scalesurfer-volume

Combined FS5/6/7 model; recommended for unknown or mixed versions.

The model predicts a 256³ FreeSurfer-compatible aparc+aseg label volume. It was trained on FreeSurfer version(s) 5, 6, 7.

Training continuation

This checkpoint starts from the original all-data ScaleSurfer model trained with voxelwise cross-entropy, then continues training on pooled FreeSurfer 5/6/7 data with the FastSurferCNN v2.5.4-style objective.

The continuation objective is the sum of soft channel-wise Dice loss and boundary/class-weighted voxelwise cross-entropy (weights 1:1, including background), matching the structure of FastSurferCNN v2.5.4's CombinedLoss while operating on ScaleSurfer's 3D output.

Files:

• transunet3d.safetensors — normalized model weights without torch.compile prefixes.
• config.json — exact TransUNet3D architecture.
• metadata.json — source checkpoint and SHA-256 provenance.
• history.csv and summary.csv — staged training records.

Source checkpoint SHA-256: c45fbe2a7b30acd6bf6f869216f8e8e57c9802dcc4f211d57abd69af56a492d3.

This model is intended for research use and is not a clinical diagnostic device.