README: concise overview of he2mif/, he2st/ and the new sdk/ folder
Browse files
README.md
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
---
|
| 2 |
license: other
|
| 3 |
license_name: mixed-per-folder
|
| 4 |
-
license_details: he2mif/ is MIT (MIPHEI-ViT Orion-CRC release); he2st/ is CC BY-NC-SA 4.0 (HEST-1k)
|
| 5 |
pretty_name: MuPaD examples
|
| 6 |
tags:
|
| 7 |
- pathology
|
|
@@ -17,119 +17,24 @@ size_categories:
|
|
| 17 |
|
| 18 |
# MuPaD examples
|
| 19 |
|
| 20 |
-
|
| 21 |
-
projects in [github.com/jinxixiang/MUPAD](https://github.com/jinxixiang/MUPAD).
|
| 22 |
-
Each folder mirrors the layout of the full dataset, so the released code runs
|
| 23 |
-
on it unchanged: the demos, prediction and scoring, and short training runs.
|
| 24 |
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
```
|
| 29 |
-
|
| 30 |
-
Download one folder at a time:
|
| 31 |
|
| 32 |
```bash
|
| 33 |
huggingface-cli download xiangjx/MuPaD-examples --repo-type dataset --include "he2mif/*" --local-dir examples
|
| 34 |
-
tar xf examples/he2mif/ORIONCRC_dataset_tile_20x.tar -C examples/he2mif
|
| 35 |
huggingface-cli download xiangjx/MuPaD-examples --repo-type dataset --include "he2st/*" --local-dir examples
|
| 36 |
```
|
| 37 |
|
| 38 |
-
|
| 39 |
-
the Hub's API rate limit on download.
|
| 40 |
-
|
| 41 |
-
## he2mif/ORIONCRC_dataset_tile_20x.tar
|
| 42 |
-
|
| 43 |
-
A 393 MB slice of the **ORION-CRC** tile dataset at 20x (512 x 512 px H&E,
|
| 44 |
-
333 x 333 px 17-channel mIF and nuclei masks), in the same layout as the full
|
| 45 |
-
[Zenodo record 15340874](https://zenodo.org/records/15340874):
|
| 46 |
-
|
| 47 |
-
```
|
| 48 |
-
he/ H&E tiles (.jpeg)
|
| 49 |
-
if/ mIF tiles, 17 channels (.tiff, uint8)
|
| 50 |
-
nuclei/ nuclei instance masks (.tiff, int32 slide-global labels)
|
| 51 |
-
csv_nuclei_pos/ per-slide nuclei positivity tables (.csv)
|
| 52 |
-
train_dataframe.csv 600 tiles, 100 from each of 6 training slides
|
| 53 |
-
val_dataframe.csv 120 tiles, 60 from each of the 2 validation slides
|
| 54 |
-
test_dataframe.csv 120 tiles, 60 from each of the 2 test slides
|
| 55 |
-
demo_dataframe.csv the 2 validation tiles shown in 5-HE2mIF/demo.ipynb
|
| 56 |
-
slide_dataframe.csv, channel_stats.json
|
| 57 |
-
```
|
| 58 |
-
|
| 59 |
-
**Selection.** Tiles with tissue (`Hoechst_prop >= 0.2`) were sampled uniformly
|
| 60 |
-
at random per slide (seed 0), so the split is representative of tissue tiles,
|
| 61 |
-
not cherry-picked. The two demo tiles are the exception: one from each
|
| 62 |
-
validation slide, both in the top few percent of the split by patch PCC
|
| 63 |
-
(0.62 and 0.53 against a split median of ~0.32), chosen to make the demo
|
| 64 |
-
legible. Treat the demo as a best case; score the val or test split here for
|
| 65 |
-
representative numbers.
|
| 66 |
-
|
| 67 |
-
**Positivity tables.** The full per-slide `csv_nuclei_pos/*.csv` run to
|
| 68 |
-
300 MB each. Here each is cut down to the nuclei that appear in the selected
|
| 69 |
-
tiles' masks (every mask label was found), 1-3 MB per slide, which is all the
|
| 70 |
-
stage-1 loss weighting needs.
