📢 New 3D Slicer Extensions: KonfAI & IMPACT-Synth Reliable AI Synthetic CT and Segmentation with Integrated QA

Support
, , ,
1

Dear All,

We are happy to announce the release of two new open-source 3D Slicer extensions, KonfAI and IMPACT-Synth, which provide graphical interfaces to run AI tasks (such as segmentation and synthetic CT generation) through KonfAI Apps directly inside Slicer.

With KonfAI, you can load patient images, select a KonfAI App, run inference, and immediately visualize the results as volumes or segmentations.

You can then perform quality assurance by evaluating predictions when a reference is available, or estimate uncertainty when no reference is available using test-time augmentation (TTA), Monte Carlo dropout, or model ensembling.

:wrench: What is KonfAI?

The KonfAI extension is the graphical interface built on top of KonfAI, a modular and fully configurable deep learning framework for medical imaging, where complete training, inference, and evaluation pipelines are defined declaratively through YAML configuration files.

This design enables reproducible, transparent, and flexible AI workflows, with native support for advanced strategies such as patch-based processing, test-time augmentation, model ensembling, and multi-output supervision, and has been successfully applied to segmentation and image synthesis tasks.

Reference:

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

:package: Built-in KonfAI Apps

KonfAI comes with several ready-to-use Apps covering segmentation and synthesis, all with built-in evaluation and uncertainty estimation.

:brain: Segmentation Apps

MRSegmentator

  • Evaluation: Dice, error maps
  • Uncertainty: uncertainty maps (via model ensembling)
  • Optimized variant of the original MRSegmentator:
    • ~3–4× faster inference
    • ~2.8× lower RAM usage (≈ 30 GB vs ≈ 83 GB)
    • Input volume size: 512 × 512 × 366

TotalSegmentator

  • Evaluation: Dice, error maps
  • Optimized variant of the original TotalSegmentator
  • Available models:
    • total
    • total 3mm
    • total-mr
    • total-mr 3mm

These segmentation Apps reuse pretrained checkpoints from the original tools, while adding built-in evaluation and uncertainty estimation and significantly improving inference speed, enabling their use as downstream tasks, for example for synthetic CT evaluation.

:test_tube: Synthesis Apps

CBCT → CT and MR → CT synthesis

Shared features:

  • Evaluation: MAE, PSNR, SSIM
  • Downstream evaluation: Dice (via automatic segmentation)
  • Uncertainty: uncertainty maps and conformity maps (via TTA and model ensembling)

These synthesis Apps are primarily intended for domain adaptation, enabling task pipelines originally designed for CT to be applied to MRI or CBCT data via synthetic CT generation, with direct applications in radiotherapy dose calculation or segmentation.

The two synthesis and segmentations Apps are integrated into a dedicated 3D Slicer extension, IMPACT-Synth, which provides an end-to-end environment for synthetic CT generation, evaluation, and quality assurance.

We also plan to further develop and release more robust synthesis models as part of this extension.

:books: Training and methodology

The synthesis models were trained on the SynthRAD2025 dataset, following the methodology described in our recent work:

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

Training relies on the IMPACT-Synth loss to enforce semantic and structural consistency, and uses IMPACT-Reg–based registration to pre-align multimodal training pairs, improving anatomical fidelity and robustness of the generated synthetic CTs.

2 Likes