| --- |
| license: apache-2.0 |
| library_name: konfai |
| pipeline_tag: image-segmentation |
| tags: |
| - medical-imaging |
| - segmentation |
| - multimodal |
| - ct |
| - mri |
| - cbct |
| - konfai |
| --- |
| |
| # ImpactSeg — Multimodal Body Segmentation |
|
|
| Multimodal (**CBCT / MR / CT**) anatomical **body segmentation** (11 structures), built with |
| [**KonfAI**](https://github.com/vboussot/KonfAI). One model handles all three modalities. |
|
|
| ## 🧩 Model |
|
|
| | Model | Input | Output | Labels | Ensemble | |
| |:--|:--|:--|:--:|:--:| |
| | `body` | Volume (CT / MR / CBCT) | Segmentation | 11 | 1 | |
|
|
| 2.5D residual-encoder UNet · patch `[1, 192, 192]` · resampled to 3 mm. |
|
|
| ## 🚀 Usage |
|
|
| ```bash |
| pip install impact_seg_konfai |
| impact-seg-konfai segment body -i input.nii.gz -o output/ |
| ``` |
|
|
| - **Generic runner:** `konfai-apps infer VBoussot/ImpactSeg:body -i input.nii.gz -o output/` |
| - **Interactive:** [**SlicerKonfAI**](https://github.com/vboussot/SlicerKonfAI) — the ⚙ *Advanced* dialog overrides patch size and batch size. |
|
|
| ## ⚡ Performance & VRAM |
|
|
| Benchmarked on a single **NVIDIA RTX PRO 5000 (24 GB)** with a real whole-body CT (295 × 259 × 219, 2 mm). The batch size is **auto-selected from your free GPU VRAM**. |
|
|
| | Free VRAM | Batch (auto) | Peak VRAM | Time / case | |
| |:--|:--|:--|:--| |
| | 8 GB | 160 | ~7 GB | — | |
| | 16 GB | 320 | ~14 GB | — | |
| | 24 GB | 512 | ~10 GB | **~7 s** | |
|
|
| Single-model body segmentation keeps **system RAM ~1.6 GB**. The thin 2-D patches never fill the card, so inference stays compute-bound (~7 s, largely batch-independent). Override with `--patch-size` / `--batch-size`. |
|
|
| ## 🔗 Links |
|
|
| - 🧠 **KonfAI:** [github.com/vboussot/KonfAI](https://github.com/vboussot/KonfAI) |
| - 📦 **PyPI:** [impact_seg_konfai](https://pypi.org/project/impact_seg_konfai/) |
| - 📄 **Paper:** [arXiv:2510.21358](https://arxiv.org/abs/2510.21358) |
|
|
