Skip to content

Dataset Access Log

Mission: Phase 1 Dataset Acquisition for Fairness-Aware Skin Cancer Detection Initiated: 2025-10-13 Status: IN PROGRESS Agent: the_didact (MENDICANT_BIAS framework)

Primary Datasets Status

1. HAM10000 (Human Against Machine with 10,000 Dermoscopic Images)

Status: ✅ ACCESS CONFIRMED - PUBLIC DATASET

Details: - Size: 10,015 dermoscopic images - Classes: 7 diagnostic categories (melanoma, basal cell carcinoma, melanocytic nevi, etc.) - Metadata: Age, sex, localization - Skin Tone Limitation: <5% FST V-VI (majority lighter skin)

Access Methods: 1. Harvard Dataverse (Primary): - URL: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T - DOI: 10.7910/DVN/DBW86T - License: Non-commercial use only - Method: Direct download (requires Terms of Use confirmation)

  1. Kaggle (Alternative):
  2. URL: https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000

  3. Method: Kaggle account required, free download

  4. ISIC Archive API:

  5. URL: https://isic-archive.com/api/v1
  6. Method: Programmatic access via API calls

Download Action: Use Kaggle CLI or Harvard Dataverse direct download Target Directory: data/raw/ham10000/ Next Steps: - Download dataset (10,015 images + metadata CSV) - Verify file integrity - Parse metadata for FST distribution analysis

2. ISIC 2019 Challenge Dataset

Status: ✅ ACCESS CONFIRMED - PUBLIC DATASET

Details: - Size: 25,331 training images (ISIC 2019), 33,126 total (ISIC 2020) - Classes: 8 diagnostic categories - Metadata: Patient demographics, lesion location - Skin Tone Limitation: <3% FST V-VI, no explicit FST labels

Access Methods: 1. ISIC Challenge Website (Primary): - URL: https://challenge.isic-archive.com/data/ - Method: Direct download (training data ~2.6GB) - Registration: Free account required

  1. AWS Open Data Registry:
  2. URL: https://registry.opendata.aws/isic-archive/

  3. Method: S3 bucket access (public)

  4. Hugging Face (Fed-ISIC-2019):

  5. URL: https://huggingface.co/datasets/flwrlabs/fed-isic2019
  6. Method: Datasets library download

Download Action: ISIC Challenge direct download for ISIC 2019 training set Target Directory: data/raw/isic2019/ Next Steps: - Download ISIC 2019 training images + ground truth CSV - Parse metadata (age, sex, anatomical site) - Note: Will require FST annotation in Phase 1 (annotation protocol in progress)

3. Fitzpatrick17k

Status: ⏳ ACCESS REQUEST REQUIRED

Details: - Size: 16,577 clinical images - FST Distribution: ~8% FST V-VI (better than baseline but still imbalanced) - Metadata: Fitzpatrick skin type (dual annotation), diagnosis, ITA values - Source: DermaAmin and Atlas Dermatologico atlases

Access Methods: 1. GitHub Repository (Metadata only): - URL: https://github.com/mattgroh/fitzpatrick17k - Available: Fitzpatrick17k.csv (annotations)

  1. Image Access (Request Required):
  2. Method: Fill out access request form (linked in GitHub README)
  3. Contact: Matthew Groh (MIT researcher)
  4. Alternative: Download from original source URLs (provided in CSV)

Citation: - Groh, M., Harris, C., Daneshjou, R., et al. (2021). Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset. arXiv:2104.09957

Access Request Draft:

Subject: Research Access Request - Fitzpatrick17k Dataset

Dear Dr. Groh,

I am writing to request access to the Fitzpatrick17k dataset for academic research purposes.

Project: Fairness-Aware AI for Skin Cancer Detection Across All Fitzpatrick Skin Types
Institution: [California State University, San Bernardino / Independent Research]
Principal Investigator: [Dr. Nabeel Alzahrani / Student: Jasmin Flores]

Research Objective:
We are developing a fairness-aware skin cancer detection system to address the 15-30%
performance gap observed in existing AI models for darker skin tones (FST IV-VI). Our
project builds upon the comprehensive survey "AI Skin Cancer Detection Across Skin Tones"
(Flores & Alzahrani, 2025) and aims to implement state-of-the-art fairness techniques
(FairSkin diffusion, FairDisCo adversarial debiasing, CIRCLe color-invariant learning).

Dataset Usage:
- Training fairness-aware deep learning models (hybrid ConvNeXt-Swin Transformer architecture)
- Stratified evaluation across FST groups (quantifying AUROC, EOD, ECE per skin tone)
- Comparative analysis with baseline datasets (HAM10000, ISIC) to demonstrate fairness improvement

We commit to:
1. Using the dataset solely for non-commercial academic research
2. Proper attribution in all publications and presentations
3. Sharing subgroup performance metrics transparently (model card documentation)
4. Not re-distributing the dataset without authorization

Timeline: Phase 1 baseline training (4 weeks), full project completion (24 weeks)

Contact Information:
- Email: [your_email@csusb.edu]
- GitHub: https://github.com/[username]/skin-cancer-fairness
- Project Framework: MENDICANT_BIAS multi-agent system

Thank you for advancing fairness-aware dermatology AI research. Your Fitzpatrick17k
dataset is foundational to addressing healthcare disparities.

