README.md

---
license: cc-by-nc-nd-4.0
language:
- en
pretty_name: TriPAH Medical Retrieval Resources
tags:
- medical
- cross-modal-retrieval
- image-text
- hashing
---

# TriPAH Medical Retrieval Resources

**TriPAH: Imbalance-Aware Tri-Prompt Affinity Hashing for Cross-Modal Medical Retrieval**

**Accepted at ICME 2026 as an Oral presentation.**

[Paper](https://arxiv.org/abs/2606.27010) · [Code](https://github.com/Medical-Multi-Agent-AI/TriPAH) · [Project page](https://medical-multi-agent-ai.github.io/TriPAH/)

This repository documents the three medical datasets used by the TriPAH implementation: ODIR-5K, IU-Xray, and MIMIC-CXR. This revision contains dataset documentation and preparation instructions only. It does **not** contain the underlying images, clinical reports, row-level labels, patient identifiers, or prompt embeddings.

## Dataset availability

| Dataset | Current prepared rows | Labels | Public distribution in this revision |
| --- | ---: | ---: | --- |
| [ODIR-5K](odir/README.md) | 3,500 | 8 | Source and preparation instructions |
| [IU-Xray](iu-xray/README.md) | 3,820 | 14 | Source and preparation instructions; Open-i asks users not to share outside their research group/organization |
| [MIMIC-CXR](mimic-cxr/README.md) | 216,320 | 14 | Credentialed source and preparation instructions; no restricted records are redistributed |

The row counts describe the current local preprocessing artifacts. They are not a replacement for the manuscript's dataset statistics or a claim that the published results were rerun with this release.

## Expected training layout

After obtaining source data through the appropriate provider and running the project converters, each dataset is expected at `TriPAH/dataset/<name>/`:

```text
dataset/<name>/
├── images/
├── caption.txt
├── index.mat
├── label.mat
├── prompt_caption.npz
└── dataset_info.json
```

The image index, captions, labels, and prompts must have matching row order. The inspected main prompt arrays have shape `[N, 77, 512]` and dtype `float32`. The current code performs a seeded sample-wise split; users must match their training and evaluation splits and should not infer patient-disjoint separation from this layout alone.

See the [repository data preparation guide](https://github.com/Medical-Multi-Agent-AI/TriPAH#prepare-the-data) and each dataset folder for source-specific instructions.

## Access and licensing

The repository's pre-existing license metadata is retained. It does not grant a new license to any upstream dataset. Dataset access, redistribution, and reuse remain governed by the original providers. Source code is licensed separately in the GitHub repository.

- ODIR-5K: obtain the data and applicable terms from the [official ODIR challenge](https://odir2019.grand-challenge.org/dataset/).
- IU-Xray: follow the [Open-i collection instructions](https://openi.nlm.nih.gov/faq#collection), including its request not to share the collection outside the research group/organization.
- MIMIC-CXR: each researcher must obtain [PhysioNet credentialed access and sign the data use agreement](https://physionet.org/content/mimic-cxr/2.1.0/). The agreement restricts sharing access to the data, and [PhysioNet's guidance on derived resources](https://physionet.org/news/post/mimic-derived-datasets-models/) directs authors to share them through PhysioNet under the source agreement. Preprocessing alone does not establish permission for unrestricted redistribution.

## Citation

```bibtex
@article{bian2026tripah,
  title={TriPAH: Imbalance-Aware Tri-Prompt Affinity Hashing for Cross-Modal Medical Retrieval},
  author={Bian, Jiaming and Li, Songming and Song, Yurui and Chen, Yunfei and Cao, Yichao and Long, Jun},
  journal={arXiv preprint arXiv:2606.27010},
  year={2026}
}
```

Please also cite the original dataset publications when using their data.