|
| 71 |
-
|
| 72 |
-
Extract it and point the code at it:
|
| 73 |
-
|
| 74 |
-
```bash
|
| 75 |
-
tar xf examples/he2mif/ORIONCRC_dataset_tile_20x.tar -C examples/he2mif
|
| 76 |
-
export ORION_DATA_ROOT=$PWD/examples/he2mif/ORIONCRC_dataset_tile_20x
|
| 77 |
-
```
|
| 78 |
|
| 79 |
-
|
| 80 |
-
**MIT license**.
|
| 81 |
-
|
| 82 |
-
> G. Balezo et al. "MIPHEI-ViT: Multiplex Immunofluorescence Prediction from
|
| 83 |
-
> H&E Images using ViT Foundation Models," 2025.
|
| 84 |
-
|
| 85 |
-
## he2st/
|
| 86 |
-
|
| 87 |
-
Two Xenium breast cancer slides from [HEST-1k](https://github.com/mahmoodlab/HEST),
|
| 88 |
-
processed for H&E → single-cell spatial transcriptomics with MuPaD-512 (`6-HE2ST`).
|
| 89 |
-
|
| 90 |
-
| slide | split | subject | tiles | cells |
|
| 91 |
-
|---|---|---|---|---|
|
| 92 |
-
| NCBI917 | train | B7 | 945 | 52,688 |
|
| 93 |
-
| NCBI920 | val | B5 | 2,527 | 187,016 |
|
| 94 |
-
|
| 95 |
-
Both come from *Single-Cell Spatial Transcriptomics of Archival Human Lung and
|
| 96 |
-
Breast Tumor FFPE Samples* (HEST v1.3.1). NCBI920 is a validation slide: the
|
| 97 |
-
model never trains on it, it only selects the checkpoint.
|
| 98 |
-
|
| 99 |
-
```
|
| 100 |
-
cache/<slide>/ the tile cache (scripts/preprocess_slides.py)
|
| 101 |
-
he.npy (n_tiles, 512, 512, 3) uint8 H&E, 100 µm tiles
|
| 102 |
-
expr.npy (n_cells, 90) float16 log1p expression, tile by tile
|
| 103 |
-
contours.npy tile-local cell contours at 512 px
|
| 104 |
-
contour_offsets.npy start of each cell's contour
|
| 105 |
-
cell_seg.npy each row's index into the xenium_seg polygons
|
| 106 |
-
tiles.parquet per tile: origin (full-res px), cell_start, n_cells
|
| 107 |
-
meta.json pixel size, subject, genes, ...
|
| 108 |
-
features/musk/<slide>/
|
| 109 |
-
emb_cell.npy (n_cells, 1024) float16 MUSK feature of every cell (he2st.embed_musk)
|
| 110 |
-
HEST1k/ the raw HEST-1k files these were built from
|
| 111 |
-
wsis/ st/ xenium_seg/ cells/ metadata/ HEST_v1_3_1.csv (these two rows)
|
| 112 |
-
splits_example.json train = NCBI917, val = test = NCBI920, for a smoke run
|
| 113 |
-
```
|
| 114 |
-
|
| 115 |
-
Rows of `expr.npy`, `emb_cell.npy` and every per-cell prediction are the same
|
| 116 |
-
cells in the same order. Tile `t` holds rows `[cell_start, cell_start + n_cells)`.
|
| 117 |
-
|
| 118 |
-
Point the code at it:
|
| 119 |
-
|
| 120 |
-
```bash
|
| 121 |
-
export HEST_ROOT=$PWD/examples/he2st/HEST1k HE2ST_CACHE=$PWD/examples/he2st/cache \
|
| 122 |
-
HE2ST_MUSK_FEATURES=$PWD/examples/he2st/features/musk
|
| 123 |
-
```
|
| 124 |
|
| 125 |
-
|
| 126 |
-
`python scripts/preprocess_slides.py --slides NCBI917,NCBI920` (into an empty
|
| 127 |
-
`HE2ST_CACHE`), which reproduces `cache/` exactly. To score a model on the
|
| 128 |
-
held-out slide, pass `--slides NCBI920` to `scripts/export_mupad_cells.py` and
|
| 129 |
-
`scripts/score.py`. To smoke-test training, set
|
| 130 |
-
`HE2ST_SPLITS=$PWD/examples/he2st/splits_example.json`.
