medarc/nanopath-evals

NanoPath evaluation data

This is the immutable data mirror used by NanoPath probe protocol v2. It contains only the exact development records consumed by medarc/nanopath: selected THUNDER training/validation images, prepared development-only slide caches, and the two PathoROB subsets. manifest.json records SHA-256 checksums and binds the snapshot to the checked-in benchmark manifests.

No official THUNDER, HEST, or CPTAC classification test record is included. HEST is absent. CPTAC appears only as the pre-existing CPTAC-PDA overall-survival development probe. PanNuke contains Fold1 and Fold2 only; Fold3 is absent. Tolkach ESCA's TCGA rows are removed. ESCA classification intentionally retains selected TCGA images in the probe's training side, while every scored ESCA validation image is from UKK.

Evaluation score (CRoMa revision, September 2026)

The final score uses 30% classification, 15% segmentation, 17.5% progression, 15% mutation, 7.5% survival, and 15% robustness. The five predictive protocols and every mirrored image and split are unchanged by this revision.

Robustness uses CRoMa on all Camelyon tiles and all non-TCGA Tolkach ESCA tiles. For each tile, exclude its slide and find the five nearest same-biology/different-center neighbors and five nearest different-biology/same-center neighbors. With mean cosine distances d_SO and d_OS, the signed margin is (d_OS - d_SO) / (d_OS + d_SO). Map each cohort median to (1 + median) / 2 and average the two cohorts. Every tile must have the required neighbors. The 10th percentile, lower-tail mean, and nonpositive fraction are diagnostics.

Nanopath's implementation uses exact chunked PyTorch searches and float64 distance arithmetic, with no CRoMa package dependency. The reference commit and score definition are recorded in manifest.json; its original payload hashes and data-selection commit are preserved. See the protocol and comparative evidence.

Runtime dataset	Download size
bach	4.54 GiB
bracs	9.78 GiB
break_his	0.63 GiB
crc	0.88 GiB
esca	0.11 GiB
mhist	0.23 GiB
pcam	0.10 GiB
spider_breast	0.37 GiB
spider_colorectal	0.38 GiB
spider_skin	0.58 GiB
spider_thorax	0.40 GiB
wilds	0.14 GiB
pannuke	1.27 GiB
segpath_epithelial	18.68 GiB
segpath_lymphocytes	20.75 GiB
ucla_lung	1.40 GiB
surgen	101.42 GiB
leopard_bcr	15.52 GiB
cptac_pda_os	14.15 GiB
pathorob	0.68 GiB

Total downloadable payload: 192.01 GiB. Classification/segmentation image trees are tarred to keep the Hub repository small; the NanoPath preparation script downloads and extracts them automatically.

The mirror does not relicense upstream datasets. Each component remains governed by its original terms and should be used for research in accordance with those terms. Sources include BACH (Zenodo 3632035), BRACS, BreaKHis, NCT-CRC-HE-100K, ESCA (Zenodo 7548828), MHIST, PCam, SPIDER, CAMELYON17-WILDS, PanNuke, SegPath (Zenodo 7412731 and 7412529), PathoBench/UCLA, SurGen, LEOPARD, CPTAC, and PathoROB. See the NanoPath and THUNDER repositories for citations and split construction.