--- pretty_name: NSCLC Thesis — WSI + Embeddings (staging) tags: - pathology - whole-slide-image - histopathology - NSCLC - embeddings - computational-pathology task_categories: - other size_categories: - 100B TCGA under wsi/TCGA-*: NCI GDC open-access / NIH GDS unrestricted-access terms. CPTAC under wsi/CPTAC-*: Creative Commons Attribution 4.0 (CC BY 4.0) via TCIA (CPTAC-LUAD, CPTAC-LSCC tissue slide images), plus TCIA Data Usage Policy. Derived embeddings in embeddings/ inherit the same source attribution obligations. extra_gated_prompt: > This dataset redistributes diagnostic WSIs from (1) TCGA via NCI GDC (wsi/TCGA-LUAD, wsi/TCGA-LUSC) and (2) CPTAC via TCIA (wsi/CPTAC-LUAD, wsi/CPTAC-LSCC; CC BY 4.0), plus derived embeddings when present. By downloading you agree: (1) not to attempt to re-identify research participants; (2) to cite TCGA/GDC for TCGA paths and CPTAC/TCIA (with DOI) for CPTAC paths; (3) to abide by NIH GDS unrestricted-access terms (TCGA) and the TCIA Data Usage Policy (https://www.cancerimagingarchive.net/data-usage-policies-and-restrictions/) including requiring the same of downstream users; (4) not to use the data for clinical diagnosis or patient care. extra_gated_fields: Affiliation: text I agree not to attempt re-identification of participants: checkbox I will cite TCGA/GDC and CPTAC/TCIA as applicable: checkbox I will abide by TCIA Data Usage Policy for CPTAC folders: checkbox --- # NSCLC-THESIS-WSI-DATASET-EMBEDDING Public staging mirror for a thesis pipeline (gated acknowledgment before download). **One dataset URL** — raw slides under `wsi/`, precomputed features under `embeddings/`. Pull only the folder you need. Repo: [`thanminh01/NSCLC-THESIS-WSI-DATASET-EMBEDDING`](https://huggingface.co/datasets/thanminh01/NSCLC-THESIS-WSI-DATASET-EMBEDDING) WSI upload of all four cohorts is **complete**. Integrity checks below passed on **2026-08-14**. Embeddings are still empty stubs (thesis Stage 2). ## Layout ```text README.md manifests/ tooling/ # download scripts + attribution (see tooling/ATTRIBUTION.md) wsi/ TCGA-LUAD/ # 541 DX .svs (+ UUID sidecars) TCGA-LUSC/ # 512 DX .svs CPTAC-LUAD/ # 1137 .svs (flat) CPTAC-LSCC/ # 1081 .svs (flat) clinical/ # per-case companion data (GDC open-access + TCIA), see below TCGA-LUAD/{pathology-reports,clinical-supplement,biospecimen-supplement,cases-json}/ TCGA-LUSC/ # same four subfolders CPTAC-LUAD/cases-json/ CPTAC-LSCC/cases-json/ CPTAC-tcia-pathology-portal-cohort.csv -index.csv # slide -> case -> artifact availability embeddings/ README.md titan/… # later — TITAN retrieval embeddings gigapath/… # later — Prov-GigaPath features wsi_llava/… # later — cached descriptions / related ``` | Path | Contents | Approx. size | Status | |---|---|---|---| | `wsi/TCGA-LUAD/` | 541 diagnostic slides | ~415 GB | uploaded, checksum PASS | | `wsi/TCGA-LUSC/` | 512 diagnostic slides | ~409 GB | uploaded, checksum PASS | | `wsi/CPTAC-LUAD/` | 1137 slides | ~429 GB | uploaded, checksum PASS | | `wsi/CPTAC-LSCC/` | 1081 slides | ~415 GB | uploaded, checksum PASS | | `clinical/` | pathology reports, clinical + biospecimen supplements, cases JSON, indexes | ~600 MB, 5,315 files | uploaded 2026-08-14 | | `embeddings/*` | precomputed case-bank artifacts | empty stubs | after GPU Stage 2 | | `manifests/` | GDC manifests, TCGA md5 list, CPTAC `.sums` + verify reports | small | uploaded | ## Selective download (use `hf`) ```bash REPO=thanminh01/NSCLC-THESIS-WSI-DATASET-EMBEDDING pip install -U "huggingface_hub[hf_transfer]" export HF_HUB_ENABLE_HF_TRANSFER=1 # one WSI cohort hf download "$REPO" --repo-type dataset \ --include "wsi/TCGA-LUAD/**" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING hf download "$REPO" --repo-type dataset \ --include "wsi/TCGA-LUSC/**" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING hf download "$REPO" --repo-type dataset \ --include "wsi/CPTAC-LUAD/**" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING hf download "$REPO" --repo-type dataset \ --include "wsi/CPTAC-LSCC/**" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING # embeddings only (when populated) hf download "$REPO" --repo-type dataset \ --include "embeddings/titan/**" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING # clinical companion data only (small) hf download "$REPO" --repo-type dataset \ --include "clinical/**" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING # manifests / card hf download "$REPO" --repo-type dataset \ --include "manifests/**" \ --include "README.md" \ --local-dir ./NSCLC-THESIS-WSI-DATASET-EMBEDDING ``` ## What is uploaded **WSI:** full cohort trees as stored after archive download. - **TCGA:** UUID directories under `wsi/TCGA-LUAD/` and `wsi/TCGA-LUSC/` (diagnostic **`DX`** `.svs` plus sidecars such as `annotations.txt` / `.parcel`). Source manifests: GDC `2023-10-03`. - **CPTAC:** flat `.svs` under `wsi/CPTAC-LUAD/` (package LUAD v12) and `wsi/CPTAC-LSCC/` (package LSCC v10), plus package `.sums` next to the slides and under `manifests/`. **Embeddings:** empty stubs. Filled later from thesis Stage 2 precompute (TITAN, Prov-GigaPath, WSI-LLaVA caches). Derived files still require citing the underlying TCGA/CPTAC sources. ## Clinical companion data (`clinical/`) Everything freely available from GDC and TCIA for the cases behind these slides, fetched 2026-08-14 with GDC-reported MD5 verified on every file (0 failures). Scope = exactly the cases of the 3,271 `.svs` on this repo. | Cohort | Cases | Cases-API JSON | Pathology report PDFs | Clinical suppl. XML/TSV | Biospecimen suppl. | |---|---|---|---|---|---| | TCGA-LUAD | 478 | 478/478 | 479 (every case ≥1) | 575 | 972 | | TCGA-LUSC | 478 | 478/478 | 478 (every case ≥1) | 546 | 966 | | CPTAC-LUAD | 244 | 229/244 | none exist on GDC | — | — | | CPTAC-LSCC | 212 | 108/212 | none exist on GDC | — | — | - `cases-json/.json` — GDC Cases API: primary diagnosis, ICD-O morphology, AJCC stage/TNM, demographics, exposures, samples down to slide-level `percent_tumor_nuclei`. - `pathology-reports/` — TCGA scanned report PDFs (GDC open-access). - `clinical-supplement/`, `biospecimen-supplement/` — per-case XML plus project-wide TSVs (drug, radiation, follow-up, slide tables). - `-index.csv` — one row per slide: case id + artifact counts. Gaps are explicit, not dropped. **Honest limits:** GDC hosts **no pathology reports or supplements for CPTAC-3** — structured Cases JSON is the maximum there. 15 CPTAC-LUAD and 104 CPTAC-LSCC cases (confirmatory-cohort slides on TCIA) have **no GDC record at all**; for those, `CPTAC-tcia-pathology-portal-cohort.csv` (per-slide histology, sex, age, tumor %, GDC/PDC links from the TCIA pathology portal) is the only free metadata. `FETCH-SUMMARY.json` lists the exact missing case IDs. ## Integrity Verified **2026-08-14**. Method: Hugging Face LFS/Xet object `sha256` + `size` compared to (a) GDC manifest sizes for TCGA, (b) streamed local SHA-256 for CPTAC. Zero size or SHA mismatches on `.svs`. | Cohort | Files | Check | Result | When | |---|---|---|---|---| | `wsi/TCGA-LUAD/` | 541 `.svs` | Hub LFS sha256 present; Hub size = GDC manifest size | PASS 541/541 | 2026-08-14 | | `wsi/TCGA-LUSC/` | 512 `.svs` | same | PASS 512/512 | 2026-08-14 | | TCGA `.svs` (prior) | 1053 | local SHA-256 = Hub LFS oid (Archive copy, since deleted) | PASS 1053/1053 | 2026-08-09 | | `wsi/CPTAC-LUAD/` | 1137 `.svs` | local SHA-256 = Hub LFS oid; size match | PASS 1137/1137 | 2026-08-14 | | `wsi/CPTAC-LSCC/` | 1081 `.svs` | local SHA-256 = Hub LFS oid; size match | PASS 1081/1081 | 2026-08-14 | Local TCGA md5 vs GDC (before upload): `manifests/md5_compare_report.txt` → `VERIFY_STATUS=PASS`, 1053/1053. Hash list: `manifests/downloaded_md5sum_hashes.txt`. CPTAC package checksums: `manifests/CPTAC-LUAD_v12.sums`, `manifests/CPTAC-LSCC.sums`. Slide indexes and local verify reports are in `manifests/` as well. TCGA UUID sidecars (`annotations.txt`, `.parcel`) are on Hub; small git blobs do not carry an LFS sha256. They are not used as the integrity gate — the `.svs` objects are. ## Intended use Research / educational computational pathology for this thesis. **Not for clinical care.** ## License / attribution (both sources) This Hub repo claims **no ownership** of the underlying archives. **Attribute both sides** when using any part of this dataset (including embeddings derived from them). ### TCGA paths (`wsi/TCGA-*`, and embeddings from them) - Source: [TCGA](https://www.cancer.gov/tcga) via [NCI GDC](https://gdc.cancer.gov/) - Terms: NIH GDS unrestricted-access; no re-identification - Acknowledgment: > The results shown here are in whole or part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga. Data were obtained from the NCI Genomic Data Commons (GDC): https://gdc.cancer.gov/. ### CPTAC paths (`wsi/CPTAC-*`, and embeddings from them) - Source: [TCIA](https://www.cancerimagingarchive.net/) — CPTAC lung histopathology - License: **CC BY 4.0** on Tissue Slide Images for [CPTAC-LUAD](https://www.cancerimagingarchive.net/collection/cptac-luad/) and [CPTAC-LSCC](https://www.cancerimagingarchive.net/collection/cptac-lscc/) - Policy: [TCIA Data Usage Policies and Restrictions](https://www.cancerimagingarchive.net/data-usage-policies-and-restrictions/) - Citations: > National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC). (2018). The Clinical Proteomic Tumor Analysis Consortium Lung Adenocarcinoma Collection (CPTAC-LUAD) (Version 13) [dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/K9/TCIA.2018.PAT12TBS > National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC). (2018). The Clinical Proteomic Tumor Analysis Consortium Lung Squamous Cell Carcinoma Collection (CPTAC-LSCC) (Version 15) [dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/K9/TCIA.2018.6EMUB5L2 ## Download tooling (transparency) Under `tooling/`: - Mirror of **[GeorgeBatch/TCGA-lung-histology-download](https://github.com/GeorgeBatch/TCGA-lung-histology-download)** (scripts + GDC manifests used for this pull). See `tooling/ATTRIBUTION.md` for license status and citations. - Thesis run notes under `tooling/thesis-tcga-download-notes/`. Upstream GitHub lists **no SPDX license** on that code repo (`license: null`); we mirror it for provenance only. **TCGA/GDC terms still govern all `wsi/TCGA-*` files.** ## Limitations - Very large binaries under `wsi/`; plan disk before a full cohort pull. - Clinical tables not bundled. - Repo is thesis staging, not an official TCGA/TCIA release.