README.md

---
license: other
tags:
  - medical
  - benchmark
  - multimodal
  - agentic
---

# AutoMedBench-Full-release

[![Website](https://img.shields.io/badge/Website-AutoMedBench-6f8f72.svg)](https://automedbench.github.io/) [![Lite Leaderboard](https://img.shields.io/badge/Lite%20Leaderboard-Open%20Space-4A7355.svg)](https://huggingface.co/spaces/MitakaKuma/AutoMedBench-Leaderboard) [![Full Leaderboard](https://img.shields.io/badge/Full%20Leaderboard-Open%20Space-4A7355.svg)](https://huggingface.co/spaces/MitakaKuma/AutoMedBench-Full-Leaderboard) [![arXiv](https://img.shields.io/badge/arXiv-2606.01961-b31b1b.svg)](https://arxiv.org/abs/2606.01961) [![GitHub](https://img.shields.io/badge/GitHub-AutoMedBench-181717?logo=github)](https://github.com/AutoMedBench/AutoMedBench)

[中文说明](README_zh.md)

AutoMedBench is a workflow-aware benchmark for autonomous medical-AI research
agents. It evaluates both the final task artifact and the S1-S5 research
workflow: Plan, Setup, Validate, Inference, and Submit.

This Full release packages **48 tasks** across **7 tracks** with Lite and
Standard tiers, for **96 task-tier combinations**. Each combination has a
pre-built Docker image. Dataset bytes are **not** bundled — users prepare data
independently and mount it at runtime.

## How to Submit

**Submission email: <jliu718@ucsc.edu>**

Suggested subject: `AutoMedBench submission - <agent/team name>`

| Priority | What to do | Why it matters |
| --- | --- | --- |
| **Required** | **Run your agent at least 5 repeated times on the benchmark.** Agent runs can vary because of model sampling, tool-use choices, network timing, and dependency setup. Repeated runs provide a more stable estimate of real performance than a single run. |
| **Required** | **Zip all output reports and conversation/transcript files into one archive, then email that zip file to <jliu718@ucsc.edu>.** | The full package lets maintainers verify final artifacts, audit the workflow, and check stability across repeated runs. |
| **Recommended** | **Use a coding agent to launch, monitor, and collect runs.** | Benchmark runs are long and may need log inspection, quick fixes, and clean collection of conversation files. |

## Quick Start

```bash
# 1. Download a track Docker image (e.g. segmentation)
#    From this repository: docker/segmentation/automedbench-segmentation-docker-image-v0.1.0.tar.gz

# 2. Load the Docker image
docker load -i automedbench-segmentation-docker-image-v0.1.0.tar.gz

# 3. Set up credentials
cp env.example .env
# edit .env with your API key

# 4. Run a task
python docker/orchestrator.py \
    --agent claude-opus-4-6 \
    --task kidney-seg-task \
    --tier lite \
    --n-patients 20
```

The Docker image contains the benchmark harness and scoring code. Credentials
are passed only at runtime through `.env` or shell environment variables.
Dataset files must be mounted separately (see [Data Acquisition](#data-acquisition)
below).

## Docker Image Packages

Prebuilt per-track Docker image packages are stored in this repository:

```text
docker/classification/automedbench-classification-docker-image-v0.1.0.tar.gz
docker/detection/automedbench-detection-docker-image-v0.1.0.tar.gz
docker/enhancement/automedbench-enhancement-docker-image-v0.1.0.tar.gz
docker/report/automedbench-report-docker-image-v0.1.0.tar.gz
docker/segmentation/automedbench-segmentation-docker-image-v0.1.0.tar.gz
docker/synthesis/automedbench-synthesis-docker-image-v0.1.0.tar.gz
docker/vqa/automedbench-vqa-docker-image-v0.1.0.tar.gz
```

Each archive is ~3.5 GB and contains all Lite + Standard task-tier images for
that track. The shared runtime archive
(`runtime/automedbench-full-runtime-0.1.0.tar.gz`, ~4 GB) supplies the common
agent runtime and seven track evaluator images.

