---
license: cc-by-nc-4.0
tags:
  - medical-imaging
  - segmentation
  - brain-extraction
  - skull-stripping
  - mri
  - monai
  - segresnet
  - pytorch
  - pig
  - veterinary
library_name: pytorch
pipeline_tag: image-segmentation
---

# Pig Brain SegResNet — brain extraction for porcine T2/CISS MRI

> ⚠️ **License is a placeholder (`cc-by-nc-4.0`).** Confirm the final license
> before making this repo public / before publication.

3D `SegResNet` (MONAI) for binary brain extraction (brain vs. background) on
sub-millimeter porcine **T2-weighted / 3D-CISS** MRI. This is the deep-learning
component of the pipeline described in *"Automated Brain Extraction Using Deep
Learning and Atlas-Guided Refinement."*

The released pipeline pairs this model with **ANTs SyN template refinement**;
the model alone provides a fast, template-free brain mask and the initial mask
that anchors the iterative curation workflow.

## Files
- `pig_brain_segresnet.pth` — model `state_dict` (load into the architecture below).
- `config.yaml` — full training/inference config (architecture + post-processing).

## Architecture
| field | value |
|---|---|
| init_filters | 32 |
| blocks_down | (1, 2, 2, 4) |
| blocks_up | (1, 1, 1) |
| in / out channels | 1 / 2 |
| spatial_dims | 3 |
| params | ~18.8 M |

## Preprocessing
LPS orientation → 1 mm isotropic resampling (bilinear) → foreground crop →
spatial pad to 128³ → per-channel non-zero z-score normalization.

## Inference
Gaussian-weighted sliding window (ROI 128³, overlap 0.75) with flip TTA, then:
softmax → Gaussian smooth → threshold 0.6 → remove small objects (min 500) →
keep largest connected component → fill holes. Spatial preprocessing is inverted
to return the mask to native resolution.

## Usage
```python
import torch
from omegaconf import OmegaConf
from huggingface_hub import hf_hub_download
from monai.networks.nets import SegResNet

repo = "madhavappaneni9/pig-brain-segresnet"
cfg = OmegaConf.load(hf_hub_download(repo, "config.yaml"))
weights = hf_hub_download(repo, "pig_brain_segresnet.pth")

model = SegResNet(
    spatial_dims=cfg.model.spatial_dims,
    init_filters=cfg.model.init_filters,
    in_channels=cfg.model.in_channels,
    out_channels=cfg.model.out_channels,
    blocks_down=tuple(cfg.model.blocks_down),
    blocks_up=tuple(cfg.model.blocks_up),
    dropout_prob=0.0,
)
model.load_state_dict(torch.load(weights, map_location="cpu", weights_only=False))
model.eval()
```
A full preprocessing + sliding-window + post-processing example is in the
[demo Space](https://huggingface.co/spaces/madhavappaneni9/pig-brain-segresnet-demo).

## Intended use & limitations
- **Intended:** research brain extraction for porcine T2/CISS MRI; baseline for
  future porcine brain MRI studies.
- **Not for clinical/diagnostic use.**
- Trained on a small cohort; defaults calibrated to ~1-month Yorkshire piglets.
  Out-of-distribution scans (other ages/cohorts/contrasts) may segment poorly and
  benefit from transfer-learning the checkpoint on a small new-cohort sample.

## Citation
Citation to be added on publication.