Fitzpatrick Skin Type (FST) Annotation Protocol

Version: 1.0 Date: 2025-10-13 Framework: MENDICANT_BIAS - the_didact research division Purpose: Standardized protocol for annotating skin tone in dermatology images to ensure fairness-aware model development

---

Executive Summary

This protocol establishes a rigorous, reproducible methodology for annotating Fitzpatrick Skin Type (FST) and Monk Skin Tone (MST) in dermatological images. Accurate skin tone annotation is CRITICAL for: - Training fairness-aware AI models - Stratified evaluation (performance metrics per FST group) - Identifying and mitigating algorithmic bias

Key Principle: Skin tone annotation is inherently subjective. This protocol combines automated algorithms (ITA), expert annotation, and crowd consensus to maximize reliability.

---

1. Annotation Scales

1.1 Fitzpatrick Skin Type (FST) - 6-Point Scale

Original Purpose: Classify skin's response to UV exposure Clinical Definition:

FST	Description	UV Response	Prevalence (Global)
I	Very fair skin, pale white	Always burns, never tans	~2%
II	Fair skin, white	Usually burns, tans minimally	~12%
III	Medium skin, white to light brown	Sometimes burns, tans uniformly	~28%
IV	Olive skin, moderate brown	Rarely burns, tans easily	~45%
V	Brown skin, dark brown	Very rarely burns, tans very easily	~11%
VI	Very dark skin, deeply pigmented	Never burns	~2%

Limitations: - Originally designed for UV sensitivity, not visual appearance - Coarse granularity (6 categories) - Poor inter-rater reliability for intermediate types (FST III-IV) - Racial bias potential (historically correlated with race)

1.2 Monk Skin Tone (MST) - 10-Point Scale

Modern Alternative: Perceptual skin tone classification (Google AI, 2022) Advantages over Fitzpatrick: - Finer granularity (10 categories vs 6) - Based on visual appearance, not UV response - Better representation of intermediate tones - Higher inter-rater agreement in recent studies

MST Scale: - Tones 1-2: Very light (corresponds to FST I-II) - Tones 3-4: Light (FST II-III) - Tones 5-6: Medium (FST III-IV) - Tones 7-8: Dark (FST IV-V) - Tones 9-10: Very dark (FST V-VI)

Recommendation: Use MST for NEW annotations, maintain FST for backward compatibility with existing datasets.

---

2. Automated Annotation: Individual Typology Angle (ITA)

2.1 ITA Algorithm Overview

ITA Formula:

ITA = [arctan((L* - 50) / b*)] × (180 / π)

Where: - L = Lightness (0-100 in CIELAB color space) - b = Blue-yellow axis (-128 to +127 in CIELAB color space)

ITA to FST Mapping (Established in Literature):

ITA Range	Fitzpatrick Type	MST Equivalent
> 55°	FST I (Very light)	MST 1-2
41° - 55°	FST II (Light)	MST 2-3
28° - 41°	FST III (Intermediate)	MST 3-5
19° - 28°	FST IV (Tan)	MST 5-7
-30° - 19°	FST V (Brown)	MST 7-9
< -30°	FST VI (Dark brown/Black)	MST 9-10

2.2 ITA Implementation Steps

Preprocessing Requirements: 1. Image standardization: Remove hair, shadows, artifacts 2. Region of Interest (ROI): Select representative skin patch (avoid lesion, hair, shadows) 3. Color space conversion: RGB → CIELAB

Automated ITA Calculation (Python pseudocode):

import cv2
import numpy as np
from skimage import color

def calculate_ita(image_path, roi=None):
    """
    Calculate Individual Typology Angle (ITA) for skin tone classification.

    Args:
        image_path: Path to dermatology image
        roi: Region of Interest coordinates (x, y, w, h). If None, use entire image.

