Research map / Vision-language and multimodal / Multimodal LLMs
RAU: Reference-based Anatomical Understanding with Vision-Language Models
Yiwei Li, Yikang Liu, Jiaqi Guo, Lin Zhao, Zheyuan Zhang, Xiao Chen, Boris Mailhe, Ankush Mukherjee, Terrence Chen, Shanhui Sun
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with RL (GRPO 2400 steps) instead of vanilla | Qwen2.5VL-7B · RAOS-test-CT VQA | labeling accuracy: 16.62% to 70.68% | Improved |
| GRPO instead of SFT (5 epochs) | Qwen2.5VL-7B · RAOS-test-CT VQA | labeling accuracy: 64.11% -> 70.68% | Improved |
| add SFT then evaluate OOD Arcade vs GRPO | Qwen2.5VL-7B · Arcade VQA | labeling accuracy: 33.92% -> 48.37% | Improved |
| add global bounding-box matching (RL-Bbox) over RL-VQA | VLM · RAOS labeling | labeling accuracy: 74.68% to 78.16% | Improved |
| bounding-box matching on structurally complex data | VLM · Arcade labeling | labeling accuracy: 64.37% -> 41.09% | Hurt |
| add SAM2 segmentation (VLM+SAM2) over RL-VQA | VLM · RAOS labeling | labeling accuracy: 74.68% -> 89.38% | Improved |
| add SAM2 segmentation (VLM+SAM2) over RL-VQA | VLM · LERA labeling | labeling accuracy: 54.57% -> 61.87% | Improved |
| VLM+SAM2 vs strongest SAM2 baseline MedSAM2-Memory-Ref-SFT | RAU · four datasets average | Dice: 0.31 to 0.71 | Improved |
| VLM+SAM2 vs strongest baseline | RAU · four datasets average | gIoU: 0.21 to 0.50 | Improved |
| add RL stage on top of SFT VLM (ablation) | RAU · Arcade segmentation | Dice: 0.4370 -> 0.6754 | Improved |
| add RL-VQA initialization instead of SFT-VQA (ablation) | VLM+SAM2 · LERA labeling | labeling accuracy: 33.95% -> 61.87% | Improved |
| add RL-VQA initialization instead of SFT-VQA (ablation) | VLM+SAM2 · CAMUS labeling | labeling accuracy: 76.36% -> 95.41% | Improved |
Closest papers by meaning
- U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding (ICLR 2026 · similarity 0.87)
- Representing Part-Whole Hierarchies in Foundation Models by Learning Localizability Composability and Decomposability from Anatomy via Self Supervision (CVPR 2024 · similarity 0.87)
- RadZero: Similarity-Based Cross-Attention for Explainable Vision-Language Alignment in Chest X-ray with Zero-Shot Multi-Task Capability (NeurIPS 2025 · similarity 0.86)
- Interpretable Bilingual Multimodal Large Language Model for Diverse Biomedical Tasks (ICLR 2025 · similarity 0.86)
- PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images? (NeurIPS 2026 · similarity 0.86)
- MedSIGHT: Towards Grounded Visual Comprehension in Medical Large Vision-Language Models (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
