You can deploy this model and start earning money today!

ahmedayman4a/
acrobat-registered-wsis-flat

Name: acrobat-registered-wsis-flat
Availability: InStock

Views

No views yet

ACROBAT Registered Whole-Slide Images (Flat JPEG)

Flat JPEG-compressed registered WSIs from the ACROBAT breast cancer dataset. All immunohistochemistry (IHC) slides — ER, PGR, HER2, KI67 — were warped into H&E spatial alignment using VALIS (2nd place overall in the ACROBAT 2022 challenge).

Dataset Summary

196 patients from the ACROBAT training set (750 total in official split)
890 TIFFs: 148 H&E + 742 IHC (ER, PGR, HER2, KI67)
Format: Flat tiled TIFF with JPEG compression (no pyramid levels, single resolution)
Resolution: 0.92 µm/px (native 10X)
Compression: JPEG Q=90 (DEFLATE sources) or Q=95 (JPEG2000 sources)

Registration Pipeline

Two-stage VALIS registration using the Gatenbee et al. recommended approach:

Global registration: ColorfulStandardizer + MicroRigidRegistrar
Micro-registration: at 8% of reference (H&E) slide max dimension, rigid refinement
All slides cropped to the H&E reference extent

File Structure

train/
  {anon_id}_{stain}_train.tif
  ...
df_acrobat_meta_registered.csv   ← filtered metadata (same schema as original ACROBAT)

Metadata Schema

The CSV (df_acrobat_meta_registered.csv) follows the original ACROBAT format:

Column	Description
`anon_id`	Anonymous patient ID (numeric string)
`stain`	HE, ER, PGR, HER2, or KI67
`filename`	WSI filename
`mpp_lvl_0..8`	Microns per pixel at each pyramid level
`magnification_lvl_0..8`	Magnification at each level
`vendor`	Scanner vendor (hamamatsu)
`model`	Scanner model (C12000-02, C12000-22, or C13220)
`set`	Dataset split (train)

Loading Code

python

import openslide

slide = openslide.OpenSlide("train/0_HE_train.tif")
img = slide.read_region((0, 0), 0, slide.dimensions)

Known Limitations

Flat TIFFs only — no pyramid levels. For pyramidal OME-TIFFs, use the companion dataset.
Train split only — validation (100 patients) and test (303 patients) are not yet registered.
JPEG2000 source files went through an additional lossy→lossy conversion (JPEG2000 Q=90 → JPEG Q=95). Visual quality is preserved (per-pixel RMSE < 2 gray levels), but this adds ~1 generation of compression.
VALIS TRE estimates are included in per-patient registration_qc.json files (not in this flat dataset — see the registration outputs directory).

Citation

If you use this dataset, please cite both:

The ACROBAT challenge: Weitz et al. "ACROBAT — A multi-stain breast cancer histological whole-slide-image data set." 2022.
VALIS: Gatenbee et al. "Virtual alignment of pathology image series." 2023.

ahmedayman4a/acrobat-registered-wsis-flat