VLM3D 2025 Task 1: Report Generation
Team: Medical Image Insights
Model submitted to the VLM3D@MICCAI 2025 Report Generation track.
Team Members
Jiawei Chenยฒ, Yiming Shiยณ, Yunhan Fangโด, Pengcheng Shiยน, Shuheng Zhangยน, Jiaqi Liuยน, Xinglin Zhangยน
ยน Medical Image Insights ยท ยฒ Xiโan Jiaotong University ยท ยณ Tsinghua University ยท โด Shanghai Jiaotong University
Official leaderboard
Overall rank: #1 โ VLM3D 2025 CT Report Generation.
https://research.forithmus.com/challenges/ct-report-generation#leaderboard
Model
- Architecture: MedM-VL
- Image encoder:
google/siglip2-so400m-patch14-384 - Language model:
Qwen/Qwen3-8B - Checkpoint:
final-MedM-VL-siglip-384-qwen3-8b-finetune - Files are stored at the repository root as 4 safetensors shards plus config and tokenizer files.
Training
- Dataset: CT-RATE.
- Pretrain: encoder and LLM frozen; connector trained for 1 epoch.
- Fine-tune: encoder, LLM, and connector trained for 1 epoch.
- Code and inference wrappers: github.com/yinghemedical/VLM3D-2025
- Classification track weights: VLM3D2025-Task2-AbnormalityClassifications
Local evaluation
CT-RATE valid_fixed, 3,039 CT volumes, conditioned on CT_CLS8 predicted labels:
| Metric | Value |
|---|---|
| BLEU-4 | 0.2146 |
| BLEU-mean | 0.3000 |
| ROUGE-L | 0.3424 |
| CIDEr | 0.1975 |
| RadBERT Macro F1 | 0.5494 |
| RadBERT Micro F1 | 0.5963 |
| CRG | 0.5431 |
Local evaluation only. CT-RATE valid was in the fine-tuning data, so these
are in-sample numbers.
Usage
Inference code: code/report_generation/inference/.
The model directory can be this repository or a local copy. SigLIP2 and Qwen3-8B are loaded from their public model cards.
Citation
If you use the CT-RATE dataset, please cite:
@article{hamamci2026generalist,
title={Generalist foundation models from a multimodal dataset for 3D computed tomography},
author={Hamamci, Ibrahim Ethem and Er, Sezgin and Wang, Chenyu and Almas, Furkan and Simsek, Ayse Gulnihan and Esirgun, Sevval Nil and Dogan, Irem and Durugol, Omer Faruk and Hou, Benjamin and Shit, Suprosanna and others},
journal={Nature Biomedical Engineering},
pages={1--19},
year={2026},
publisher={Nature Publishing Group UK London}
}
Data and license
CT-RATE images, reports, and patient identifiers are not included. Model weights are released under CC BY-NC 4.0 (non-commercial use only). Upstream model/data licenses apply to their respective assets.
- Downloads last month
- 95