# SimCortex v2: Pre-trained Models and Reproducibility Files

This archive contains the finalized pre-trained model weights and reproducibility files for SimCortex v2, a deep-learning pipeline for simultaneous reconstruction of left/right white-matter and pial cortical surfaces from T1-weighted MRI.

The archive contains inference-ready PyTorch state dictionaries, configuration files, the training/validation/test split used for model development, and the evaluation cohort definition used for the final benchmark.

## Recommended checkpoint

The primary deformation checkpoint is:

`deform/deform_best_rmse.pth`

This is the recommended checkpoint and the one used for the final SimCortex v2 paper evaluation.

The archive also includes:

`deform/deform_best_model.pth`

as a secondary checkpoint retained for reproducibility and comparison.

The segmentation checkpoint is:

`seg/seg_best_dice.pt`

## Archive contents

```text
SimCortex_v2_pretrained_models/
├── README.md
├── MODEL_CARD.md
├── provenance.json
├── SHA256SUMS
├── configs/
│   ├── deform/
│   │   ├── eval.yaml
│   │   ├── inference.yaml
│   │   └── train.yaml
│   ├── initsurf/
│   │   └── generate.yaml
│   └── seg/
│       ├── eval.yaml
│       ├── inference.yaml
│       └── train.yaml
├── deform/
│   ├── deform_best_model.pth
│   └── deform_best_rmse.pth
├── seg/
│   └── seg_best_dice.pt
└── splits/
    ├── dataset_split.csv
    └── evaluation_sample40.csv
```

## Model weights

### Segmentation

`seg/seg_best_dice.pt`

Pre-trained weights for the 3D U-Net segmentation stage. The network predicts nine classes used by the subsequent cortical-surface initialization stage.

### Deformation

`deform/deform_best_rmse.pth`

Primary and recommended deformation checkpoint.

`deform/deform_best_model.pth`

Secondary deformation checkpoint retained for reproducibility and comparison.

The released `.pt` and `.pth` files contain model state dictionaries intended for inference. Optimizer state and training-resume checkpoints are not included.

## Data splits

`splits/dataset_split.csv` contains the HCP-YA and OASIS-1 subject split used for model development:

- 515 subjects total
- 361 training subjects
- 77 validation subjects
- 77 test subjects

Dataset composition:

- HCP-YA: 100 subjects
- OASIS-1: 415 subjects

`splits/evaluation_sample40.csv` defines the final evaluation cohort:

- 560 cases
- 14 datasets
- 40 cases per dataset

The evaluation cohort file contains only public case identifiers and does not contain workstation-specific paths.

## Configuration files

The `configs/` directory contains the frozen YAML configuration files associated with the released pipeline:

- segmentation training, inference, and evaluation
- initial-surface generation
- deformation training, inference, and evaluation

These files are also maintained in the SimCortex source-code repository.

## Pipeline

The released model corresponds to the following processing stages:

1. preprocessing of T1-weighted MRI into MNI152 space;
2. nine-class 3D U-Net segmentation;
3. initialization of left/right white and pial cortical surfaces;
4. simultaneous deformation of the four cortical surfaces.

The deformation stage reconstructs the following four surfaces:

```text
lh_pial
lh_white
rh_pial
rh_white
```

## Reference-space data

The MNI152 template itself is not redistributed in this archive.

Users should obtain the required reference-space resources through the software dependencies or sources documented by the SimCortex code repository.

## Dataset availability

The MRI datasets and derived neuroimaging data are not redistributed in this model archive.

The associated SimCortex dataset release is archived separately in the Federated Research Data Repository (FRDR).

FRDR dataset DOI:

`10.20383/103.01769`

## File integrity

SHA256 checksums for the released files are provided in:

`SHA256SUMS`

Additional checkpoint and cohort provenance is recorded in:

`provenance.json`

## Citation

Please cite the SimCortex v2 paper and the corresponding Zenodo release when using these pretrained models.

The final publication DOI and Zenodo DOI should be used once available.