After downloading, load all images:

```bash
docker load -i runtime/automedbench-full-runtime-0.1.0.tar.gz
docker load -i docker/segmentation/automedbench-segmentation-docker-image-v0.1.0.tar.gz
# ... repeat for other tracks as needed
```

## Released Tasks (48 tasks, 96 Lite/Standard combinations)

| Track | Tasks | Metric focus |
| --- | ---: | --- |
| Segmentation | 14 | Dice (organ + lesion) |
| Enhancement | 7 | SSIM |
| VQA | 9 | Answer accuracy |
| Report generation | 5 | Clinical report quality |
| Detection | 4 | mAP@0.5 |
| Classification | 5 | Label accuracy |
| Synthesis | 4 | SSIM |

## Data Acquisition

**Dataset bytes are not included in this release.** Each task references an
upstream dataset that must be downloaded independently. The table below lists
every task's data source, download link, license, and access requirements.

### Segmentation (14 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `kidney-seg-task` | KiTS19 | [kits19.grand-challenge.org](https://kits19.grand-challenge.org/) | CC BY-NC-SA 4.0 | Public | ~200 MB/case | Acquisition ready |
| `pancreas-seg-task` | PanTS | [github.com/MrGiovanni/PanTS](https://github.com/MrGiovanni/PanTS) | CC BY-NC-ND 4.0 | Public | ~300–346 GB | Acquisition ready |
| `pancreas-oar-seg-task` | PanTS | [github.com/MrGiovanni/PanTS](https://github.com/MrGiovanni/PanTS) | CC BY-NC-ND 4.0 | Public | ~300–346 GB | Acquisition ready |
| `liver-seg-task` | MSD Task03 Liver | [medicaldecathlon.com](https://medicaldecathlon.com/) | Provenance conflict | Public | ~28.9 GB | Staged verified |
| `aeropath-seg-task` | AeroPath | [zenodo.org/records/10069289](https://zenodo.org/records/10069289) | CC BY 4.0 | Public | ~5.04 GB | Acquisition ready |
| `tsg-multiorgan-seg-task` | TotalSegmentator v2.0.1 | [zenodo.org/records/10047263](https://zenodo.org/records/10047263) | CC BY 4.0 | Public | ~3.24 GB | Staged verified |
| `colon-seg-task` | MSD Task10 Colon | [medicaldecathlon.com](https://medicaldecathlon.com/) | CC BY-SA 4.0 | Public | ~6.24 GB | Staged verified |
| `hepaticvessel-seg-task` | MSD Task08 HepaticVessel | [medicaldecathlon.com](https://medicaldecathlon.com/) | CC BY-SA 4.0 | Public | ~9.35 GB | Staged verified |
| `spleen-seg-task` | MSD Task09 Spleen | [medicaldecathlon.com](https://medicaldecathlon.com/) | CC BY-SA 4.0 | Public | ~1.61 GB | Staged verified |
| `heart-seg-task` | MSD Task02 Heart | [medicaldecathlon.com](https://medicaldecathlon.com/) | CC BY-SA 4.0 | Public | ~456 MB | Staged verified |
| `prostate-seg-task` | MSD Task05 Prostate | [medicaldecathlon.com](https://medicaldecathlon.com/) | CC BY-SA 4.0 | Public | ~240 MB | Staged verified |
| `feta-seg-task` | FeTA | [zenodo.org/records/4541606](https://zenodo.org/records/4541606) | Research agreement | Gated | — | Externally blocked |
| `panther-t1-seg-task` | PANTHER | [zenodo.org/records/15192302](https://zenodo.org/records/15192302) | CC BY-NC 4.0 + access controls | Restricted | — | Externally blocked |
| `panther-t2-seg-task` | PANTHER | [zenodo.org/records/15192302](https://zenodo.org/records/15192302) | CC BY-NC 4.0 + access controls | Restricted | — | Externally blocked |

