---
license: apache-2.0
language: en
library_name: pytorch
tags:
  - medical-imaging
  - brain-mri
  - segmentation
  - implicit-neural-representation
  - meta-learning
  - siren
  - parameter-efficient
datasets:
  - oasis
metrics:
  - dice
pipeline_tag: image-segmentation
---

# MetaSeg-SIREN 2D 5-class (OASIS)

Meta-learned **SIREN implicit neural network** for 5-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) | **0.925 ± 0.013** | Reproduction of paper Table 1 row 1 (paper: 0.93 ± 0.012, within 1σ) |
| INR parameters | 83 K | – |
| Seg head parameters | 645 | – |
| Inference time (per slice, A100 BF16) | ~0.3 s | inner_steps=100 |

Output classes:

| ID | Class |
|---:|---|
| 0 | background |
| 1 | CSF / ventricles |
| 2 | cortex |
| 3 | white matter |
| 4 | deep grey matter (thalamus, caudate, putamen, hippocampus, etc.) |

## Usage

```python
from inr_brain_seg import InrBrainSeg

model = InrBrainSeg.from_pretrained("basimazam/metaseg-siren-2d-5cls")
mask  = model.segment("path/to/T1.nii.gz")
# mask is a numpy.ndarray of integer class labels, same spatial shape as the input.
```

The model expects a **mid-coronal slice at 192² resolution**, percentile-normalised
to `[0, 1]`. If you pass a full 3D NIfTI volume, the wrapper extracts the mid-coronal
slice automatically; see `preprocessing.json` for the canonical recipe.

CLI:

```sh
inr-seg-single \
    --model basimazam/metaseg-siren-2d-5cls \
    --input  /data/T1.nii.gz \
    --output /data/T1_seg.nii.gz
```

## Training details

* **Backbone**: SIREN with `omega_0=30`, 3 hidden layers of width 128.
* **Outer loop**: 5,000 MAML iterations, lr 1e-4, inner_steps=2 at lr 1e-4.
* **Classifier finetune**: 4,001 epochs, lr 5e-5, BCE loss.
* **Validation budget**: 50 inner-loop steps.
* **Test-time inner loop**: K=100 steps at lr 1e-4 (paper default).

Training data: 314 neurite-OASIS subjects (train split). Validation: 21
subjects. Held-out test: 80 subjects.

## Intended use

Research only. The model is not validated for clinical use. Best for in-domain
neurite-OASIS-style coronal T1 slices.

For **cross-site** scans (different scanner / vendor / field strength), the
zero-shot Dice drops to ~0.30. Use the adapter variant
[`basimazam/metaseg-siren-adapter-v2-ixi`](https://huggingface.co/basimazam/metaseg-siren-adapter-v2-ixi)
which lifts cross-site Dice to 0.69 with a 65,920-parameter MLP adapter.

## Reproducibility

The full reproduction log lives at
[`results/RESULTS.md`](https://github.com/basim-azam/inr-brain-seg/blob/main/results/RESULTS.md)
in the accompanying repository, including hardware, wall-clock, seed values,
and paper-deviation discussion.

## 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}
}

@inproceedings{vyas2025metaseg,
  title     = {Fit Pixels, Get Labels: Meta-Learned Implicit Networks for Image Segmentation},
  author    = {Vyas, Kushal and others},
  booktitle = {MICCAI},
  year      = {2025}
}

@inproceedings{sitzmann2020siren,
  title     = {Implicit Neural Representations with Periodic Activation Functions},
  author    = {Sitzmann, Vincent and others},
  booktitle = {NeurIPS},
  year      = {2020}
}
```

## License

Apache License 2.0.