August 31, 2024Open Access

Fairness Matters: Evaluating and Mitigating Bias in Skin Cancer Classification

Key Points

The study highlights significant bias in AI models used for skin cancer detection, especially for darker skin tones.
Using the fitzpatrick dataset, accuracy discrepancies were shown among various skin types, notably affecting diagnoses.
Five AI models were evaluated, and four augmentation techniques were implemented to address identified biases effectively.
Results indicate that adjustments in training parameters can influence bias outcomes in AI skin cancer detection models.

Abstract

Almost 1 in 5 Americans develop skin cancer by the age of 70, and it is the most commonly diagnosed cancer in the United States, with most cases being preventable. Artificial intelligence (AI) has shown great potential to accelerate the diagnosis of skin cancer for early treatment. However, the fairness of using AI for skin cancer detection has raised concerns due to the lower accuracy of "darker" skin tone detection. This paper conducts a comprehensive study on the bias problem using the fitzpatrick dataset, while analyzing the effects of different methods towards mitigating this bias. Throughout our experiments, we found that not only was the darkest skin type biased, one of the specific light skin types also had a relatively low accuracy. Five AI models were evaluated and compared in terms of bias, and four augmentation techniques were applied to mitigate bias. Moreover, we studied the impact of training parameters (e.g. batch size, data splitting) on bias.

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper