Skip to content

brainssl

Two studies on public neuroimaging data that ask the same question from different directions: how much can a model gain from structure in the data that the standard baseline ignores, when annotations are scarce or noisy?

Structural MRI. Brain tumour segmentation on four-contrast MRI from the Medical Segmentation Decathlon. A 3D U-Net gets one extra block per encoder level, which lets every voxel compare itself with its mirror image across the brain's midline, and is pretrained without labels by reconstructing hidden tissue from its contralateral twin. It is compared with the same U-Net trained from scratch, at 5, 10, 25 and 100% of the annotated scans.

Resting-state fMRI. Autism versus control classification from functional connectomes in ABIDE I. The reference pipeline (tangent-space connectivity and a penalised linear classifier) treats each subject's connectome as an exact point; here each scan's own estimation noise, measured by re-estimating the connectome on windows of the scan, becomes the training distribution. It is compared with that reference pipeline and with three published deep models.

Where to start

Repository layout

src/brainssl/
  data/          dataset download (parallel, resumable, streamed tar extraction)
  smri/          preprocessing, bilateral U-Net, masking, training, inference, metrics
  fmri/          ABIDE loading, SPD geometry, windows, classifiers, cross-validation, study
  export/        ONNX export, model cards, deployment bundles
  report/        tables and paired tests from run artefacts
  stats.py       intervals and paired tests
rust/            brainssl-infer, a Rust binary for both models (ONNX Runtime)
dashboard/       the static site behind the dashboard link
main.nf, modules/, conf/   the Nextflow pipeline
assets/          run matrices and the pre-registered model choices
docs/            this documentation