Brain tumour segmentation on multi-contrast MRI and autism classification from resting-state fMRI, each with a model built around a property of the data that the usual baseline ignores: the left-right symmetry of the brain, and the estimation noise of a connectome measured from a few minutes of signal.
MSD Task01 (BraTS): 750 scans, four contrasts each. Every encoder level can read the features of its mirror image across the midline, and the network is pretrained without labels by reconstructing hidden tissue from its contralateral twin.
ABIDE I: 871 scans from 20 sites. Each scan's connectome is re-estimated on six sub-windows, and the classifier is trained on that spread rather than on one point per subject, which penalises connections that are not reproducible within a scan.
Four co-registered contrasts with the reference annotation and the predictions of both models trained on 10% of the labels. Drag to rotate in 3D, scroll to move through slices.
Dice (higher is better) and HD95 in mm (lower is better)
The same U-Net backbone, data pipeline and training budget at every point; only the bilateral blocks and the pretraining differ. Bars are 95% bootstrap intervals over test scans.
Each dot is a test scan; above the diagonal, ours is better.
10% of labels, mean Dice
HD95 in mm, mean over test scans
The connections the classifier weighs most, drawn between region centres in MNI space. Warm-coloured edges push the prediction towards autism, blue ones towards typical development. Weights are those of the model fitted on all 871 subjects, in units of standardised features.
These are weights of a linear classifier on standardised features, not group differences, and a single weight should not be read as an effect size.
largest absolute weights
All models are scored on the same 50 train/test splits (5 repeats of 10-fold cross-validation, stratified by site and diagnosis), and again with each of the 20 sites held out in turn.
ours against the strongest baseline on each atlas
mean over repeats with 95% bootstrap interval
difference to the strongest baseline on each split
accuracy on each held-out site, all five atlases
predicted probability against observed frequency, all five atlases
Before relying on a prediction: does the network know when it is wrong, do its probabilities mean what they say, can its volumes be followed over time, where does it fail, and what happens when the images get worse. Each line of the scorecard has a written rule; the evidence is below it.
each dot is a test scan; further right means more uncertain
mean error of the scans that remain
within five voxels of the predicted or true region
predicted against reference volume, log scale
Dice by tertile of whole-tumour volume
mean Dice on the same scans under the same perturbations
Dice lost when one contrast is removed, ours
for comparison
share of gradient × input inside the segmented tumour, by contrast
how much the exchange changes the features, and whether left-right asymmetry alone finds the tumour
Uncertainty and gradient maps of the showcase scans are in the tumour viewer, under Overlay. Rules and limits: the plan page.
An AUC alone does not say whether the probabilities can be believed, whether the classifier is reading the site or the head motion instead of the brain, whether it holds in every group, or what it looks at. The scorecard answers with rules written down in advance; the evidence follows.
predicted probability of autism against the observed fraction
raw and recalibrated; the recalibration is fitted on other folds only
accuracy when only the most confident subjects are scored
classifiers that see only nuisance variables, on the same folds
AUC of the noise-aware model inside each group
strongest regions of the shipped classifier
Features are entries of the matrix logarithm of the connectivity matrix, so a single entry is not one edge. Regions and lobe blocks are the safer reading.
weight per feature in each block of regions
30 bootstrap refits of the classifier
AUC after deleting the features an explanation names
The connectome classifier runs here in plain JavaScript: shrinkage covariance, matrix logarithm by Jacobi rotations, standardisation and the linear score, on scans of subjects that model never saw during training. The segmentation network runs through ONNX Runtime Web.
held-out subjects of one cross-validation fold
the bilateral U-Net, 10.8M parameters, fetched from Hugging Face
Three pictures that carry most of the method. The documentation has the derivations.
Every position reads its mirror image across the midline. Move the probe.
Each hidden cube has a visible twin on the other side. Compare with ordinary random masking.
Each subject is a cloud of window connectomes rather than one point.
One Nextflow pipeline runs both studies, locally, in Docker or on a cluster; a test profile runs every stage on synthetic data in a few minutes.
nextflow run mahdisabetkish/brainssl -profile docker,gpu
pip install -e . brainssl download abide && brainssl download msd brainssl prepare abide && brainssl prepare msd python scripts/run_smri.py python scripts/run_fmri.py brainssl report
brainssl-infer segment --model exports/segmentation \ --input scan.nii.gz --output seg.nii.gz