EricPeter/malaria-microscopy-vqa

Malaria Microscopy Visual Question Answering (VQA) Dataset

Dataset Description

This dataset contains multi-choice visual question answering pairs for malaria microscopy images, focusing on parasite detection, classification, species identification, and severity assessment.

Dataset Summary

The Malaria Microscopy VQA Dataset is designed for training and evaluating AI systems on medical microscopy image analysis tasks. It contains 67 questions across 59 unique microscopy images.

Key Features:

• Multi-choice questions with 3-6 answer options
• Comprehensive explanations for correct answers
• Multiple difficulty levels (easy, medium, hard)
• 6 question types covering clinical workflows
• High-quality filtering and validation

Supported Tasks

• Visual Question Answering: Multi-choice VQA for microscopy images
• Medical Image Classification: Parasite detection and classification
• Species Identification: Plasmodium species recognition
• Clinical Assessment: Parasitemia level and severity evaluation

Languages

English

Dataset Structure

Data Instances

{
  "question_id": "img_001_q1",
  "image_id": "id_example.jpg", 
  "question": "Is there evidence of malaria parasites in this blood smear?",
  "question_type": "detection",
  "choices": ["Yes, parasites are present", "No, no parasites visible", "Cannot determine"],
  "correct_answer": "Yes, parasites are present",
  "explanation": "Malaria trophozoites are visible within red blood cells.",
  "difficulty": "easy"
}

Data Fields

• question_id: Unique identifier for each question
• image_id: Identifier for the associated microscopy image
• question: The question text
• question_type: Type of question (detection, classification, species, severity, count, localization)
• choices: List of multiple choice options
• correct_answer: The correct answer from choices
• explanation: Detailed explanation for the correct answer
• difficulty: Question difficulty level (easy, medium, hard)

Data Splits

Split	Questions	Images
Train	67	59
Validation	14	13
Test	15	15

Question Type Distribution

Train Split

• detection: 35
• classification: 18
• species: 5
• severity: 9

Metadata Split

Val Split

• severity: 1
• detection: 6
• classification: 5
• species: 1
• count: 1

Test Split

• detection: 7
• classification: 6
• species: 2

Medical Domain Coverage

Parasite Types

• Trophozoite: Mature, amoeboid forms within red blood cells
• Ring forms: Early developmental stages with characteristic ring appearance
• Schizont: Dividing forms containing multiple nuclei
• Gametocyte: Sexual forms for mosquito transmission

Plasmodium Species

• P. falciparum: Most dangerous species, causes severe malaria
• P. vivax: Wide geographic distribution, causes relapsing malaria
• P. malariae: Causes quartan fever with 72-hour cycles
• P. ovale: Similar to P. vivax, less common

Clinical Parameters

• Parasitemia Levels: Low (<1%), Moderate (1-5%), High (5-15%), Severe (>15%)
• Cell Types: Red blood cells (RBCs), white blood cells (WBCs), platelets
• Morphological Features: Cytoplasm, nucleus, pigment, vacuoles

Dataset Creation

Source Data

• Lacuna/Zindi Challenge: Bounding box annotations for trophozoites and cell types
• NIH Malaria Cell Images: Parasitized vs uninfected cell classifications
• Research Datasets: Additional microscopy data sources

Quality Control

Multi-stage filtering pipeline inspired by Medical-CXR-VQA:

1. Language Quality: Grammar, clarity, choice consistency
2. Medical Accuracy: Domain terminology validation
3. Answer Consistency: Logical question-answer alignment
4. Complexity Analysis: Difficulty-appropriate complexity scoring
5. Diversity Control: Prevents repetitive questions

Annotation Process

Questions generated using:

• Template-based generation with medical domain knowledge
• Automated answer derivation from image annotations
• Quality filtering with configurable thresholds
• Expert-informed question templates and explanations

Uses

Direct Use

• Training VQA models for medical microscopy
• Benchmarking medical AI systems
• Educational tools for parasitology training
• Clinical decision support development

Research Applications

• Multi-modal medical AI research
• Few-shot learning in medical domains
• Domain adaptation for microscopy analysis
• Automated diagnostic system evaluation

Considerations for Use

Social Impact

This dataset supports development of AI systems that could:

• Positive: Improve malaria diagnosis in resource-limited settings
• Positive: Accelerate medical training and education
• Positive: Reduce diagnostic errors and improve patient outcomes

Limitations

• Limited to specific microscopy image types and staining methods
• Questions focus on common malaria scenarios (may not cover rare cases)
• Geographic/demographic bias from source datasets
• Simplified species identification (real diagnosis requires more context)

Recommendations

• Use in conjunction with expert medical validation
• Consider additional clinical context for real-world applications
• Validate performance across different microscopy protocols
• Regular updates as medical knowledge evolves

Additional Information

Dataset Curators

Generated using automated VQA pipeline with medical domain expertise.

Licensing Information

MIT License - Free for research and commercial use with attribution.

Citation Information

@dataset{malaria_vqa_2025,
  title={Malaria Microscopy VQA Dataset},
  author={Generated by Malaria VQA Pipeline},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/[YOUR_USERNAME]/malaria-vqa}
}

Contributions

Dataset created using advanced quality filtering pipeline inspired by Medical-CXR-VQA methodology.