|
| 131 |
|
| 132 |
-
|
| 133 |
-
Transcriptomics and Histology Image Analysis", NeurIPS 2024), redistributed
|
| 134 |
-
under its **CC BY-NC-SA 4.0** license: non-commercial use, with attribution,
|
| 135 |
-
and derivatives under the same license.
|
|
|
|
| 1 |
---
|
| 2 |
license: other
|
| 3 |
license_name: mixed-per-folder
|
| 4 |
+
license_details: he2mif/ is MIT (MIPHEI-ViT Orion-CRC release); he2st/ is CC BY-NC-SA 4.0 (HEST-1k); sdk/ redistributes two H&E slides under their original releases' terms (ORION-CRC, HEST-1k)
|
| 5 |
pretty_name: MuPaD examples
|
| 6 |
tags:
|
| 7 |
- pathology
|
|
|
|
| 17 |
|
| 18 |
# MuPaD examples
|
| 19 |
|
| 20 |
+
Example data for [MuPaD](https://github.com/jinxixiang/MUPAD) (paper code) and the [MuPaD SDK](https://github.com/jinxixiang/mupad-sdk) (whole-slide inference).
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
+
| folder | contents | size | licence |
|
| 23 |
+
|---|---|---|---|
|
| 24 |
+
| `he2mif/` | ORION-CRC tiles for `5-HE2mIF`: 840 H&E / 17-channel mIF / nuclei tiles in the Zenodo [15340874](https://zenodo.org/records/15340874) layout, as one tar | 0.4 GB | MIT |
|
| 25 |
+
| `he2st/` | two HEST-1k Xenium breast slides for `6-HE2ST` (NCBI917 train, NCBI920 val): raw HEST files, the 100 µm tile cache and per-cell MUSK features | 12.4 GB | CC BY-NC-SA 4.0 |
|
| 26 |
+
| `sdk/` | the two held-out H&E whole-slide images the SDK's end-to-end checks run on: ORION-CRC CRC11 (0.325 µm/px) and HEST-1k NCBI920 (0.087 µm/px); H&E only, see `sdk/README.md` | 6.9 GB | original terms |
|
|
|
|
| 27 |
|
| 28 |
```bash
|
| 29 |
huggingface-cli download xiangjx/MuPaD-examples --repo-type dataset --include "he2mif/*" --local-dir examples
|
| 30 |
+
tar xf examples/he2mif/ORIONCRC_dataset_tile_20x.tar -C examples/he2mif # tiles ship as one tar
|
| 31 |
huggingface-cli download xiangjx/MuPaD-examples --repo-type dataset --include "he2st/*" --local-dir examples
|
| 32 |
```
|
| 33 |
|
| 34 |
+
**he2mif.** Tiles with tissue were sampled uniformly per slide: 600 train, 120 val, 120 test, plus the two demo tiles of `5-HE2mIF/demo.ipynb` (a best case, not typical). Point the code at it with `ORION_DATA_ROOT=$PWD/examples/he2mif/ORIONCRC_dataset_tile_20x`.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
|
| 36 |
+
**he2st.** `cache/<slide>/` is the tile cache, `features/musk/<slide>/emb_cell.npy` the per-cell MUSK features, `HEST1k/` the raw inputs they were built from (`wsis/ st/ xenium_seg/ cells/ metadata/`). Rows of `expr.npy`, `emb_cell.npy` and every prediction are the same cells in the same order. Set `HEST_ROOT=$PWD/examples/he2st/HEST1k HE2ST_CACHE=$PWD/examples/he2st/cache HE2ST_MUSK_FEATURES=$PWD/examples/he2st/features/musk`; `splits_example.json` makes a one-slide smoke run.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
+
**sdk.** `python scripts/e2e_mif_orion.py` and `python scripts/e2e_expression_ncbi920.py` in the SDK download these slides themselves; the ground truth (Orion mIF, Xenium counts) comes from the original ORION-CRC and HEST-1k releases.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
+
Sources: ORION-CRC tiles from the MIPHEI-ViT release (Balezo et al., 2025; MIT); HEST-1k (Jaume et al., NeurIPS 2024; CC BY-NC-SA 4.0, non-commercial, attribution, share-alike).
|
|
|
|
|
|
|
|
|