FastSegmentator

Fast inference pipeline for nnU-Net and TotalSegmentator, the most popular frameworks for medical image segmentation. This project provides a clean, minimal inference module with only the necessary components, plus an end-to-end GPU fast-path: every stage — resampling (cucim + a GPU cubic-B-spline that matches scipy order=3 to ~1e-13), normalization, the sliding-window forward pass, logits→label conversion, cropping, and connected-component postprocessing — runs on the GPU. Across 24 parity-validated modes it reproduces official TotalSegmentator output at ≥0.999 DSC on the headline modes (≥0.995 on 21 of 24) while running 2–9× faster (forward-pass-bound; the rest is fixed import/model-load overhead amortized in batch).

Requirements

• Python: 3.10
• CUDA: 12.4
• uv for environment management

Installation

1. Clone this repo. The vendored nnunetv2 lives in src/, and TotalSegmentator is expected as a sibling checkout (see [tool.uv.sources] in pyproject.toml):

git clone https://github.com/JunMa11/FastSegmentator.git
git clone https://github.com/wasserth/TotalSegmentator.git   # sibling of FastSegmentator
cd FastSegmentator

2. Create the environment and install everything (including the FastSegmentator command) in one step:

uv sync

uv sync builds the editable nnunetv2 package from src/, installs the pinned CUDA 12.1 torch wheels, cupy/cucim, and registers the FastSegmentator console script into .venv/.

3. Activate the environment so the FastSegmentator command is on your PATH:

source .venv/bin/activate

Data and Model Weights

Download the dataset and model weights from the Google Drive link.

• Place the dataset in FastSegmentator/nnUNet_data/
• Place the model weights in FastSegmentator/model_weights/

TotalSegmentator weights default to ~/.totalsegmentator/nnunet/results (override with --weights_dir).

Running Inference

With the environment activated, the FastSegmentator command dispatches to one of two backends:

FastSegmentator <command> [options]

Without activating, you can equivalently run uv run FastSegmentator ... or .venv/bin/FastSegmentator ....

totalseg — TotalSegmentator modes (config-driven)

FastSegmentator totalseg \
    -i <path_to_input_images> \
    -o <path_to_output_segmentations> \
    --task total

Flag	Default	Description
-i, --input_path	(required)	Folder of *.nii.gz input images
-o, --output_path	(required)	Folder to write multilabel output NIfTIs
--task	total	Mode (e.g. total, total_mr, body_mr, …)
--weights_dir	~/.totalsegmentator/nnunet/results	TotalSegmentator weights path
--device	cuda	Device (cuda or cpu)

Run FastSegmentator totalseg --help for the full list of --task modes.

nnunet — generic nnU-Net model folder

FastSegmentator nnunet \
    -i <path_to_input_images> \
    -o <path_to_output_segmentations> \
    --model_path <path_to_model_weights>

Flag	Default	Description
-i, --input_path	(required)	Path to the input image folder
-o, --output_path	(required)	Path to save output segmentations
--model_path	(required)	Path to the trained model directory
--fold	all	Fold to use for inference
--checkpoint	checkpoint_final.pth	Checkpoint filename
--use_softmax	False	Apply softmax to output probabilities
--device	cuda	Device (cuda or cpu)

Trainers. By design, the nnunet branch resolves only the standard nnU-Net trainers (nnUNetTrainer, nnUNetTrainerNoMirroring, nnUNetTrainerTopkLoss). To use a model trained with a custom trainer, point the nnUNet_extTrainer environment variable at the directory containing your trainer class so it can be resolved at checkpoint load:

export nnUNet_extTrainer=/path/to/your/trainers

Example

FastSegmentator nnunet \
    -i ./nnUNet_data/Dataset701_AbdomenCT/imagesVal \
    -o ./seg \
    --model_path ./model_weights/701/nnUNetTrainerMICCAI_repvgg__nnUNetPlans__3d_fullres

Parity with official TotalSegmentator

The fast-path is validated to match official TotalSegmentator on the same input (parity, not vs. ground truth) across 24 modes. Overview + interactive figures: report/index.html; full per-mode report: report/validation_report.html.

Of the 24 validated modes, 21 reach ≥0.995 DSC — every previously-failing pathology mode is now ≥0.999 — and 3 thin/sparse modes carry small, characterized caveats, all at 2–9× speedup:

Mode	Task	DSC vs official	Speedup
total	291–295	1.0000	9.6×
liver_lesions	591	1.0000	4.9×
liver_lesions_mr¹	589	1.0000	6.2×
liver_segments_mr	576	1.0000	6.7×
trunk_cavities	343	1.0000	4.1×
lung_vessels	117	0.9999	2.8×
teeth	113	0.9999	3.9×
total_mr	850,851	0.9999	9.5×
lung_nodules	913	0.9999	7.5×
vertebrae_mr	756	0.9999	4.6×
lung_vessels_LEGACY	258	0.9999	4.2×
craniofacial_structures	115	0.9998	4.0×
body	299	0.9996	2.3×
pleural_pericard_effusion	315	0.9990	9.3×
head_muscles	777	0.9989	5.1×
head_glands_cavities	775	0.9987	5.0×
liver_segments	570	0.9984	5.2×
abdominal_muscles	952	0.9981	4.6×
headneck_bones_vessels	776	0.9968	4.5×
oculomotor_muscles	351	0.9960	4.8×
body_mr	597	0.9956	4.9×
headneck_muscles	778,779	0.9945	4.9×
kidney_cysts	789	0.9919	4.9×
liver_vessels	8	0.9880	5.9×

¹ liver_lesions_mr — an 86-voxel lesion on the crop boundary, nondeterministic on both pipelines (official itself flips 86/52 voxels across runs); our deterministic output matches official's same-draw at DSC 1.0.

Three fixes brought the harder modes to parity (each isolated by bisecting against official's per-function intermediates):

1. GPU cubic-B-spline input resample — replaced order-1 trilinear (F.interpolate) with a separable order-3 cubic B-spline matching nnU-Net's skimage.resize(order=3) to ~1e-13. (pleural, lung_nodules)
2. dtype=np.int32 on the cucim input resample — matches official's pre-model int truncation. (total_mr, liver_segments_mr, liver crops)
3. Per-mode softmax→argmax convert for low-confidence lesion modes. (liver_lesions, liver_lesions_mr)

Plus GPU-ported crop + connected-component postprocess (bit-identical to the scipy originals) and cuDNN-deterministic forward for reproducibility.