Tuesday 11 March 2025
A recent study has shed light on a worrying issue in medical image analysis: deep learning models may be biased against darker skin tones. The researchers used generative models, which are trained to generate synthetic images of skin pathologies, and found that even when the training set is balanced with an equal number of light and dark skinned individuals, the model still performs better on lighter skin tones.
The study used a dataset of high-quality skin images from various skin tones, including Fitzpatrick skin types 1-6. The researchers trained a Variational Autoencoder (VAE) to generate synthetic images of skin pathologies such as acne, psoriasis and melanoma. They then evaluated the performance of the model on test sets of light and dark skinned individuals.
The results showed that the VAE performed better on lighter skin tones, even when the training set was balanced. This suggests that there is more at play than just a simple representation bias. The researchers also found that the uncertainty estimates produced by the VAE were ineffective in assessing the model’s fairness.
This study highlights the importance of considering skin tone diversity in medical image analysis. Current datasets are dominated by light-skinned individuals, which can lead to biased models that may not perform well on darker skin tones. The authors suggest that there is a need for more representative dermatological datasets and better understanding of the sources of bias in such models.
The study’s findings have implications for the development of artificial intelligence (AI) in healthcare. AI-powered diagnosis tools are increasingly being used in clinical settings, but if they are biased against certain groups, this can lead to inaccurate diagnoses and poor patient outcomes.
The researchers also highlight the need for better understanding of the relationship between skin tone and disease prevalence. They found that there were differences in the distribution of conditions across skin tones, which could affect the performance of the model.
The study’s findings suggest that more work is needed to ensure that AI-powered diagnosis tools are fair and accurate for all individuals, regardless of their skin tone.
Cite this article: “Biases in Deep Learning Models for Medical Image Analysis: A Study on Skin Tone Diversity”, The Science Archive, 2025.
Deep Learning, Medical Image Analysis, Bias, Skin Tone, Dermatology, Artificial Intelligence, Healthcare, Diagnosis Tools, Fairness, Accuracy
