Hugging Face's logo

Datasets:
MahmoodLab
/
hest

Tasks:
Image Classification
Feature Extraction
Image Segmentation
Languages:
English
Size:
100B<n<1T
ArXiv:
Tags:
spatial-transcriptomics
histology
pathology
License:
Dataset card Files
xet
 Community
23

Duplicated samples across cohorts in HEST-1k v1.3.0 (42 samples, 19 cross-study)

#28
by Alesaraband - opened
Discussion
Alesaraband

Hi, I'm using HEST-1k for patient-stratified cross-validation on the breast IDC cohort. I've found what appears to be systematic sample duplication across source cohorts and wanted to flag it, as it affects patient-level CV.

I retrieved the LFS SHA256 of every file via the Hub API and grouped by content hash.

Findings in v1.3.0:

  • 21 groups of WSIs share identical content; 19 span two different source
    studies; 42 samples are affected.
  • The same duplication appears independently in st/ (15 groups),
    patches/ (16) and cellvit_seg/ (35).

Four mechanisms:

  1. One study deposited to two repositories, ingested twice (MEND41-55 <-> NCBI527-540, Villacampa et al.).
  2. Same tissue released under targeted and whole-transcriptome panels. (TENX28/29, TENX30/31, TENX23/24).
  3. Preservation-method comparison studies (TENX77/84, TENX78/83, TENX79/82).
  4. Same section profiled on two platforms and listed as separate samples (TENX169/TENX190 and TENX171/TENX189, Visium HD <-> Xenium 5k; NCBI886/887, Xenium <-> Xenium 5k).

The breast IDC case is the one with patient-level consequences. Twelve sample pairs across He et al. 2020 and Andersson et al. 2021 correspond to the same physical sections under different patient labels:

BC23810 = Patient F (SPA99/100/101 <-> SPA125/126/127)
BC23567 = Patient E (SPA102/103/104 <-> SPA128/129/130)
BC23903 = Patient G (SPA96/97/98 <-> SPA122/123/124)
BC24044 = Patient H (SPA93/94/95 <-> SPA119/120/121)

Evidence: wsis/SPA100.tif and wsis/SPA126.tif are byte-identical (SHA256 8406eeb1...). Eight of the twelve pairs have byte-identical cellvit_seg outputs, which can only arise from identical input images. All twelve share 71-78% of exact patch coordinates, against a maximum of 0.27% among sections from the same donor within one study (n=139) and 0.49% among
different donors (n=2769) - a ~150x separation with no intermediate values.

The implication is that the breast IDC cohort has 27 distinct donors rather than 31. Holding out BC23810 leaves the same sections in training under Patient F, and likewise for the other three.

With thanks for HEST - it has been central to my work.

Sign up or log in to comment