Which method wins on your slide?

ARI vs ground-truth domains, averaged over 3 seeds, for every registered domain-detection method on every benchmark dataset. Pick a dataset, pick a family, and the site tells you what to reach for first.

–methods

–datasets

–platforms

–spots / cells

Dataset Platform / assay Tissue Domain count Matrix sparsity Family Task Sort

Leaderboard

Numbers are mean ARI over 3 seeds. win marks the best method for a dataset.

ARI heatmap

Rows = methods, columns = datasets. Colour intensity encodes ARI (0 to max in the visible slice).

Per-method ARI across datasets

One panel per method; dots are per-dataset means, whiskers span the min→max across seeds.

Methods included

About

Metric. Adjusted Rand Index (ARI) against the annotated ground-truth domain label for each dataset. Higher is better. 3 seeds per (method, dataset) cell.

Ground truth. DLPFC slices use spatialLIBD_layer from Maynard et al. (2021); MERFISH mouse hypothalamus uses the coarse Cell_class compartment collapse from Moffitt et al. (2018); the SlideseqV2 mouse hippocampus uses the cluster labels from Stickels et al. (2021), collapsed to 6 compartments (see domain_mappings.json). "Ambiguous" cells in MERFISH are bucketed as other.

Reproducibility. Every data point is a checkpointed (method, dataset, seed) cell. The generate.py script in this folder rebuilds data.json from the CSV artefacts in 5x15_spatial_aware/ and benchmark_crossplatform/; the GH Pages workflow (.github/workflows/pages.yml) runs it on every push to main.

Built from – · –