MRSegmentator-KonfAI

KonfAI-accelerated adaptation of MRSegmentator — multi-organ MRI segmentation (40 structures), built with KonfAI.

🧩 Model

Model	Modality	Labels	Ensemble
MRSegmentator	MRI	40	5

3D residual UNet · patch [96, 128, 160] · resampled to 1.5 mm.

🚀 Usage

pip install mrsegmentator-konfai
mrsegmentator-konfai segment -i input_mr.nii.gz -o output/

• Generic runner: konfai-apps infer VBoussot/MRSegmentator-KonfAI:MRSegmentator -i input_mr.nii.gz -o output/
• Interactive: SlicerKonfAI — the ⚙ Advanced dialog overrides patch size and batch size.

⚡ Performance & VRAM

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 KonfAI's benchmarks/perf/bench_apps.py (2026-09-09).

Case (voxels)	Tool	Time	Peak RAM	Peak VRAM
S (248 × 246 × 141)	KonfAI	15 s	5.2 GB	18.2 GB
	Original	19 s	7.2 GB	3.0 GB
M (249 × 246 × 246)	KonfAI	21 s	5.3 GB	16.3 GB
	Original	24 s	8.5 GB	3.8 GB
L (512 × 512 × 531)	KonfAI	86 s	6.9 GB	20.6 GB
	Original	143 s	37.4 GB	14.9 GB

1.2–1.7× faster, 1.4–5.4× less host RAM, the gap widening 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. The batch size is auto-selected from your free VRAM; override with --patch-size / --batch-size.

🔗 Links

• 🧠 Original MRSegmentator: github.com/hhaentze/MRSegmentator
• 🧠 KonfAI: github.com/vboussot/KonfAI
• 📦 PyPI: mrsegmentator-konfai