The dataset is licensed under CC BY 4.0. Please attribute the authors and dataset, link to the license, and indicate changes when redistributing adapted data.

AutoPath liver 10× patches · AutolysisBench

Archived A2F_10x liver training/test patches from the AutoPath project, associated with Through the Schrödinger Bridge: Benchmarking Antemortem Image Restoration from Postmortem Autolysis to Enhance Forensic Diagnostics (MICCAI 2026).

Shuang Hao, Jiacheng Yue, Yaxuan Zhao, Fan Wang, Jianhua Ma, Erwen Huang, and Chunfeng Lian.

Paper · Hugging Face Paper page · Training and inference code · Model weights · Download guide

Scope and counts

This release provides the original training and test PNGs for the study's unpaired liver autolysis restoration model. Domain A contains autolyzed tissue; domain B contains non-autolyzed tissue. The original split membership is unchanged.

Split	Domain	Patches	Cases
`trainA`	Autolyzed	4,962	79
`trainB`	Non-autolyzed	4,964	79
`testA`	Autolyzed	500	78
`testB`	Non-autolyzed	500	79

Total: 10,926 patches from 79 cases. Patches are 1024 × 1024 RGB PNGs extracted at the study's 10× scale. Files are packaged as WebDataset TAR shards containing matching PNG and JSON members. The PNG bytes are preserved; no additional compression or resizing is applied to the images.

Access and download

Sign in to Hugging Face, submit the access request on this dataset page, and access is approved automatically. Then authenticate the downloading machine with hf auth login.

Download as trainA / trainB / testA / testB folders

python -m pip install -r requirements-data.txt
python download_data.py --repo-id fengluo/AutolysisBench --output data/liver10x

Run these commands from the code repository. The downloader verifies each archive and PNG with SHA-256, preserves the original PNG bytes, and reconstructs trainA/, trainB/, testA/, and testB/. The output directory must not already exist. Use --splits testA testB to download only the test data. --revision accepts a dataset tag or commit; each download resolves one revision and records it in download_record.json.

Streaming access

from datasets import load_dataset

train_a = load_dataset("fengluo/AutolysisBench", split="trainA", streaming=True)
sample = next(iter(train_a))
image = sample["png"]
metadata = sample["json"]

Metadata and checksums

Each JSON sample record includes sample_id, case_id, domain, split, width, height, and the original PNG's sha256. Case IDs and filenames have been pseudonymized consistently across all four splits. Original accession-style names and internal paths are excluded. case_id supports grouping and overlap checks; it does not provide registration or a paired target.

release_manifest.jsonl lists all samples, public shard paths, PNG byte counts, and SHA-256 checksums. release_summary.json contains the release totals and per-shard checksums. SHA256SUMS covers the published files.

All 10,926 PNGs were fully decoded and checked as 1024 × 1024 RGB images. The original PNG payloads total 16,848,298,432 bytes (16.85 GB). Their bytes are preserved in the archives. SHA-256 checking found 0 duplicate-content groups, including 0 groups spanning training and test data.

Split limitations and study accounting

The original patch-level membership is preserved. There are 79 distinct cases; 79 occur in both the combined training and combined test splits. The splits have no overlapping source file paths, but they are not case-independent. These historical splits must not be described as a held-out-case benchmark. Use case_id to construct case-disjoint partitions for new generalization experiments. Autolyzed and non-autolyzed patches are unpaired: a shared case ID does not establish pixel-level or patch-level registration.

Counts in this release follow the preserved training/test manifests. trainB includes 4,964 images, including the two-image difference from the 4,962-per-domain summary used in the historical poster. The release's 79-case accounting may differ from a differently defined paper/poster cohort; it is reported explicitly here. No images were removed or reassigned to match those summary numbers.

Only the liver 10× training/test patches are released here. This repository does not include whole-slide files, other organs, generated restorations, paired pixel-level targets, or diagnostic annotations. Model-generated restorations should not be treated as observed ground-truth antemortem tissue.

Model preprocessing

Training uses 286 × 286 resizing followed by random 256 × 256 cropping. Released model inference bicubic-resizes to 256 × 256. Downloading this dataset preserves the original 1024 × 1024 images.

License and attribution

Dataset: Creative Commons Attribution 4.0 International; the complete terms are in LICENSE. Attribute the AutolysisBench authors, link to the dataset and license, and indicate modifications when redistributing adapted data. The code and released model weights are separately licensed under MIT in the code repository.

Citation

@inproceedings{autolysis_schrodinger_bridge_2026,
  title     = {Through the Schrödinger Bridge: Benchmarking Antemortem Image Restoration from Postmortem Autolysis to Enhance Forensic Diagnostics},
  author    = {Hao, Shuang and Yue, Jiacheng and Zhao, Yaxuan and Wang, Fan and Ma, Jianhua and Huang, Erwen and Lian, Chunfeng},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year      = {2026},
  note      = {Early Accept}
}

Downloads last month: 17

Total file size:

16.9 GB

Paper for ShuangHaoXJTU/AutolysisBench

Through the Schrödinger Bridge: Benchmarking Antemortem Image Restoration from Postmortem Autolysis to Enhance Forensic Diagnostics

Paper • 2608.21813 • Published Aug 22