### Enhancement (7 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `ldct-denoising-task` | AAPM Low Dose CT Grand Challenge | [aapm.org/grandchallenge/lowdosect](https://www.aapm.org/grandchallenge/lowdosect/) | No explicit redistribution grant | Public library | ~42 MB staged | Download required |
| `lidc-idri-denoising-task` | LIDC-IDRI | [cancerimagingarchive.net/collection/lidc-idri](https://www.cancerimagingarchive.net/collection/lidc-idri/) | CC BY 3.0 | Public | ~136.55 GB | Acquisition ready |
| `deeplesion-denoising-task` | NIH DeepLesion | [nihcc.app.box.com/v/DeepLesion](https://nihcc.app.box.com/v/DeepLesion) | No explicit license | Public | Unknown | Acquisition ready |
| `mri-sr-task` | fastMRI | [fastmri.med.nyu.edu](https://fastmri.med.nyu.edu/) | Agreement prohibits redistribution | Application required | — | Externally blocked |
| `brats-t1c-sr-task` | BraTS 2023 GLI | [synapse.org/Synapse:syn51156910](https://www.synapse.org/Synapse:syn51156910/wiki/) | Controlled access | Controlled | — | Externally blocked |
| `ixi-t1-sr-task` | IXI T1 | [brain-development.org/ixi-dataset](https://brain-development.org/ixi-dataset/) | CC BY-SA 3.0 | Public | Not specified | Acquisition ready |
| `nih-cxr-sr-task` | NIH ChestXray14 | [nihcc.app.box.com/v/ChestXray-NIHCC](https://nihcc.app.box.com/v/ChestXray-NIHCC) | CC0 (NIH Kaggle) | Public | ~45.08 GB | Acquisition ready |

### VQA (9 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `pathvqa-task` | PathVQA | [huggingface.co/datasets/flaviagiammarino/path-vqa](https://huggingface.co/datasets/flaviagiammarino/path-vqa) | MIT | Public | ~785 MB | Staged verified |
| `vqa-rad-task` | VQA-RAD | [huggingface.co/datasets/flaviagiammarino/vqa-rad](https://huggingface.co/datasets/flaviagiammarino/vqa-rad) | CC0 1.0 | Public | ~34.5 MB | Staged verified |
| `medframeqa-task` | MedFrameQA | [huggingface.co/datasets/SuhaoYu1020/MedFrameQA](https://huggingface.co/datasets/SuhaoYu1020/MedFrameQA) | CC BY 4.0 | Public | ~571 MB | Staged verified |
| `slake-task` | SLAKE | [huggingface.co/datasets/BoKelvin/SLAKE](https://huggingface.co/datasets/BoKelvin/SLAKE) | CC BY 4.0 | Public | ~217 MB | Staged verified |
| `medxpertqa-mm-task` | MedXpertQA | [huggingface.co/datasets/TsinghuaC3I/MedXpertQA](https://huggingface.co/datasets/TsinghuaC3I/MedXpertQA) | MIT | Public | ~523 MB | Staged verified |
| `vqa-kvasir-task` | Kvasir-VQA | [huggingface.co/datasets/SimulaMet-HOST/Kvasir-VQA](https://huggingface.co/datasets/SimulaMet-HOST/Kvasir-VQA) | CC BY-NC 4.0 + benchmark permission | Permission required | ~15.18 GB | Externally blocked |
| `vqa-omnimedvqa-task` | OmniMedVQA | [huggingface.co/datasets/foreverbeliever/OmniMedVQA](https://huggingface.co/datasets/foreverbeliever/OmniMedVQA) | Mixed (no global license) | Mixed access | ~10.70 GB | Externally blocked |
| `vqa-pmc-vqa-task` | PMC-VQA | [huggingface.co/datasets/RadGenome/PMC-VQA](https://huggingface.co/datasets/RadGenome/PMC-VQA) | CC BY-SA | Public | ~21.77 GB | Staged verified |
| `vqa-mmmu-medical-task` | MMMU medical subsets | [huggingface.co/datasets/MMMU/MMMU](https://huggingface.co/datasets/MMMU/MMMU) | Apache-2.0 | Public | ~2.09 GB | Staged verified |

