da567:sm_120

Example output

No example rendered for this checkpoint yet.