fidoptSynthSeg 1.0
SynthSeg 1.0, a U-Net trained only on synthetic scans drawn from brain label maps, so that one model reads any MR contrast and resolution: 32 brain structures, run as its authors' SynthSeg_predict.py --v1 runs it.
SegmentationCPUversion SynthSeg 2a2aa3b, models/synthseg_1.0.h5 of the repository (sha256 7d2e32d2)brain · MRApache-2.0
Run it
With its KonfAI app, at the benchmark's settings (the fidopt app repositories are private for now: a Hugging Face token with access opens them):
pip install konfai-apps konfai-apps infer fideus-labs/fidopt-segmentation:SynthSeg -i image.nii.gz -o out
The app on Hugging Face: its model, its settings and how it is scored.
Where it works
Its place on each kind of problem it has results on, among the methods ranked there. A method that leads one problem can trail another: read the one closest to yours.
| Problem | Place | Ranked | Datasets |
|---|---|---|---|
| Brain MR, segmentation | not ranked yet | 3 of 6 |
Results
| Dataset | # | Read by | Score | Failed | Time s |
|---|---|---|---|---|---|
| Baby Open Brains, 31 structures | 1 | Dice | 0.626 | 0 | 43 |
| M-CRIB neonatal brain, 94 regions | 1 | Dice | 0.480 | 0 | 26 |
| MRBrainS18 | 1 | Dice | 0.840 | 0 | 91 |
| UltraCortex, cortex and white matter at 9.4 T | 1 | Dice | 0.848 | 0 | 47 |
- Reference
- Billot et al., Medical Image Analysis 2023