    Returns:
        ita_value: ITA angle in degrees
        fst: Fitzpatrick Skin Type (I-VI)
        mst: Monk Skin Tone (1-10)
    """
    # Load image
    img = cv2.imread(image_path)
    img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

    # Extract ROI (if specified)
    if roi:
        x, y, w, h = roi
        img_rgb = img_rgb[y:y+h, x:x+w]

    # Convert to CIELAB color space
    img_lab = color.rgb2lab(img_rgb)

    # Calculate mean L* and b* values
    L_mean = np.mean(img_lab[:, :, 0])
    b_mean = np.mean(img_lab[:, :, 2])

    # Calculate ITA
    ita_value = np.degrees(np.arctan((L_mean - 50) / b_mean))

    # Map ITA to FST
    if ita_value > 55:
        fst = 1
        mst = 1
    elif 41 <= ita_value <= 55:
        fst = 2
        mst = 2
    elif 28 <= ita_value < 41:
        fst = 3
        mst = 4
    elif 19 <= ita_value < 28:
        fst = 4
        mst = 6
    elif -30 <= ita_value < 19:
        fst = 5
        mst = 8
    else:  # ita_value < -30
        fst = 6
        mst = 10

    return ita_value, fst, mst

Limitations of ITA: - Sensitive to illumination conditions - Requires clean skin region (no lesions, hair, shadows) - May misclassify erythema (redness) as lighter skin - Less reliable for intermediate tones (FST III-IV)

Mitigation: Use ITA as initial estimate, validate with human annotation.

---

3. Human Annotation Protocol

3.1 Annotation Platform: LabelBox

Why LabelBox: - Used in Fitzpatrick17k dataset creation (validated in literature) - Supports custom classification taxonomies - Facilitates multi-annotator consensus workflows - Tracks inter-rater reliability metrics

Alternative Platforms: - CVAT (Computer Vision Annotation Tool): Open-source, self-hosted - Label Studio: Open-source, ML-assisted labeling - Custom annotation tool: Python + Flask + React (if specific needs arise)

3.2 Annotator Selection

Expertise Levels: 1. Expert Annotators (Gold Standard): - Board-certified dermatologists - Minimum 2 years clinical experience - Trained in Fitzpatrick/Monk scales - Target: 2-3 experts for validation set

1. Trained Layperson Annotators:
2. Completed FST/MST training module (2-hour online course)
3. Passed calibration test (Kappa > 0.7 with expert consensus)
4. Target: 3-5 annotators per image for crowd consensus

Training Module Content: - Fitzpatrick Scale history and clinical context - Monk Skin Tone Scale visual guide - Common annotation pitfalls (erythema, shadows, lighting) - Practice annotation with expert-validated examples

3.3 Annotation Workflow

Step 1: Automated Pre-Annotation - Run ITA algorithm on all images - Generate initial FST/MST estimates - Flag images with ambiguous ITA values (e.g., FST III-IV boundary)

Step 2: Multi-Annotator Review - Each image annotated by 3 independent annotators - Annotators see image + ITA estimate (as reference, not ground truth) - Annotators classify using both FST (6-point) and MST (10-point)

Step 3: Consensus Resolution - High Agreement (3/3 or 2/3 same label): Accept majority label - Disagreement (all different): Route to expert dermatologist for adjudication - Borderline Cases: Accept range (e.g., "FST III-IV") for stratified evaluation

Step 4: Quality Control - Random 10% sample re-annotated by expert dermatologist - Calculate Cohen's Kappa (inter-rater reliability) - Target: Kappa > 0.7 (substantial agreement)

---

4. Inter-Rater Reliability Metrics

4.1 Cohen's Kappa

Formula:

κ = (P_observed - P_expected) / (1 - P_expected)

Interpretation: - κ < 0.00: No agreement - κ = 0.00-0.20: Slight agreement - κ = 0.21-0.40: Fair agreement - κ = 0.41-0.60: Moderate agreement - κ = 0.61-0.80: Substantial agreement - κ = 0.81-1.00: Almost perfect agreement

Benchmark: Fitzpatrick17k achieved κ = 0.70-0.75 between dermatologists.

