The paper will be available after the work's presentation at the IEEE iAIMS 2026 conference
Synthetic Data Replacement in Dermatology: Effects on External Generalization found that quality checks for synthetic data are needed. Synthetically-generated clinical data by frontier AI models contained errors where lesion texture, skin features, or surrounding pigmentation would be distorted.
Predictive accuracy across trained AI lesion classification models dropped below 3-class guessing (0.33) when synthetic data was involved in training data
Subtler features in samples such as Erythema were removed during synthetic generation, and removal occurred the most among synthetic edits requesting for dark skin complexions.
Paper-writing work session