Synthetic Skin Lesion Images for Fairness-Aware ML, 10,000 Images

A synthetic collection of 10,000 skin lesion images generated by Safrina Kabir using a tone-conditioned Generative Adversarial Network (GAN). The images, at 128x128 pixel resolution, are equally divided between medium and dark skin tones and follow the class distribution of the HAM10000 dataset. This dataset, last updated in April 2026, is intended to support fairness-aware experimentation and data augmentation in dermatological machine learning.

Use Cases

• Fairness-aware model evaluation based on synthetic images representing darker skin tones.
• Data augmentation for skin lesion classification models based on the described synthetic image generation.
• Benchmarking generative model quality using reported Fréchet Inception Distance (FID) scores.
• Controlled experimentation on dataset bias and representation in medical imaging.

Strengths

• Contains 10,000 generated images, providing a substantial synthetic sample.
• Images are equally divided between medium and dark skin tones, addressing a stated representation gap.
• Image resolution is explicitly stated as 128x128 pixels.
• Includes quantitative characterization via Fréchet Inception Distance (FID) scores and a baseline classification experiment.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is very small at 5.5 KB, suggesting it contains metadata or labels rather than the image files themselves.
• The actual image files are not directly available in the described XLS format; users must locate them separately.

Provenance

Source

Safrina Kabir via figshare.

Collection Method

Generated using a tone-conditioned Generative Adversarial Network (GAN).

Time Range

Synthetic generation date not specified; metadata updated April 2026.

Freshness

Last updated 2026-04-17 17:40:38; freshness should be verified.

Geography

Spatial coverage is not specified.