Background: Artificial intelligence (AI) image recognition models have been relatively successful in diagnosing cutaneous manifestations in individuals with light skin tone. However, when these models are tested on the same cutaneous manifestations in individuals with darker or brown skin tone, the performance of the model drops due to a paucity of such images available for model training.
Objective: The objective of this study was to improve the performance of AI models in recognizing cutaneous diseases in individuals with darker skin tone.
Methods: Unsupervised computer darkening of skin color with preservation of the dermatological disease/lesion characteristics in images of light-skinned individuals with basal cell carcinoma (BCC), and melanoma was performed.
Results: Training an AI model on these artificially "darkened" images as compared to training on the original "light-skinned" images resulted in a higher sensitivity, specificity, positive predictive value, negative predictive value, F1 score and area under the receiver-operating characteristic curve of the AI model in differentiating between BCC and melanoma in individuals with brown skin tone.
Conclusion: Use of unsupervised image to image translation in medical AI image recognition models has the potential to significantly improve their accuracy in diagnosing diseases in individuals with racially diverse skin tone.
Keywords: Machine learning; cutaneous manifestations; skin of color.