File Information

This dataset simulates real-world medical data complexities to provide hands-on experience in data cleaning and preprocessing for AI in healthcare. Deliberately retains common data issues (common data issues) found in real clinical settings, serving as a practical sandbox for:

• 🚮 Data Cleaning Mastery: Tackle duplicates, inconsistent resolutions, and naming heterogeneity.
• ⚖️ Class Imbalance Solutions: Experiment with techniques like SMOTE or augmentation.
• 🎯 Medical AI Readiness: Prepare raw clinical data for XAI (Explainable AI) models.

Content

11,784 ultrasound images intentionally uncleaned:

• infected/: 6,784 images (PCOS-positive cases)
• noninfected/: 5,000 images (Healthy ovaries)

🚩 Real-World Challenges Included:

1. Duplicates: 1,956+ groups of identical images (intra-class & cross-class)
2. Multi-Resolution Mix: Images from 255x247px to 984x848px
3. Metadata Gaps: No clinical patient data (simulating HIPAA-restricted scenarios)
4. Class Imbalance: 57.5% vs 42.5% distribution
5. Noise Artifacts: Blurring, rotation variants, and naming inconsistencies

Sources & Methodology

• Curated from: Retrospective ultrasound studies across 3 clinics (2018-2022)
• Ethical Compliance: Patient identifiers removed, DICOM metadata stripped
• Annotated: By radiology residents under consultant supervision

Inspiration

Born from the need to bridge the gap between:

• 📊 Clean Tutorial Datasets (MNIST, CIFAR)
• 🏥 Messy Real Clinical Data
  Use this to practice the unglamorous but critical 80% of ML work - data wrangling!

Potential Applications

• 🚮 Data cleaning pipelines for medical imaging
• 🔍 Duplicate detection algorithms
• ⚖️ Benchmarking class-balancing techniques
• 📏 Resolution standardization methods

Note to Practitioners

"The true test of an ML engineer isn't model architecture choice, but transforming messy data into trainable gold." - Use these imperfections as your training ground