### Report Generation (5 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `mimic-cxr-report-task` | MIMIC-CXR | [physionet.org/content/mimic-cxr/2.1.0](https://physionet.org/content/mimic-cxr/2.1.0/) | DUA prohibits sharing | Credentialed DUA | 377,110 images | Externally blocked |
| `iu-xray-report-task` | IU Open-i | [openi.nlm.nih.gov](https://openi.nlm.nih.gov/) | No clear bulk redistribution license | Public search | Unknown | Acquisition ready |
| `chexpert-plus-cxr-task` | CheXpert Plus | [aimi.stanford.edu/datasets/chexpert-plus](https://aimi.stanford.edu/datasets/chexpert-plus) | No license declared by mirror | Public mirror | ~636 MB | Staged verified |
| `pathology-caption-100-task` | PathCap | [huggingface.co/datasets/jamessyx/PathCap](https://huggingface.co/datasets/jamessyx/PathCap) | CC BY-NC 2.0 + click-through | Auto-gated | ~13.7 GB | Staged verified |
| `pathology-caption-500-task` | PathCap | [huggingface.co/datasets/jamessyx/PathCap](https://huggingface.co/datasets/jamessyx/PathCap) | CC BY-NC 2.0 + click-through | Auto-gated | ~13.7 GB | Staged verified |

### Detection (4 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `vindr-cxr-det-task` | VinDr-CXR | [physionet.org/content/vindr-cxr/1.0.0](https://physionet.org/content/vindr-cxr/1.0.0/) | PhysioNet Credentialed Health Data License | Credentialed DUA | — | Externally blocked |
| `bccd-det-task` | BCCD | [github.com/Shenggan/BCCD_Dataset](https://github.com/Shenggan/BCCD_Dataset) | MIT | Public | ~4.95 MB | Staged verified |
| `dentex-det-task` | DENTEX | [huggingface.co/datasets/ibrahimhamamci/DENTEX](https://huggingface.co/datasets/ibrahimhamamci/DENTEX) | CC BY-NC-SA 4.0 | Public | ~11.84 GB | Staged verified |
| `grazpedwri-det-task` | GRAZPEDWRI-DX | [figshare.com/articles/dataset/GRAZPEDWRI-DX/14825193](https://figshare.com/articles/dataset/GRAZPEDWRI-DX/14825193) | CC BY 4.0 | Public | ~16.26 GB | Staged verified |

### Classification (5 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `braintumor-cls-task` | Brain Tumor MRI | [kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset](https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset) | CC BY 4.0 | Public | ~167.85 MB | Acquisition ready |
| `crc-histology-cls-task` | NCT-CRC-HE-100K | [zenodo.org/records/1214456](https://zenodo.org/records/1214456) | CC BY 4.0 | Public | ~11.7 GB | Acquisition ready |
| `patchcamelyon-cls-task` | PatchCamelyon | [github.com/basveeling/pcam](https://github.com/basveeling/pcam) | CC0 | Public | ~7.7 GB | Acquisition ready |
| `chest-xray-pneumonia-cls-task` | Kermany chest X-ray | [data.mendeley.com/datasets/rscbjbr9sj/2](https://data.mendeley.com/datasets/rscbjbr9sj/2) | CC BY 4.0 | Public | Not specified | Acquisition ready |
| `skin-lesion-cls-task` | ISIC Archive | [isic-archive.com](https://www.isic-archive.com/) | Per-image (not frozen) | Public | ~6.26 MB staged | Staged verified |

