MRSegmentator-KonfAI

Fast and lightweight CLI for whole-body MRI segmentation using MRSegmentator models within the KonfAI framework.

🧩 Overview

MRSegmentator-KonfAI is a lightweight command-line interface (CLI) for running MRSegmentator models through the KonfAI deep learning framework.

It provides fast and efficient inference for whole-body MRI segmentation, including on low-resource hardware.

Pretrained models are automatically downloaded from Hugging Face Hub.

⭐ Key Advantages

📦 Lightweight model distribution

• ~128 MB per model, with up to 5 folds available
• Download only the folds you need
• Total size with 5 folds: ~640 MB
• 🔁 Compared to ~1.07 GB for the original full MRSegmentator model distribution

➡️ Faster setup, smaller disk footprint

⚡ Efficient inference

🔬 Performance comparison

Same input, the five folds ensembled on both sides (-f 5), same PyTorch build (2.12.1, cu13.0), single NVIDIA RTX PRO 5000 (24 GB), MRSegmentator 2.0.0. Peak RAM = process-tree resident set; peak VRAM = over baseline. Measured with benchmarks/perf/bench_apps.py (2026-09-09).

📈 Key observations

• 1.2–1.7× faster, 1.4–5.4× less host RAM; the gap widens with the volume.
• The GPU-resident accumulator trades more VRAM for the speed and low host RAM, while streaming keeps it bounded: on the large case host RAM stays at 6.9 GB where the original grows to 37.4 GB.
• The default -f 2 ensembles two folds and runs faster than the table.

🧠 Features

• ⚡ Fast inference powered by KonfAI
• 🤗 Automatic model download from Hugging Face
• 🧩 Multi-model ensembling
• 🧠 Supports evaluation workflows with reference data, and uncertainty estimation without reference
• 🧾 Multi-format compatibility: supports all major medical image formats handled by ITK

🚀 Installation

From PyPI:

python -m pip install mrsegmentator-konfai

From source:

git clone https://github.com/fideus-labs/KonfAI.git
cd KonfAI
# konfai and konfai-apps must come from the same checkout: this app pins both to its own
# setuptools_scm version, which only exists on PyPI at a release tag.
python -m pip install -e . -e konfai-apps -e apps/mrsegmentator

⚙️ Usage

The CLI is organised into sub-commands, mirroring the KonfAI Apps operations:

Run segmentation on an MRI scan:

mrsegmentator-konfai segment -i path/to/input.nii.gz -o ./Output/

Evaluate against a reference, or run everything at once:

mrsegmentator-konfai eval -i input.nii.gz --gt reference.nii.gz -o ./Output/
mrsegmentator-konfai pipeline -i input.nii.gz --gt reference.nii.gz --gpu 0 -f 3 -uncertainty

Arguments

📖 Reference

If you use MRSegmentator-KonfAI in your work, please cite the original MRSegmentator work in addition to this CLI tool.

• Häntze, H. et al. (2025).
  Segmenting Whole-Body MRI and CT for Multiorgan Anatomic Structure Delineation.
  Radiology: Artificial Intelligence, 7(6). https://doi.org/10.1148/ryai.240777

• Boussot, V., & Dillenseger, J.-L. (2025).
  KonfAI: A Modular and Fully Configurable Framework for Deep Learning in Medical Imaging.
  arXiv preprint arXiv:2508.09823

⚡ Performance & VRAM

Benchmarked on a single NVIDIA RTX PRO 5000 (24 GB) with a real whole-body MR (295 × 259 × 219, 2 mm), patch [96, 128, 160], 5-fold ensemble, half precision (autocast). The app measures its batch size on your GPU (a forward of one patch, then of two, then the largest power of two that fits half of the free VRAM); override it in SlicerKonfAI (⚙ Advanced) or on the CLI with --patch-size / --batch-size.

On a 24 GB card the accumulator stays on the GPU, keeping host RAM low with a byte-identical result. The plan stops short of filling the card: a still-larger batch (12 → ~24 GB) saturates the allocator and slows inference ~2× without running faster. Inference scales with the case size.

🔗 Links

• 🧠 Original MRSegmentator: github.com/hhaentze/MRSegmentator
• 🤗 Model Hub: huggingface.co/VBoussot/MRSegmentator-KonfAI
• 📦 PyPI Package: pypi.org/project/mrsegmentator-konfai