Research artifact

BioCluster

A benchmark of when unsupervised and self-supervised visual representations recover reproducible, biologically relevant structure from biomedical imagery -- and how stable that structure is across seeds, algorithms, embedding choices, and acquisition sites.

What this site is not

Nothing here is a diagnostic tool or a clinical finding. Clusters are described as latent structure or morphological groupings, never as discovered diseases. See Methods for full scope and limitations.

Datasets

Representations benchmarked

Clustering runs

Git SHA

unknown

Central question

Under what conditions do unsupervised and self-supervised visual representations recover reproducible, biologically relevant structure from biomedical imagery, and how stable are those conclusions across datasets, seeds, algorithms, and embedding choices? See the full research questions and hypotheses.

Datasets in this release

PathMNIST (histopathology)

histopathology (H&E-stained colorectal tissue patches)

License: CC BY 4.0
Classes: 9
Images (train/val/test): 89996/10004/7180

Full provenance →

BloodMNIST (blood cell microscopy)

blood cell microscopy (stained peripheral blood smear, single-cell crops)

License: CC BY 4.0
Classes: 8
Images (train/val/test): 11959/1712/3421

Full provenance →

Explore

Representation comparison

External validity, stability, linear-probe performance, retrieval quality side by side.

Stability benchmark

Seed-to-seed and perturbation robustness for every representation.

Technical confounding

Does clustering track clinical site rather than biology?

Embedding explorer

UMAP/t-SNE views, with an explicit visualization-only warning.

Cluster inspector

Per-cluster size, label enrichment, and technical metadata.

Manuscript

Full write-up, claims-to-evidence mapping, and reviewer rebuttal.