Research map / Vision-language and multimodal / Visual question answering
A Structured, Tagged, and Localized Visual Question Answering Dataset with Full Sentence Answers and Scene Graphs for Chest X-ray Images
Philip Müller, Friederike Jungmann, Georgios Kaissis, Daniel Rueckert
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| fine-tune on 1M fine-tuning-grade QA pairs instead of 1M pre-training-grade (ablation) | Llama 3.2 3B + Rad-DINO · Structured VQA | RadStrucVQA Logical Prec.: 0.67 -> 0.76 | Improved |
| add 1M fine-tuning-grade stage after 1M pre-training-grade (ablation) | Llama 3.2 3B + Rad-DINO · Structured VQA | RadStrucVQA Logical Prec.: 0.67 -> 0.78 | Improved |
| PT(1M) then FT(1M) instead of 2M pre-training samples (ablation) | Llama 3.2 3B + Rad-DINO · Structured VQA | RadStrucVQA Logical Prec.: 0.68 -> 0.78 | Improved |
| double pre-training samples from 1M to 2M (ablation) | Llama 3.2 3B + Rad-DINO · Structured VQA | RadStrucVQA Logical Prec.: 0.67 -> 0.68 | No clear effect |
| fine-tune on FT-grade data instead of PT-grade (ablation) | Llama 3.2 3B + Rad-DINO · Structured VQA | Grounding Rec.: 0.92 -> 0.89 | Hurt |
| train on CXR-QBA vs zero-shot MAIRA-2 | Llama 3.2 3B + Rad-DINO · Structured VQA | RadStrucVQA Logical Prec.: 0.25 -> 0.78 | Improved |
| train on CXR-QBA vs zero-shot Qwen3-VL | Llama 3.2 3B + Rad-DINO · Structured VQA | RadStrucVQA Grounding Rec.: 0.51 -> 0.90 | Improved |
| LLM-based scene graph with fine-grained tags vs Chest ImaGenome graphs | CXR-QBA scene graphs · CXR-LT 2024 gold long-tail classes | MCC: 0.59 -> 0.71 | Improved |
| our scene graph boxes vs Chest ImaGenome | CXR-QBA scene graphs · MS-CXR | IoU@30: 0.45 -> 0.51 | Improved |
| our scene graph boxes vs Chest ImaGenome | CXR-QBA scene graphs · MS-CXR | IoT@30: 0.98 -> 0.94 | Hurt |
| our scene graph boxes vs Chest ImaGenome | CXR-QBA scene graphs · MS-CXR | IoP@30: 0.48 -> 0.56 | Improved |
| Llama 3.1 8B judge vs larger judges in RadStrucVQA (ablation) | RadStrucVQA metric · entailment sub-metrics | correlation: 0.88 to 0.97 | No clear effect |
Closest papers by meaning
- 3D-RAD: A Comprehensive 3D Radiology Med-VQA Dataset with Multi-Temporal Analysis and Diverse Diagnostic Tasks (NeurIPS 2025 · similarity 0.88)
- CXReasonBench: A Benchmark for Evaluating Structured Diagnostic Reasoning in Chest X-rays (NeurIPS 2025 · similarity 0.87)
- DiN: Diffusion Model for Robust Medical VQA with Semantic Noisy Labels (CVPR 2025 · similarity 0.87)
- Alignment, Mining and Fusion: Representation Alignment with Hard Negative Mining and Selective Knowledge Fusion for Medical Visual Question Answering (CVPR 2025 · similarity 0.87)
- Falcon Medical Visual Question Answering (AAAI 2025 · similarity 0.87)
- OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.87)
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.
