Meta-learned 3D SIREN implicit neural network for 5-class brain MRI volume segmentation on neurite-OASIS at 80×80×100 resolution. Trained with the MetaSeg recipe (Vyas et al., MICCAI 2025).

Performance

Metric	Value	Source
Mean Dice (OASIS test split, n=80)	0.915 ± 0.008	Reproduction of paper Table 1 row 3 (paper: 0.91 ± 0.011, within 1σ)
INR parameters	330 K	–
Inference time (per volume, A100 BF16)	~5 s	inner_steps=100

Parameter efficiency context: SegResNet at 4.7 M params hits 0.963 on this task (the empirical ceiling); SwinUNETR at 62 M and 248 M both plateau at 0.965. This 330 K-parameter MetaSeg model is only 0.05 Dice below the ceiling at 750× fewer parameters than SwinUNETR.

Output classes: same five as metaseg-siren-2d-5cls (bg / CSF / cortex / WM / deep grey).

Usage

python

from inr_brain_seg import InrBrainSeg
model = InrBrainSeg.from_pretrained("basimazam/metaseg-siren-3d-5cls")
mask  = model.segment("path/to/T1_volume.nii.gz")
# Returns a 3D numpy.ndarray, shape (80, 80, 100).

Training details

Backbone: 3D SIREN with omega_0=30, 3 hidden layers of width 256.
Volume shape: 80×80×100 (resampled from neurite-OASIS's native grid).
Outer loop: 5,000 MAML iterations, lr 1e-4, inner_steps=2.
Classifier finetune: 4,001 epochs, lr 5e-5.

Citation

See repository results/RESULTS.md.

License

Apache License 2.0.