Meta-learned SIREN implicit neural network for 25-class brain MRI segmentation on neurite-OASIS coronal slices. Trained with the MetaSeg recipe (Vyas et al., MICCAI 2025) — 5,000 outer-loop MAML iterations + 4,001 classifier-only finetune epochs.

Performance

Metric	Value	Source
Mean Dice (OASIS test split, n=80)	0.855 ± 0.035	Reproduction of paper Table 1 row 2 (paper: 0.86 ± 0.032, within 1σ)
INR parameters	1.06 M	–
Inference time (per slice, A100 BF16)	~0.4 s	inner_steps=100

Output classes: 1 background + 24 foreground anatomical regions (left/right cortex, left/right WM, thalamus, caudate, putamen, pallidum, hippocampus, amygdala, accumbens, brainstem, cerebellum). The 25-vs-24 difference is a naming convention: the paper counts 24 foreground labels; this checkpoint includes class 0 (background) for a total of 25 softmax bins. Both refer to the same seg24 files in the neurite-OASIS release.

Usage

python

from inr_brain_seg import InrBrainSeg
model = InrBrainSeg.from_pretrained("basimazam/metaseg-siren-2d-25cls")
mask  = model.segment("path/to/T1.nii.gz")

Training details

Backbone: SIREN with omega_0=30, 3 hidden layers of width 512.
Outer loop: 5,000 MAML iterations, lr 1e-4, inner_steps=2.
Classifier finetune: 4,001 epochs, lr 5e-5.
Test-time inner loop: K=100 steps at lr 1e-4.

Training data: 314 neurite-OASIS subjects. Validation: 21. Test: 80.

Intended use

Research only. See the 5-class card for cross-site / clinical-use caveats.

Citation

bibtex

@inproceedings{azam2026inrbrainseg,
  title     = {Coordinate-Field Implicit Networks for Cross-Site Brain MRI Segmentation},
  author    = {Azam, Basim},
  booktitle = {ACCV},
  year      = {2026}
}

License

Apache License 2.0.