---
license: cc-by-nc-4.0
pretty_name: TCGA-LUAD Prov-GigaPath Tile Embeddings (20x, 256px)
task_categories:
  - image-feature-extraction
tags:
  - pathology
  - computational-pathology
  - histopathology
  - foundation-model
  - TCGA
  - TCGA-LUAD
  - lung-adenocarcinoma
  - whole-slide-image
  - embeddings
size_categories:
  - 100K<n<1M
---

# TCGA-LUAD — Prov-GigaPath Tile Embeddings

Pre-extracted **1536-dimensional tile embeddings** for TCGA-LUAD diagnostic H&E
whole-slide images, produced with the **Prov-GigaPath** tile encoder
(`prov-gigapath/prov-gigapath`). Part of a four-encoder benchmark in which every
model was run on the **identical tiles** for a fair comparison.

## Tiling (identical across all encoders in the study)
- Source: raw TCGA-LUAD `.svs` (GDC, open-access diagnostic slides).
- 20× / **0.5 microns per pixel**, MPP-aware (handles mixed 20×/40× scans).
- **256 × 256 px** tiles, non-overlapping, **full tissue coverage** (Otsu on HSV saturation).
- Prov-GigaPath's official transform (Resize 256 → CenterCrop 224 → ImageNet norm) applied per tile.

## Format (per slide, CLAM-compatible)
`<slide_id>.h5`:
- `features` — `(N, 1536)` float32
- `coords` — `(N, 2)` int32, level-0 (x, y) of each tile
- attrs: `patch_size=256`, `patch_level=0`, `target_mpp=0.5`, `model`, `embed_dim`

`slide_id` = TCGA barcode + GDC file UUID. Task labels (PDS / TP53 / EGFR / KRAS)
are in `labels/` (schema: `case_id, slide_id, label`).

## Cohort
531 slides / 478 patients (diagnostic FFPE, primary tumor).

## Attribution & license
- **Source model:** Prov-GigaPath — Xu et al., *A whole-slide foundation model for
  digital pathology from real-world data*, **Nature** 2024. Model under **Apache-2.0**.
- These **derived embeddings** are released under **CC-BY-NC 4.0**, honoring the
  model's research-only / non-clinical intent. **Not for clinical use.**
- Source images: TCGA-LUAD (NIH/GDC), open-access.

Provided for non-commercial academic research. Please cite the Prov-GigaPath paper
and TCGA when using these features.