zenodo-v2/MODEL_CARD.md

SimCortex v2 Model Card

Model overview

SimCortex v2 is a deep-learning pipeline for reconstructing four cortical surfaces from T1-weighted structural MRI:

• left pial surface
• left white-matter surface
• right pial surface
• right white-matter surface

The pipeline operates in MNI152 space and combines volumetric segmentation, cortical-surface initialization, and simultaneous deformation of the four cortical surfaces.

Intended use

The released models are intended for research use in cortical surface reconstruction and downstream morphometric analysis.

They are not intended for standalone clinical diagnosis or treatment decisions.

Pipeline stages

1. T1-weighted MRI preprocessing into MNI152 space.
2. Nine-class 3D U-Net segmentation.
3. Initial cortical-surface generation.
4. Simultaneous deformation of left/right white and pial surfaces.

Released checkpoints

Segmentation

File:

seg/seg_best_dice.pt

Role:

Final segmentation checkpoint used by the released pipeline.

SHA256:

2cea6dc1606958dbdb2aea50748b73ff73e604b0c367b3b1588cef7a112c8e06

Primary deformation checkpoint

File:

deform/deform_best_rmse.pth

Role:

Recommended deformation checkpoint used for the final SimCortex v2 paper evaluation.

SHA256:

3fb84c917953e479e72d59291837bd6b08f40fb197534046b95dc5a452e86b95

Secondary deformation checkpoint

File:

deform/deform_best_model.pth

Role:

Alternative deformation checkpoint retained for reproducibility and comparison.

SHA256:

d30d3218069a6eb395dacd3ff0dd085a93f6f0102616eed809b90c362b885b40

Model file format

The released .pt and .pth files are PyTorch state dictionaries intended for inference.

Training-resume checkpoints, optimizer states, and scheduler states are not included.

Training data split

The released training split contains 515 subjects from HCP-YA and OASIS-1:

• training: 361
• validation: 77
• test: 77

By dataset:

HCP-YA:

• 70 train
• 15 validation
• 15 test
• 100 total

OASIS-1:

• 291 train
• 62 validation
• 62 test
• 415 total

The exact split is stored in:

splits/dataset_split.csv

SHA256:

68f42ef3ed09373685de5d1dd144c7d3d72b7f8083850057dcccd5780756188d

Evaluation cohort

The public evaluation cohort definition is stored in:

splits/evaluation_sample40.csv

It contains:

• 560 cases
• 14 datasets
• 40 cases per dataset

SHA256:

2c41b21afea327756c8d7fccc30bd03e461670d59dc00f941e8cba06fcf59842

Configuration files

The archive includes the YAML configurations used by the released pipeline:

• configs/seg/train.yaml
• configs/seg/inference.yaml
• configs/seg/eval.yaml
• configs/initsurf/generate.yaml
• configs/deform/train.yaml
• configs/deform/inference.yaml
• configs/deform/eval.yaml

Surface ordering

The deformation implementation uses the following four-surface order:

lh_pial
lh_white
rh_pial
rh_white

Input and output

Input:

T1-weighted MRI processed into the reference space expected by the SimCortex pipeline.

Output:

Left/right white-matter and pial cortical surface meshes.

Data availability

The MRI datasets themselves are not included in this model archive.

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

FRDR dataset DOI:

10.20383/103.01769

Limitations

The released models were developed and evaluated using the datasets and processing conventions described in the SimCortex v2 study.

Performance may vary for data acquired with different scanners, protocols, populations, preprocessing pipelines, or image quality.

The models are provided for research use and should be independently validated before use in new settings.

Reproducibility

The archive includes:

• pretrained model state dictionaries
• frozen YAML configurations
• training/validation/test split
• final evaluation cohort definition
• provenance metadata
• SHA256 checksums

The source-code repository should be used together with this archive for inference and reproduction of the released pipeline.