IMPACT-Synth-KonfAI

Fast and lightweight CLI for synthetic CT generation using IMPACT-Synth models within the KonfAI framework.

🧩 Overview

IMPACT-Synth-KonfAI is the command-line interface (CLI) for performing inference and uncertainty estimation with the IMPACT-Synth models.
It provides a streamlined way to generate synthetic CT (sCT) images from MR or CBCT scans, leveraging the KonfAI framework for efficient inference, test-time augmentation (TTA), model ensembling, and uncertainty quantification.

The underlying IMPACT-Synth models are a family of supervised convolutional neural networks (CNNs) dedicated to sCT generation. They build upon the research presented in “Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration” (Boussot et al., 2025).
These models are trained on carefully aligned MR–CT pairs, where alignment is optimized through the IMPACT-Reg loss to minimize spatial bias. Their training further integrates the IMPACT-Synth loss, a perceptual loss derived from semantic representations of segmentation networks. Together, precise spatial alignment and semantic perceptual supervision reinforce anatomical fidelity and realistic tissue contrast in the synthesized CT images.

The official IMPACT-Synth models are available on Hugging Face and can be executed directly through this CLI. This CLI is Apache-2.0; the model weights are CC BY-NC 4.0, the terms of their training data (see the model card).

🚀 Installation

From PyPI:

python -m pip install impact-synth-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/impact_synth

⚙️ Usage

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

Generate a synthetic CT:

impact-synth-konfai synthesize MR -i path/to/input.nii.gz -o ./Output/

Evaluate against a reference CT, or run everything at once:

impact-synth-konfai eval MR -i input.nii.gz --gt reference_ct.nii.gz -o ./Output/
impact-synth-konfai pipeline CBCT -i patient01.nii.gz --gt ct.nii.gz -o patient01 --gpu 0 --tta 2 --ensemble 5 -uncertainty

Arguments

🧠 Features

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

📚 References

If you use IMPACT-Synth-KonfAI in your work, please cite:

• Boussot, V., Hémon, C., Nunes, J.-C., & Dillenseger, J.-L. (2025).
  Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration.
  arXiv preprint arXiv:2510.21358

• 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 [1, 512, 512]. 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.

Single-model sCT keeps system RAM ~2 GB. The plan leaves memory headroom: a larger batch saturates the card and slows inference (batch 48 → ~22 GB). A full 5-model ensemble runs in ~82 s. Inference scales with the case size.

Measured on the S/M/L cases of benchmarks/perf/bench_apps.py (2026-09-09, MR, default ensemble, same GPU). Peak RAM = process-tree resident set; peak VRAM = over baseline.

🔗 Links

• 🤗 Model Hub: huggingface.co/VBoussot/ImpactSynth
• 📦 PyPI Package: pypi.org/project/impact_synth_konfai