TotalSegmentator-KonfAI

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

🧩 Overview

TotalSegmentator-KonfAI is a lightweight command-line interface (CLI) for running TotalSegmentator models for multi-organ medical image segmentation, through the KonfAI deep learning framework.

It provides fast and efficient inference for segmentation tasks, including on low-resource hardware. Pretrained models are automatically downloaded from Hugging Face Hub.

⭐ Key Advantages

📦 Lightweight model distribution

• ~125 MB per model 1.5 mm model
• 🔁 Compared to ~234 MB per model for the original TotalSegmentator
• ~66.2 MB 3 mm models
• 🔁 Compared to ~135 MB (original)

➡️ Faster setup, smaller disk footprint

⚡ Efficient inference

🔬 Performance comparison

Same input, same weights (Datasets 291–295, 1.5 mm, 5-model total), same PyTorch build (cu13.0), single NVIDIA RTX PRO 5000 (24 GB). Peak RAM = process-tree resident set; peak VRAM = over baseline.

📈 Key observations

• 1.5–3.6× faster whole-body inference across sizes, 2.7–4.1× less host RAM.
• KonfAI trades more VRAM on small/medium cases (larger patches, GPU accumulation) for the speed-up, while it stays inside a 24 GB card.
• On large cases the streaming reassembly bounds VRAM (10.4 GB) where the original nears the card limit (23.3 GB / 24 GB) — KonfAI is then lighter on both RAM and VRAM.

🧠 Features

• ⚡ Fast inference powered by KonfAI
• 🤗 Automatic model download from Hugging Face
• 🧠 Supports evaluation workflows with reference data
• 🧾 Multi-format compatibility: supports all major medical image formats handled by ITK

🚀 Installation

From PyPI:

python -m pip install totalsegmentator-konfai

From source:

git clone https://github.com/fideus-labs/KonfAI.git
python -m pip install -e apps/totalsegmentator

⚙️ Usage

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

Perform segmentation on an input volume:

totalsegmentator-konfai segment total -i path/to/image.nii.gz -o ./Output/

Evaluate against a reference, or run both at once:

totalsegmentator-konfai eval total -i image.nii.gz --gt reference.nii.gz -o ./Output/
totalsegmentator-konfai pipeline total -i image.nii.gz --gt reference.nii.gz --gpu 0

Arguments

Note: TotalSegmentator models do not expose an uncertainty workflow, so there is no uncertainty sub-command.

📖 Reference

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

• Wasserthal, J. et al. (2023).
  TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images.
  Radiology: Artificial Intelligence, 5(5). https://doi.org/10.1148/ryai.230024

• Akinci D’Antonoli, T. et al. (2025).
  TotalSegmentator MRI: Robust Sequence-independent Segmentation of Multiple Anatomic Structures in MRI.
  Radiology, 314(2). https://doi.org/10.1148/radiol.241613

• 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 CT (295 × 259 × 219, 2 mm), patch [96, 128, 160], 5-model ensemble (total), half precision (autocast). The app auto-selects the batch size from your free GPU VRAM (vram_plan); override it in SlicerKonfAI (⚙ Advanced) or on the CLI with --patch-size / --batch-size.

The 5-model total head (117 classes) needs ~20 GB for its forward, so the ensemble targets a 24 GB card — on smaller cards use total-3mm (1 model, 3 mm). Its whole-volume accumulator is too large for the GPU, so reassembly runs on the host (~19 GB RAM). A larger batch saturates the card and slows inference. Inference scales with the case size.

🔗 Links

• 🧠 Original TotalSegmentator: github.com/wasserth/TotalSegmentator
• 🤗 Model Hub: huggingface.co/VBoussot/TotalSegmentator-KonfAI
• 📦 PyPI Package: pypi.org/project/totalsegmentator-konfai