4.2 Confusion Matrix Analysis

Track common misclassifications: - FST III vs IV (most frequent disagreement) - FST II vs III (fair vs intermediate skin) - Erythema (redness) leading to lighter FST classification

---

5. Annotation Guidelines & Best Practices

5.1 Visual Assessment Rules

What to Annotate: - Classify the patient's skin tone, NOT the lesion itself - Use perilesional skin (adjacent healthy skin) as primary reference - Consider anatomical site: Arms, legs may be tanner than torso

What to Avoid: - Do NOT classify based on lesion color (melanomas can be dark on light skin) - Do NOT use hair color, eye color as proxy for skin tone - Do NOT let demographic assumptions (name, location) bias annotation

5.2 Handling Edge Cases

Case 1: Erythema (Redness) - Problem: Inflammation can artificially lighten ITA estimate - Solution: Focus on non-inflamed perilesional skin, adjust FST upward if erythema present

Case 2: Tanning / Sun Damage - Problem: Anatomical site (e.g., face, arms) may be darker than constitutional skin type - Solution: Annotate "effective skin tone" (observed), not constitutional. Note in metadata if sun-exposed site.

Case 3: Vitiligo / Pigmentation Disorders - Problem: Skin tone is non-uniform - Solution: Annotate predominant skin tone, flag as "pigmentation disorder" in metadata

Case 4: Poor Lighting / Shadows - Problem: Shadows artificially darken skin, overexposure lightens - Solution: Mark as "ambiguous" if lighting precludes reliable annotation. Request image recapture if possible.

5.3 Monk Skin Tone Reference Chart

Visual Guide: Use official MST reference card (10 color swatches) - Print on calibrated color printer (sRGB, D65 illuminant) - View under standardized lighting (D65 daylight simulator, 6500K) - Compare perilesional skin to nearest MST swatch

Digital Alternative: Use MST digital palette (hex codes available on Google AI website)

---

6. FST Stratification for Model Evaluation

6.1 Binary Grouping (Simplified Evaluation)

Light Skin (FST I-III): Majority group in existing datasets Dark Skin (FST IV-VI): Historically underrepresented group

Use Case: Initial fairness gap quantification (AUROC light vs dark)

6.2 Granular Grouping (Comprehensive Evaluation)

Three Groups: - FST I-II: Very light to light - FST III-IV: Intermediate to tan - FST V-VI: Brown to very dark

Use Case: Detailed subgroup analysis, identify specific underperforming tones

6.3 Monk Skin Tone Grouping (Finest Resolution)

Five Groups: - MST 1-2: Very light - MST 3-4: Light - MST 5-6: Medium - MST 7-8: Dark - MST 9-10: Very dark

Use Case: Research publications, maximum fairness transparency

---

7. Dataset-Specific Annotation Plans

7.1 Datasets with Existing FST Labels

Fitzpatrick17k, DDI, MIDAS, SCIN: - Action: Use existing labels (validated by dermatologists) - Quality Check: Spot-check 10% with ITA algorithm, flag major discrepancies - No re-annotation required (trust expert curation)

7.2 Datasets WITHOUT FST Labels

HAM10000, ISIC 2019: - Phase 1 (Immediate): Run automated ITA on all images - Phase 2 (Week 2-3): Human annotation for validation set (1,000 images) - Phase 3 (Week 4): Train FST classifier (ResNet18) on validated set, pseudo-label remaining images - Validation: Expert review of 100 random pseudo-labeled images

Pseudo-Labeling Strategy: 1. Manually annotate 5,000 diverse images (covering all FST types) 2. Train FST classifier: ResNet18, 6-class (FST I-VI) 3. Predict FST for remaining 60,000+ images 4. Use predicted labels for stratified splits, report as "estimated FST" in publications

---

8. Implementation Timeline

Week 1: - Set up LabelBox project (or alternative platform) - Create annotator training module - Run ITA algorithm on HAM10000 + ISIC 2019

