ICME 2026 · Oral

TriPAH

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

Jiaming Bian · Songming Li · Yurui Song · Yunfei Chen^* · Yichao Cao^* · Jun Long^*

Big Data Institute, Central South University
* Co-corresponding authors

Compact binary codes for retrieving medical images and clinical reports, learned from complementary image, text, and patient-context prompt views.

Method

Patient-context prompts complement noisy reports. Mamba–Transformer blocks fuse the three views, while class-balanced learning and progressive quantization preserve disease information in binary codes.

TriPAH framework: context-adaptive prompting, CLIP feature extraction, Mamba–Transformer tri-view fusion, and imbalance-aware multi-task hashing. — TriPAH framework · Scroll to exploreOpen full-size figure ↗

Results

Image-to-text mAP, averaged over 32, 64, and 128 bits. Results from Table I of the paper.

ODIR-5K

0.937

MIMIC-CXR

0.820

IU-Xray

0.896

Full benchmark results

BestSecond bestmAP ↑

ODIR-5K · Mean average precision (mAP ↑)
Method	Image → Text			Text → Image
Method	32 bits	64 bits	128 bits	32 bits	64 bits	128 bits
DCHMT	0.502	0.518	0.522	0.707	0.711	0.718
MITH	0.501	0.506	0.524	0.432	0.443	0.436
DSPH	0.449	0.487	0.488	0.351	0.385	0.401
CMCL	0.531	0.527	0.542	0.390	0.393	0.407
CICH	0.542	0.528	0.525	0.813	0.819	0.931
MSACH	0.665	0.669	0.690	0.888	0.889	0.901
TriPAH Ours	0.936	0.939	0.937	0.989	0.971	0.972

Swipe or scroll the table to see all bit lengths.

ODIR-5K: binocular fundus images with 8 disease labels.

Table I of the paper. Best and second-best values are marked within each column; ties share the same rank. A dash indicates an unreported result.

Inside the experiments

Qualitative ODIR-5K example from Figure 3: retrieval between binocular fundus images and diagnostic descriptions. — Retrieval between fundus images and diagnostic descriptions. ODIR-5K example from Figure 3. Scroll the figure sideways to read it.

1 / 4

Data

Hugging Face ↗

Dataset cards and preparation guides for the three medical benchmarks.

ODIR-5K Binocular fundus photography MIMIC-CXR Chest radiographs and reports IU-Xray Chest X-rays and reports

Citation / BibTeX

Download 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}
}

TriPAH

Method

Results

ODIR-5K

MIMIC-CXR

IU-Xray

Inside the experiments

ODIR-5K retrieval

IU-Xray retrieval

Bidirectional retrieval

Component ablation

Data