A Generative AI Approach for Reducing Skin Tone Bias in Skin Cancer Classification

1. Problem Statement

The paper addresses a critical issue in automated dermatological diagnostics: dataset bias regarding skin tone.

• The Disparity: Current AI diagnostic tools are predominantly trained on datasets (such as the International Skin Imaging Collaboration, or ISIC, dataset) that are heavily skewed toward lighter skin tones. The ISIC dataset contains over 70% images of light skin (Fitzpatrick types I–III) and fewer than 8% of dark skin (Fitzpatrick types V–VI).
• The Consequence: Models trained on this imbalanced data exhibit reduced accuracy, sensitivity, and fairness when diagnosing skin cancer in patients with darker skin tones, potentially leading to missed or delayed diagnoses in underserved populations.
• The Challenge: Collecting additional real-world clinical images of dark-skin lesions is difficult due to privacy regulations, institutional barriers, and the scarcity of annotated data from underrepresented groups.

2. Methodology

The authors propose a Generative AI-based augmentation pipeline to synthesize high-quality dermoscopic images of dark skin, thereby balancing the training dataset. The methodology consists of six stages:

A. Data Analysis and Subset Identification

• Dataset: The study utilizes the ISIC dataset (17,728 images).
• Skin Tone Quantification: The authors computed the Individual Typology Angle (ITA) for every image to objectively map skin tone to the Fitzpatrick Skin Type (FST) scale.
• Target Subset: They isolated the underrepresented dark-skin subset (FST V and VI), which comprised only 1,407 images (approx. 7.94% of the total), containing 223 melanoma and 1,184 non-melanoma cases.

B. Generative Model Fine-Tuning (LoRA)

• Base Model: A pre-trained Stable Diffusion model was selected for its ability to generate high-fidelity images.
• Adaptation Technique: Instead of full fine-tuning, the authors used Low-Rank Adaptation (LoRA). This injects small, trainable low-rank matrices into the attention layers of the U-Net backbone, allowing for efficient domain adaptation with minimal parameter overhead while keeping the base model frozen.
• Training Data: The LoRA adapter was trained exclusively on the 1,407 dark-skin images.

C. Synthetic Image Generation

• Conditional Generation: The model generated 808 synthetic dermoscopic images conditioned on specific text prompts describing the lesion type (melanocytic vs. non-melanocytic) and skin tone.
  • 144 synthetic melanocytic lesions.
  • 664 synthetic non-melanocytic lesions.
• Validation & Filtering: Generated images were validated using statistical metrics (color histograms, Structural Similarity Index [SSIM], and Grey Level Co-occurrence Matrix [GLCM]) to ensure they matched the distribution of real dark-skin images. Outliers and artifacts were removed.

D. Dataset Integration

• The 808 validated synthetic images were merged with the original 17,728 images, creating an augmented dataset of 18,536 images.
• The dataset was split into training (70%), validation (15%), and test (15%) sets using stratified sampling at the patient level to prevent data leakage.

E. Downstream Evaluation Tasks

The utility of the augmented dataset was tested on two tasks:

1. Lesion Segmentation: To verify if synthetic images contain genuine structural information.
   • Approach: A Fully Convolutional Network (FCN) was trained on the augmented dataset.
   • Crucial Constraint: The model was evaluated strictly on a held-out set of real images (no synthetic images in the test set) to prove the model learned transferable features rather than memorizing synthetic patterns.
2. Binary Classification: To assess diagnostic performance.
   • Model: EfficientNet-B0 (initialized with ImageNet weights).
   • Training: Fine-tuned on the augmented dataset (Real + Synthetic) vs. a baseline trained only on real data.

3. Key Contributions

• Novel Pipeline: A framework for targeted synthetic augmentation of underrepresented skin tones using LoRA-adapted Stable Diffusion.
• Bias Mitigation: Successfully increased the representation of Fitzpatrick V–VI skin types in the ISIC dataset by nearly doubling the available dark-skin samples (from 1,407 to ~2,215 total).
• Dual-Task Validation: Provided complementary evidence of synthetic data utility through:
  • Segmentation: Proving that models trained on synthetic data improve boundary accuracy on real images.
  • Classification: Demonstrating improved diagnostic metrics on the augmented distribution.
• Efficiency: Showed that parameter-efficient fine-tuning (LoRA) is a viable strategy for adapting large generative models to small, specialized medical sub-datasets.

4. Results

Segmentation Performance (Validation on Real Data)

The augmented model demonstrated superior performance on real-world test data, confirming the clinical relevance of the synthetic images:

• Mean IoU: Increased from 0.82 (Baseline) to 0.85 (Augmented).
• Dice Coefficient: Increased from 0.88 to 0.90.
• Hausdorff Distance: Decreased from 12px to 10px, indicating tighter boundary adherence.
• Note: Both CNN variants significantly outperformed the non-learning Max-Flow algorithm (IoU 0.75).

Classification Performance

Using EfficientNet-B0, the augmented model showed significant improvements in standard metrics:

• Accuracy: Improved from 85.45% (Baseline) to 92.14% (Augmented).
• Precision: Increased from 0.82 to 0.89.
• Recall: Increased from 0.78 to 0.87.
• F-Score: Improved from 0.80 to 0.88.
• AUC Observation: Validation AUC slightly decreased (0.9765 to 0.9480). The authors attribute this to the increased distributional complexity of the synthetic images making probability ranking harder, though threshold-based classification accuracy improved.

Computational Efficiency

• Inference Time: The augmented segmentation model added negligible overhead (120ms to 130ms per image).
• Generation Time: Synthetic image generation took ~4–6 seconds per image, suitable for offline dataset augmentation but not real-time generation.

5. Significance and Future Directions

• Equitable Healthcare: The study demonstrates that Generative AI can effectively mitigate demographic bias in medical imaging, offering a scalable solution to the scarcity of diverse clinical data without violating patient privacy.
• Clinical Validity: The improvement in segmentation metrics on real test images provides strong evidence that the synthetic data captures morphologically correct features, not just visual noise.
• Limitations & Future Work:
  • Clinical Validation: The synthetic images have not yet been validated by dermatologists for clinical plausibility.
  • Generalization: Future work requires testing on strictly real-only test sets to confirm generalization to unseen patient populations.
  • Scale: The current augmentation volume (4.6% increase) is modest; future iterations should aim for larger, more class-balanced synthetic sets.
  • Inter-domain Application: The methodology is applicable to other medical fields (radiology, pathology) facing similar data imbalances.

In conclusion, this paper establishes a robust proof-of-concept that Generative AI, specifically LoRA-adapted diffusion models, can successfully reduce skin tone bias in skin cancer detection, leading to more accurate and fair diagnostic tools for diverse patient populations.