AI & ML

The modelling core: what makes this problem hard, both training pipelines end to end, the network architectures and loss suite, and how predictions are evaluated against the competition metric.

Pipelines The custom MONAI training pipeline end to end, the nnU-Net v2 lifecycle, and a direct comparison of the two. Models & Losses Network architectures, the dual-head distance-field design, and the full loss suite with every component and its default weight. Evaluation & Metrics The competition metric definitions, the composite weighting, and the offline post-processing lab that turns probability volumes into a score.

1. What makes this problem hard

Four properties of the task drive essentially every modelling decision in this section. They are worth stating plainly before the architectures, because each one explains a design choice that otherwise looks arbitrary.

Property	Consequence for the model
Extreme class imbalance. The surface occupies far less than one percent of voxels in a patch.	Patch sampling must be foreground-weighted, and the loss must be composed rather than a single term — plain cross-entropy would be dominated by empty space.
The target is a thin manifold. A sheet is a couple of voxels thick across hundreds of voxels of extent.	Downsampling along the sheet’s normal direction destroys it, so the encoder is deliberately asymmetric — it reduces resolution in the cross-sheet plane while preserving it along the scroll axis.
Topology is scored discretely. Betti numbers are integers, so a near-miss earns nothing.	A binary mask alone gives the optimiser no gradient for “almost connected”. The network therefore also regresses a continuous distance field, and the loss includes explicitly topological terms.
Adjacent scroll wraps look alike. Two neighbouring sheets of papyrus produce similar local intensity.	Distance supervision and continuity terms discourage the model from bridging across the gap between wraps — the single most damaging topological error.

2. How the pages fit together

Pipelines describes process: how a run is launched, what the training loop does, and the sequence of steps in each of the two pipelines. Models & Losses describes mechanism: the networks themselves and the objective they are trained against. Evaluation & Metrics closes the loop: how a trained model is measured, and the substantial post-processing machinery that sits between a raw prediction and a final score.