neuripsqgr2026/Med4VQA

Med4-VQA (Anonymous Submission)

This repository contains the Med4-VQA dataset, a multimodal medical visual question answering dataset designed for joint reasoning and spatial grounding.

---

📌 Overview

Med4-Grounded VQA is constructed to support research in grounded medical VQA, where models are required to not only answer clinical questions but also localize relevant visual evidence.

The dataset covers four imaging modalities:

• CT
• MRI
• X-ray
• Ultrasound

It includes:

• Medical images
• Question-answer pairs
• Segmentation masks
• Bounding box annotations
• Grounded reasoning annotations

to enable both semantic reasoning and spatial localization.

---

📊 Dataset Statistics

Split	Images	VQA Pairs
Train	6,170	123,400
Test	728	14,560

• Total images: 6,898
• Total VQA pairs: 137,960
• Questions per image: 20
  • 10 open-ended
  • 10 closed-ended

---

🧠 Annotations

Each sample includes:

• Medical image
• Question
• Answer
• Grounded reasoning
• Semantic region labels
• Bounding boxes
• Segmentation masks

The dataset contains 27 clinically relevant labels across modalities.

---

🏥 Modalities

Modality	Description
CT	Multi-organ anatomical structures and lesions
MRI	Soft tissue variations and signal properties
X-ray	Pneumothorax-focused findings
Ultrasound	Lesion characterization (e.g., breast, thyroid)

---

❓ Question Types

Each image includes:

• Open-ended questions (free-form reasoning)
• Closed-ended questions (structured answers)

Example questions:

• “What abnormalities are present in this image?”
• “Where is the dominant abnormality located?”
• “Is the lesion homogeneous or heterogeneous?”

---

📂 Metadata Columns

Column	Description
row_id	Unique row index
q_id	Unique question identifier
image	Path to the medical image
mask_path	Path to segmentation mask
modality	Imaging modality (CT, MRI, Ultrasound, X-ray)
type	Dataset sample type
eval_type	Evaluation category
qa_style	Question-answer style
question	Medical visual question
answer	Ground-truth answer
reason	Grounded medical reasoning explanation
visual_regions	Ground-truth region labels
visual_locations	Ground-truth localization annotations
visual_regions_text	Textual region grounding
visual_locations_text	Textual localization grounding
choices	Multiple-choice candidate answers
answer_closed	Closed-form answer label

---

⚙️ Annotation Format

The dataset uses structured XML-style annotations:

<answer>...</answer>
<reason>...</reason>
<visual_regions>...</visual_regions>
<visual_locations>...</visual_locations>