mugezhang/medical-diff-vqa-longitudinal

Medical-Diff-VQA Longitudinal (Enhanced)

Dataset Description

This is an enhanced, filtered subset of Medical-Diff-VQA with pre-computed longitudinal imaging context.

Key Features

✅ Pre-computed context images: Each question includes DICOM IDs for historical images ✅ Temporal information: Days from first visit for all images ✅ Most recent N=4 strategy: Context images are the 4 most recent studies before the reference pair ✅ Complete image metadata: View positions, dates, and paths included ✅ Ready for experiments: No need to recompute context extraction

Filtering Criteria

From the original Medical-Diff-VQA (700K questions):

1. Question Type: Only "difference" subtype (164K questions)
2. Longitudinal Context: Only patients with ≥1 additional study before the reference pair (124K questions)

This ensures every question can be evaluated in both:

• Baseline condition: 2 reference images only
• Longitudinal condition: 2 reference images + up to 4 historical context images

Dataset Statistics

• Total questions: 123,980
• Unique patients: 19,410
• Train: 99,194 (80.0%)
• Validation: 12,349 (10.0%)
• Test: 12,437 (10.0%)

Context Image Distribution:

• 1 context image: 29,871 questions (24.1%)
• 2 context images: 19,119 questions (15.4%)
• 3 context images: 13,556 questions (10.9%)
• 4 context images: 61,434 questions (49.6%)
• Average: 2.86 context images per question

Data Structure

Fields

Field	Type	Description
id	int	Unique question identifier
subject_id	int	MIMIC-CXR patient ID
question_type	string	Always "difference"
question	string	VQA question text
answer	string	Ground truth answer
split	string	train/val/test
num_context_images	int	Number of context images (1-4)
context_images	JSON string	Array of context image objects
main_image	JSON string	Current/study image object
ref_image	JSON string	Reference image object

Image Object Structure

Each image object (in context_images, main_image, ref_image) contains:

{
  "study_id": 55088298,
  "dicom_id": "61976388-5e534624-f6465079-76ea9caf-116f9938",
  "view_position": "PA",
  "study_date": "2177-10-02",
  "days_from_first_visit": 5,
  "image_path": "/fs/scratch/.../p18/p18936629/s55088298/61976388-...-.jpg"
}

Example

{
  "id": 0,
  "subject_id": 18936629,
  "question": "what has changed compared to the reference image?",
  "answer": "the main image has an additional finding of pneumothorax...",
  "split": "train",
  "num_context_images": 2,
  "context_images": "[{\"study_id\": 51847070, \"dicom_id\": \"6b1bf785-...\", ...}, ...]",
  "main_image": "{\"study_id\": 55088298, \"dicom_id\": \"61976388-...\", ...}",
  "ref_image": "{\"study_id\": 54989200, \"dicom_id\": \"15a2bf67-...\", ...}"
}

Usage

Load the Dataset

from datasets import load_dataset
import json

# Load dataset
dataset = load_dataset("mugezhang/medical-diff-vqa-longitudinal")

# Access a question
example = dataset['train'][0]

# Parse nested JSON fields
context_images = json.loads(example['context_images'])
main_image = json.loads(example['main_image'])
ref_image = json.loads(example['ref_image'])

print(f"Question: {example['question']}")
print(f"Context images: {len(context_images)}")
for i, ctx in enumerate(context_images, 1):
    print(f"  {i}. Study {ctx['study_id']} (Day {ctx['days_from_first_visit']}, {ctx['view_position']})")

Access MIMIC-CXR Images

You need access to MIMIC-CXR-JPG v2.0.0:

from PIL import Image

def load_image(image_obj, mimic_root):
    """Load image from MIMIC-CXR-JPG dataset."""
    subject_id = str(image_obj['subject_id'])
    study_id = str(image_obj['study_id'])
    dicom_id = image_obj['dicom_id']

    # Construct path
    prefix = f"p{subject_id[:2]}"
    patient_dir = f"p{subject_id}"
    study_dir = f"s{study_id}"

    path = f"{mimic_root}/files/{prefix}/{patient_dir}/{study_dir}/{dicom_id}.jpg"
    return Image.open(path)

# Example usage
example = dataset['train'][0]
main_img = json.loads(example['main_image'])
ref_img = json.loads(example['ref_image'])

# Load images
main_image = load_image(main_img, "/path/to/mimic-cxr-jpg/2.0.0")
ref_image = load_image(ref_img, "/path/to/mimic-cxr-jpg/2.0.0")

Build VQA Prompts

def create_longitudinal_prompt(example):
    """Create prompt with longitudinal context."""
    context_images = json.loads(example['context_images'])
    main_image = json.loads(example['main_image'])
    ref_image = json.loads(example['ref_image'])

    prompt = "HISTORICAL CONTEXT:\n"
    for i, ctx in enumerate(context_images, 1):
        prompt += f"Image {i}: Day {ctx['days_from_first_visit']} ({ctx['view_position']})\n"

    prompt += f"\nREFERENCE PAIR:\n"
    prompt += f"Reference: Day {ref_image['days_from_first_visit']}\n"
    prompt += f"Current: Day {main_image['days_from_first_visit']}\n"

    prompt += f"\nQuestion: {example['question']}"

    return prompt

Context Selection Strategy

Most Recent N=4: For each question, we select up to 4 most recent studies that occurred before both images in the reference pair.

Why "most recent"?

1. ✅ Clinically relevant (recent history matters most)
2. ✅ Matches the task (understanding immediate changes)
3. ✅ Standard in clinical practice
4. ✅ Reproducible and deterministic

Temporal Ordering

Context images are ordered chronologically (earliest to most recent), making it easy to understand disease progression.

Experimental Setup

This dataset supports comparing two conditions:

Condition	Images Provided	Expected Performance
Baseline	2 (ref + main)	Lower baseline
Longitudinal	2 + N context (up to 6 total)	Improved with context

The hypothesis: Providing longitudinal context helps VLMs better understand changes.

Source Data

• Original Dataset: Medical-Diff-VQA v1.0.1
• Images: MIMIC-CXR-JPG v2.0.0
• Filtering: 75.4% of "difference" questions have longitudinal context

License & Access

⚠️ Requires PhysioNet Credentialing

This dataset inherits restrictions from Medical-Diff-VQA and MIMIC-CXR:

• Requires PhysioNet credentialed access
• For research use only
• Must comply with MIMIC-CXR data use agreement

Citation

If you use this dataset, please cite:

@article{medvqa2023,
  title={Medical-Diff-VQA: A Large-Scale Benchmark for Difference Visual Question Answering in Medical Imaging},
  journal={PhysioNet},
  year={2023}
}

@article{johnson2019mimic,
  title={MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports},
  author={Johnson, Alistair EW and others},
  journal={Scientific Data},
  volume={6},
  number={1},
  pages={317},
  year={2019}
}

Changelog

v2.0 (Enhanced) - December 2025

• ✅ Added pre-computed context images with DICOM IDs
• ✅ Added temporal information (days from first visit)
• ✅ Added view positions and image paths
• ✅ Implemented "most recent N=4" selection strategy
• ✅ Reduced from 275MB JSON to 33MB Parquet

v1.0 (Basic) - December 2025

• Basic filtering with context counts only

Contact

For questions or issues, please open an issue in the repository.