### Synthesis (4 tasks)

| Task ID | Upstream Dataset | Download Link | License | Access | Approx. Size | Status |
| --- | --- | --- | --- | --- | --- | --- |
| `synthrad2025-mrct-task` | SynthRAD2025 | [zenodo.org/records/15373853](https://zenodo.org/records/15373853) | CC BY-NC 4.0 | Public | ~14.8 GB | Acquisition ready |
| `ctorg-ctsr-task` | CT-ORG | [cancerimagingarchive.net/collection/ct-org](https://www.cancerimagingarchive.net/collection/ct-org/) | CC BY 3.0 | Public | Not exposed | Acquisition ready |
| `msd-pancreas-ctsr-task` | MSD Task07 Pancreas | [medicaldecathlon.com](https://medicaldecathlon.com/) | CC BY-SA 4.0 | Public | ~12.3 GB | Staged verified |
| `totalsegmentator-ctsr-task` | TotalSegmentator | [github.com/wasserth/TotalSegmentator](https://github.com/wasserth/TotalSegmentator) | CC BY 4.0 | Public | Not frozen | Acquisition ready |

## Data Acquisition Summary

| Status | Count | Description |
| --- | ---: | --- |
| Staged verified | 23 | Locally staged and layout-hashed; recipe frozen |
| Acquisition ready | 16 | Source and staging recipe frozen; user must download |
| Externally blocked | 9 | Access approval, DUA, or upstream clarification required |

## Dataset Licenses

See [docs/DATA_POLICY.md](docs/DATA_POLICY.md) for the full per-task policy
records. The release rule is conservative: **no dataset bytes are embedded** in
any task package, runtime image, or task-tier image. Every uploaded artifact is
code-only.

## Package Design

Each of the 48 task packages supports both `lite` and `standard` tiers; the
tiers share the same harness and runtime contract. The common container layer
remains eight shared, data-free runtime images:

- one common agent runtime used by every track; and
- seven track-specific evaluator runtimes, one for each track.

The execution release adds 96 thin images on top of that shared layer: one
immutable image for every task-tier cell. Their build IDs and API verification
records are frozen in `manifests/cell_images.csv` and
`manifests/cell_verification.csv`. The shared image inventory remains frozen in
`manifests/runtime_images.csv`.

Following AutoMedBench-Lite, the 96 images are exported under seven track
folders as `docker/<track>/automedbench-<track>-docker-image-v0.1.0.tar.gz`;
the local `synthetic` track uses Lite's public `synthesis` folder name. Each
archive is accompanied by a SHA-256 sidecar and an image inventory. Exact
remote sizes, hashes, and revisions are frozen in
`manifests/cell_archives.csv`.

## Security

No credentials, API keys, model-service tokens, personal home paths, or
populated runtime configuration belong in this repository or its images.
Source worktrees are read-only inputs; builds use fresh allow-listed staging
contexts made from immutable Git blobs.

## Paper

Project page: <https://automedbench.github.io/>

Paper: Junqi Liu et al., "AutoMedBench: Towards Medical AutoResearch with
Agentic AI Models", arXiv:2606.01961, 2026.
<https://arxiv.org/abs/2606.01961>

DOI: <https://doi.org/10.48550/arXiv.2606.01961>

```bibtex
@article{liu2026automedbench,
  title={AutoMedBench: Towards Medical AutoResearch with Agentic AI Models},
  author={Liu, Junqi and Song, Selena and Wang, Yuhan and Mao, Jiawei and Chen, Hardy and Huang, Xiaoke and Qi, Tianhao and Guo, Pengfei and Tang, Yucheng and He, Yufan and Zhao, Can and Myronenko, Andriy and Yang, Dong and Xu, Daguang and Zhou, Yuyin},
  journal={arXiv preprint arXiv:2606.01961},
  year={2026}
}
```