Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
KJD-25K is intended for research use only. Access may require ethics/IRB approval, a signed data use agreement, and compliance with applicable medical data regulations. Please do not attempt to re-identify individuals. This dataset must not be used for clinical diagnosis, emergency decision-making, or direct patient care.
Log in or Sign Up to review the conditions and access this dataset content.
KJD-25K: A 3D Multi-modal Dataset and Benchmark for Knee Joint Diagnosis
Overview | Dataset Details | Benchmark Metrics | Ethics-Approved Access List | Usage Restrictions | Changelog
Overview
KJD-25K is a large-scale, expert-annotated 3D multi-modal knee MRI dataset and benchmark for clinically grounded knee joint diagnosis. It contains approximately 25K MRI volume–report pairs, with 20K training and 5K test examples. The dataset includes expert-curated Chinese radiology reports, structured diagnostic labels, and severity annotations. It covers four standard clinical knee MRI sequences: T1WI, T2WI, T2WIFS, and PDWIFS, acquired on 3.0T scanners with dedicated 8-channel knee coils.
KJD-25K is designed to support training and systematic evaluation of multi-modal large language models (MLLMs) for knee MRI report generation, visual question answering, structured diagnostic reasoning, and clinical consistency assessment. The construction pipeline includes expert-guided case selection, multi-modal data structuring, AI-assisted label generation, and rigorous quality, privacy, and traceability review. All data were de-identified and processed under Institutional Review Board (IRB) approval and data use agreements.
Important: KJD-25K is for research use only. It must not be used for clinical diagnosis, treatment decisions, emergency care, or direct patient management. Model outputs are not a substitute for professional radiological judgment.
Dataset Details
| Item | Description |
|---|---|
| Total samples | Approximately 25K MRI volume–report pairs |
| Training set | 20K |
| Test set | 5K |
| MRI sequences | T1WI, T2WI, T2WIFS, PDWIFS |
| Report language | Chinese |
| Annotation content | Chinese radiology reports, structured diagnostic labels, severity annotations |
| Main tasks | Report generation, visual question answering, classification/structured diagnosis, clinical consistency evaluation |
| Modality | 3D MRI + text |
| Intended use | Research on medical MLLMs and knee MRI understanding |
Construction Summary
- Curated by a 15-person professional team: five board-certified orthopedic radiologists with over ten years of experience, seven trained physician assistants, and three clinically trained engineers.
- Human–AI collaborative pipeline with over 4,000 hours of expert/clinical and computational effort, including approximately 2,000 clinician-hours and 2,000 GPU-hours.
- Four-step pipeline: expert-guided case selection → multi-modal data structuring → AI-assisted label generation → rigorous dataset validation and compliance review.
- All data were de-identified before retention; only MRI volumes and associated clinical reports were retained.
Benchmark Metrics
The paper benchmarks eight representative MLLMs on KJD-25K and establishes baselines. Evaluation considers both linguistic similarity and clinical correctness:
- LS (Lesion Sensitivity): proportion of ground-truth pathological findings correctly identified
- DS (Diagnostic Specificity): proportion of normal anatomical structures correctly recognized as normal, penalizing over-diagnosis
- SRC (Structured Report Compliance): adherence to predefined reporting fields, including examination method, MRI findings, and diagnostic impression
LS, DS, and SRC are clinically oriented LLM-assisted structured metrics scored on a 0–100 scale. The scoring criteria were reviewed and validated by experienced orthopedic radiologists.
Access and Use
KJD-25K is provided for research use only. Access may require applicants to provide a research purpose statement, institutional information, ethics/IRB approval information, and a signed Data Use Agreement (DUA). Please refer to DATA_USE_AGREEMENT.md and LICENSE in this repository. If the dataset is not yet available for public download, please contact the maintainers.
The following applicants/teams have obtained ethics/IRB approval and have been granted access:
| Applicant/Team | Affiliation | Status | Public Note |
|---|---|---|---|
| Li*** | Shenzhen University | Approved | Ethics/IRB approval obtained; DUA signed |
| Ying*** | Zhengzhou University | Approved | Ethics/IRB approval obtained; DUA signed |
| Wang*** | Shenzhen Second People's Hospital | Approved | Ethics/IRB approval obtained; DUA signed |
| Guo*** | Shenzhen Second People's Hospital | Approved | Ethics/IRB approval obtained; DUA signed |
| Yu*** | Zhe Jiang University | Approved | Ethics/IRB approval obtained; DUA signed |
Last updated: 2026-10-23
Citation
@InProceedings{LiYih_KJD25K_MICCAI2026,
author = {Li, Yihui and He, Along and Wang, Yuli and Guo, Kai and Wu, Yanlin and Niu, Ben},
title = {{KJD-25K: A 3D Multi-modal Dataset and Benchmark for Knee Joint Diagnosis}},
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
pages = {141--151}
}
- Downloads last month
- 54