Week 2: - Recruit and train annotators (3-5 laypersons) - Begin multi-annotator review (target: 500 images/week) - Establish expert adjudication workflow

Week 3: - Complete 1,500+ annotations - Calculate inter-rater reliability (Cohen's Kappa) - Identify and document common disagreement patterns

Week 4: - Train FST classifier for pseudo-labeling - Validate pseudo-labels (expert review) - Finalize stratified train/val/test splits

---

9. Software Tools & Libraries

9.1 ITA Calculation

# Dependencies
pip install opencv-python scikit-image numpy

# Implementation: src/data/fst_annotation.py

9.2 Annotation Platforms

• LabelBox: https://labelbox.com (free tier available)
• CVAT: https://github.com/opencv/cvat (self-hosted)
• Label Studio: https://labelstud.io (open-source)

9.3 Inter-Rater Reliability

# Cohen's Kappa calculation
from sklearn.metrics import cohen_kappa_score

kappa = cohen_kappa_score(annotator1_labels, annotator2_labels)

---

10. Ethical Considerations

10.1 Privacy & Consent

• All images must be de-identified (no patient names, dates, metadata)
• Use only datasets with proper informed consent
• FST labels should NOT enable re-identification

10.2 Avoiding Racial Essentialism

CRITICAL PRINCIPLE: FST/MST are visual appearance classifications, NOT proxies for race/ethnicity.

• Do NOT conflate skin tone with race (e.g., "FST VI = Black race" is FALSE)
• Do NOT use FST to infer other demographic attributes
• Transparent disclosure: Models use FST for fairness-aware training, NOT for racial profiling

10.3 Annotator Bias Mitigation

• Train annotators to recognize implicit bias (e.g., assuming lighter skin for certain names)
• Blind annotators to patient demographics during labeling
• Diverse annotator pool (recruit annotators of varied skin tones)

---

11. Quality Metrics

Target Benchmarks: - Inter-rater Kappa: > 0.70 (substantial agreement) - ITA-Human agreement: > 75% (within 1 FST category) - Expert validation accuracy: > 90% (for pseudo-labeled data)

Monitoring: - Weekly inter-rater reliability reports - Monthly expert validation audits - Annotator performance tracking (flag outliers)

---

12. References

Key Literature: 1. Groh, M., et al. (2021). Evaluating Deep Neural Networks Trained on Clinical Images in Dermatology with the Fitzpatrick 17k Dataset. arXiv:2104.09957. - Relevance: LabelBox annotation, inter-rater reliability analysis

1. Chardon, A., Cretois, I., & Hourseau, C. (1991). Skin colour typology and suntanning pathways. International Journal of Cosmetic Science, 13(4), 191-208.

2. Relevance: Original ITA formula and FST mapping

3. Monk, E. (2022). The Monk Skin Tone Scale. Google AI.

4. Relevance: MST 10-point scale, inter-rater agreement improvements

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

6. Relevance: SCIN dataset, eFST and eMST annotation methodology

7. Del Bino, S., et al. (2006). Clinical and biological characterization of skin pigmentation diversity and its consequences on UV impact. International Journal of Dermatology, 45(3), 314-319.

8. Relevance: Limitations of Fitzpatrick scale, need for objective measurements

Automated Annotation Tools: - Scarletred Vision: https://www.scarletred.com/scarletredderm (ITA commercial platform) - OpenCV + scikit-image: Open-source ITA implementation

---

13. Contact & Support

Protocol Questions: the_didact@mendicant-bias.ai Annotation Platform Issues: [Platform-specific support] Expert Adjudication Requests: [Designated dermatologist email]

---

Version Control: - v1.0 (2025-10-13): Initial protocol established - Next Review: 2025-10-20 (post-pilot annotation batch)

Maintained by: the_didact (MENDICANT_BIAS framework) Approved by: mendicant_bias (Supreme Orchestrator)