Sincerely,
[Your Name]
[Affiliation]

Target Directory: data/raw/fitzpatrick17k/ Next Steps: - Submit access request via GitHub form - Follow up in 3-5 business days - Alternative: Use CSV URLs to download from original atlases (if permitted)

4. DDI (Diverse Dermatology Images)

Status: ✅ ACCESS AVAILABLE - RESEARCH USE AGREEMENT REQUIRED

Details: - Size: 656 images (570 unique patients) - FST Distribution: 34% FST V-VI (EXCELLENT diversity) - Metadata: Pathologically confirmed diagnoses, clinician-rated FST (gold standard) - Source: Stanford Clinics (2010-2020 retrospective selection)

Access Methods: 1. Stanford AIMI Shared Datasets Portal: - URL: https://aimi.stanford.edu/datasets/ddi-diverse-dermatology-images - URL: https://stanfordaimi.azurewebsites.net/datasets/35866158-8196-48d8-87bf-50dca81df965 - DOI: https://doi.org/10.71718/kqee-3z39 - License: Research Use Agreement (non-commercial, no re-identification)

Research Use Agreement Key Terms: - Personal, non-commercial research only - No commercial use, sale, or monetization - No attempt to re-identify individual data subjects - Indemnification of Stanford from claims/damages

Access Action: - Navigate to Stanford AIMI portal - Accept Research Use Agreement - Download dataset directly (no institutional approval needed for research use)

Citation: - Daneshjou, R., Barata, C., Betz-Stablein, B., et al. (2022). Disparities in dermatology AI performance on a diverse, curated clinical image set. Science Advances, 8(25), eabq6147.

Target Directory: data/raw/ddi/ Next Steps: - Accept Research Use Agreement on Stanford AIMI portal - Download 656 images + metadata - Verify pathology-confirmed labels - Prioritize for fairness evaluation (highest FST V-VI representation)

5. MIDAS (MRA-MIDAS: Multimodal Image Dataset for AI-based Skin Cancer)

Status: ✅ ACCESS AVAILABLE - STANFORD AIMI PORTAL

Details: - Size: Dual-center, prospectively recruited dataset - Modalities: Paired dermoscopic + clinical images - FST Distribution: ~28% FST V-VI (estimated based on Stanford diversity metrics) - Metadata: Patient-level clinical metadata, histopathologic confirmation - Unique Feature: Prospective recruitment (higher real-world fidelity vs retrospective)

Access Methods: 1. Stanford AIMI Portal: - URL: https://aimi.stanford.edu/datasets/mra-midas-Multimodal-Image-Dataset-for-AI-based-Skin-Cancer - DOI: https://doi.org/10.71718/15nz-jv40 - License: Non-commercial research use

Citation: - McCoy, L. G., Naik, B., Saunders, H., et al. (2024). Multimodal Image Dataset for AI-based Skin Cancer (MIDAS) Benchmarking. medRxiv. DOI: 10.1101/2024.06.27.24309562

Access Action: - Navigate to Stanford AIMI portal - Accept Research Use Agreement (same as DDI) - Download multimodal dataset (dermoscopic + clinical pairs)

Target Directory: data/raw/midas/ Next Steps: - Download dataset from Stanford AIMI - Parse multimodal structure (separate dermoscopic/clinical folders) - Leverage clinical images for metadata fusion architecture (Phase 3)

6. SCIN (Skin Condition Image Network)

Status: ✅ ACCESS CONFIRMED - OPEN GITHUB REPOSITORY

Details: - Size: 10,000+ images - FST Distribution: ~33% FST V-VI, balanced FST distribution (major advantage) - Metadata: Self-reported demographics, symptoms, dermatologist labels, estimated FST (eFST), estimated MST (eMST) - Source: Crowdsourced from US internet users via Google Search Ads - Unique Feature: Real-world, in-the-wild images (not clinical dermoscopy)

Access Methods: 1. GitHub Repository: - URL: https://github.com/google-research-datasets/scin - License: Open access for research, education, development - Method: Git clone or direct download

Key Features: - Dermatologist estimates of Fitzpatrick Skin Type (eFST) - Layperson labeler estimates of Monk Skin Tone (eMST) - Common allergic, inflammatory, and infectious conditions (not just tumors) - Crowdsourced with informed consent

Citation: - Jain, A., Lipman, M., Liu, Y., et al. (2024). Crowdsourcing Dermatology Images with Google Search Ads: Creating a Real-World Skin Condition Dataset. arXiv:2402.18545

Access Action: Git clone repository Target Directory: data/raw/scin/ Next Steps: - Clone GitHub repository: git clone https://github.com/google-research-datasets/scin.git - Review dataset schema (dataset_schema.md in repo) - Explore scin_demo.ipynb for loading examples - Prioritize for Phase 2 training (excellent FST balance + real-world diversity)

