Hugging Face's logo

Datasets:
qklee
/
PathSeek-MAS

Tasks:
Tabular Classification
Image Classification
Size:
n<1K
Tags:
pathology
ovarian-cancer
HGSOC
platinum-resistance
virtual-mIF
multimodal
License:
Dataset card Data Studio Files
xet
 Community
PathSeek-MAS / README.md
qklee's picture
qklee
docs: simplify WSI source column in cohort table
672ba8f verified
preview code
|
Raw
History Blame Contribute Delete
2.82 kB
metadata 
license: cc-by-4.0
task_categories:
  - tabular-classification
  - image-classification
tags:
  - pathology
  - ovarian-cancer
  - HGSOC
  - platinum-resistance
  - virtual-mIF
  - multimodal
  - computational-pathology
pretty_name: PathSeek-MAS (HGSOC Platinum-Response Multimodal Features)
size_categories:
  - n<1K

PathSeek-MAS — Multimodal WSI Features for HGSOC Platinum-Response Prediction

Derived per-slide features supporting the PathSeek-MAS study (npj Digital Medicine). The task is platinum response prediction in high-grade serous ovarian carcinoma (HGSOC) from H&E whole-slide images (WSIs), using a dual-agent model:

  • Virtual Protein Agent — virtual multiplex immunofluorescence (mIF): 21 protein markers predicted per patch by GigaTIME (gigatime_csv/).
  • Morphology Agent — UNI foundation-model patch embeddings (uni_pt/).

Cohorts (503 WSIs)

Cohort Slides Sensitive (label 1) Resistant (label 0) WSI source
TCGA-OV 156 117 39 GDC
PTRC-HGSOC 347 202 145 Chowdhury et al.

Label convention (both cohorts): label == 1 = Sensitive, label == 0 = Resistant. Ground truth: TCGA PlatinumStatus (Zhang et al., mmc2.xlsx); PTRC Tumor response (sensitive→1, refractory→0).

Files 

TCGA-OV/   PTRC-HGSOC/
  gigatime_csv/   # one CSV per slide: virtual-mIF predicted protein values per patch
  uni_pt/         # one .pt per slide: UNI patch embeddings (dict: 'features' [N,1024], 'coords' [N,2])
  labels.csv      # slide_id, case_id, label

gigatime_csv/*.csv schema (27 columns)

patch_name, slide_id, row, col, then 23 channels (21 protein markers + 2 staining channels): DAPI, TRITC, Cy5, PD-1, CD14, CD4, T-bet, CD34, CD68, CD16, CD11c, CD138, CD20, CD3, CD8, PD-L1, CK, Ki67, Tryptase, Actin-D, Caspase3-D, PHH3-B, Transgelin. row, col are the patch grid coordinates (used to place values spatially on the WSI). Note: DAPI and Cy5 are staining/imaging channels, not protein markers; the remaining 21 are antibody-targeted proteins.

uni_pt/*.pt

torch.load(...) returns a dict with features ([n_patches, 1024] float) and coords ([n_patches, 2] int patch coordinates).

Preprocessing used by the model

  • Virtual-protein input: TCGA → log1p, PTRC → logit, then per-patch row-standardization.
  • Both agents pool patches via gated attention; fused by a cross-modal gating head.

Not included

  • Raw WSIs (~200 GB): TCGA-OV via the NCI GDC; PTRC-HGSOC via the original study. Not re-hosted.
  • Original third-party clinical supplementary tables — please cite the source publications.

Code

Analysis code: https://github.com/qklee/PathSeek-MAS

Citation

PathSeek-MAS (npj Digital Medicine). Citation to be added upon publication.