Modelpublic

basimazam/metaseg-siren-adapter-v2-ixi

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes10downloads

Model Card

MetaSeg-SIREN + v2 Adapter (Cross-Site IXI)

Cross-site brain MRI segmentation: a frozen OASIS-trained MetaSeg-SIREN pipeline with a 65,920-parameter MLP adapter trained on 200 IXI subjects. Beats a 7.7 M-parameter supervised U-Net by +0.14 Dice with 117× fewer trainable parameters.

Performance

Held-out IXI test split (n=331). All scans MNI152-affine-registered.

Method	Trainable params	Test Dice
MetaSeg-SIREN, zero-shot	0	0.301 ± 0.096
U-Net (Buda) — supervised on OASIS	7,762,498	0.549 ± 0.056
This model (MetaSeg + adapter v2)	65,920	0.688 ± 0.088
Parameter ratio vs U-Net	117× fewer	+0.14 Dice

Per-site breakdown:

Site	Vendor	Field	n	Dice
Guys	Philips	1.5 T	183	0.714
HH	Philips	3.0 T	105	0.662
IOP	GE	1.5 T	43	0.642

Architecture

coords ──► SIREN INR (frozen, 83K params, OASIS-trained)
        ──► features (128-d/pixel)
        ──► MLP adapter (TRAINABLE, 65,920 params, this model)
        ──► seg head (frozen, 645 params)
        ──► class logits

The adapter is a residual two-layer MLP: Linear(128, 256) → GELU → Linear(256, 128), with the output projection zero-initialised so the adapter starts as identity at epoch 0. Houlsby-style.

Usage

python

from inr_brain_seg import InrBrainSeg

model = InrBrainSeg.from_pretrained("basimazam/metaseg-siren-adapter-v2-ixi")
mask  = model.segment("/path/to/preprocessed_T1.nii.gz")

Preprocessing: this model expects scans that have been through the full IXI preprocessing recipe — SynthSeg pseudo-labels are NOT required at inference, but the input must be MNI152-affine-registered (via FSL flirt -dof 12) before the mid-coronal slice is extracted. Without this step, Dice drops back to ~0.31 (see paper §4.5). The full pipeline is implemented in inr_brain_seg.data.preprocessing.run_full_pipeline.

Training recipe

—Base model: basimazam/metaseg-siren-2d-5cls (frozen).
—Adapter: MLPAdapter(feat_dim=128, hidden_dim=256) → 65,920 params.
—Loss: DiceFocal (1:1 weighted; γ=2 for focal).
—Optimizer: AdamW, lr=1e-3, weight_decay=0, 500 epochs.
—Inner loop at training time: K=100 (features cached once).
—Training set: 200 IXI subjects with SynthSeg pseudo-labels (random stratified split across sites).
—Validation set: 50 IXI subjects.

The K-sweep result (negative finding)

The MetaSeg paper claims that the test-time MAML inner loop functions as implicit test-time adaptation (TTA). On real cross-site MRI shift this is not true: with the same OASIS-trained MetaSeg-SIREN model on IXI,

K	Dice
0	0.3115 ± 0.0350
10	0.3155 ± 0.0227
50	0.3238 ± 0.0261
100	0.3241 ± 0.0278
200	0.3242 ± 0.0295
500	0.3239 ± 0.0315

Dice never moves more than 0.013 across two orders of magnitude of K. The inner loop saturates by K=50 and does not bridge cross-site shift on its own.

This is what motivated the population-level adapter approach used here.

Intended use

Research only. Validated on IXI (3 London hospitals, 2 vendors, 2 field strengths). Generalisation to other cross-site cohorts is not guaranteed.

For in-domain neurite-OASIS scans, use basimazam/metaseg-siren-2d-5cls directly (it gets 0.925 Dice there with no adapter).

Citation

bibtex

@inproceedings{azam2026inrbrainseg,
  title     = {Coordinate-Field Implicit Networks for Cross-Site Brain MRI Segmentation},
  author    = {Azam, Basim},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  year      = {2026}
}

License

Apache License 2.0.