Synthetic Augmentation Datasets (Phase 2)

7. FairSkin / DermDiff Synthetic Generation

Status: ⏳ IMPLEMENTATION RESEARCH IN PROGRESS

Objective: Generate 60,000 synthetic images with balanced FST distribution

Models Identified:

  1. FairSkin (Oct 2024):
  2. Paper: https://arxiv.org/abs/2410.22551
  3. Method: Three-level resampling, class diversity loss, balanced sampling
  4. Code: NOT YET AVAILABLE (paper just published)

  5. Implementation: Will require custom development using Hugging Face Diffusers

  6. DermDiff (March 2025):

  7. Paper: https://arxiv.org/abs/2503.17536
  8. Method: Skin tone detector + race-conditioned diffusion + multimodal text-image learning
  9. Implementation: PyTorch with HuggingFace Diffusers + OpenAI APIs
  10. Generated Dataset: 60k synthetic images (30k benign, 30k malignant)

  11. Code: NOT EXPLICITLY AVAILABLE (check arXiv code links)

  12. From Majority to Minority (June 2024):

  13. Paper: https://arxiv.org/html/2406.18375
  14. GitHub: https://github.com/janet-sw/skin-diff
  15. Award: MICCAI ISIC Workshop 2024 Honorable Mention
  16. Method: Stable Diffusion via Textual Inversion + LoRA
  17. Implementation: AVAILABLE (Hugging Face Diffusers)
  18. RECOMMENDED FOR PHASE 2 IMPLEMENTATION

Next Steps for Phase 2: - Review janet-sw/skin-diff GitHub repository - Implement tone-conditioned Stable Diffusion using HuggingFace Diffusers - Train on HAM10000 + ISIC 2019 + Fitzpatrick17k (once acquired) - Generate 60k synthetic images with target FST distribution: 25% FST V-VI - Validate quality: FID <20, LPIPS <0.1, expert dermatologist review

Documentation: See docs/synthetic_augmentation.md

Summary Statistics

Dataset Size FST V-VI % Access Status Priority
HAM10000 10,015 <5% ✅ Public HIGH (baseline)
ISIC 2019 25,331 <3% ✅ Public HIGH (baseline)
Fitzpatrick17k 16,577 ~8% ⏳ Request HIGH (FST labels)
DDI 656 34% ✅ Available CRITICAL (diversity)
MIDAS Variable ~28% ✅ Available HIGH (multimodal)
SCIN 10,000+ ~33% ✅ Public HIGH (real-world)
Synthetic (Phase 2) 60,000 25% target ⏳ Pending MEDIUM (augmentation)

Total Training Dataset Target (Phase 2): ~130,000 images with 25%+ FST V-VI representation

Action Items

IMMEDIATE (Week 1): - [ ] Download HAM10000 from Kaggle (using Kaggle API) - [ ] Download ISIC 2019 from challenge website - [ ] Submit Fitzpatrick17k access request to Dr. Matthew Groh - [ ] Download DDI from Stanford AIMI portal - [ ] Download MIDAS from Stanford AIMI portal - [ ] Clone SCIN from GitHub repository

SHORT-TERM (Week 2-3): - [ ] Verify all dataset file integrity (checksums, image loading) - [ ] Parse and merge metadata CSVs - [ ] Analyze FST distribution across datasets - [ ] Implement FST annotation for ISIC 2019 (no native labels) - [ ] Create stratified train/val/test splits (balanced FST)

MEDIUM-TERM (Week 4-6): - [ ] Implement FST annotation protocol using ITA + Monk Skin Tone - [ ] Annotate ISIC 2019 images (automated ITA + manual validation) - [ ] Prepare Phase 2 synthetic data generation pipeline

Risk Assessment

High Risk: - Fitzpatrick17k access delay (mitigation: use CSV URLs to download from original atlases if form delayed)

Medium Risk: - ISIC 2019 download size (2.6GB+, may require bandwidth/storage management) - FST annotation quality for datasets without native labels (mitigation: dual annotation + ITA validation)

Low Risk: - Public datasets (HAM10000, ISIC, SCIN) - minimal access barriers - Stanford AIMI datasets (DDI, MIDAS) - streamlined Research Use Agreement

Contact Information

Dataset Curators: - HAM10000: Peter Tschandl (Medical University of Vienna) - Fitzpatrick17k: Matthew Groh (MIT Media Lab) - DDI: Roxana Daneshjou (Stanford Dermatology) - MIDAS: Leo McCoy, Bhavik Naik (Stanford AIMI) - SCIN: Abhishek Jain (Google Health)

Institutional Contact: - PI: Dr. Nabeel Alzahrani (CSUSB) - Researcher: Jasmin Flores (CSUSB)

Last Updated: 2025-10-13 Next Review: 2025-10-14 (daily updates during acquisition phase) Maintained by: the_didact (MENDICANT_BIAS framework)