Unintended Bias in Artificial Intelligence Driven Diagnosis of Melanoma: A Systematic Review

Abstract

Melanoma remains a public health crisis, with incidence rates increasing rapidly in the past decades. Improving diagnostic accuracy to decrease misdiagnosis using Artificial intelligence (AI) continues to be documented. Unfortunately, unintended racially biased outcomes a product of lack of diversity in the dataset used, with a noted class imbalance favoring lighter vs. darker skin tone have increasingly been recognized as a problem. Resulting in noted limitations of the accuracy of the Convolutional neural network (CNN) models. CNN models are prone to biased output due to biases in the dataset used to train them. Although the incidence of melanoma is lower in patients with darker skin tone, it is associated with a worse prognosis than in Caucasians, underscoring the need for accurate early diagnosis in these patients. Our objective in this systematic review was to assess to what degree race/ethnicity, specifically Black/ African American patient cohort were included in training datasets used in generating machine learning algorithms for automated melanoma diagnosis. Our review documents the fact that there is a remarkable lack of and inconsistent reporting of patient demographics especially race/ethnicity with notable under-representation of patients of color, highlighting a currently unmet critical need of lack of diversity in the publicly available skin image datasets. These publicly available skin image datasets, are an inherently unbalanced unintentionally biased datasets from which AI models created for the diagnosis of melanoma skin cancer have restricted applicability to real life clinical scenarios and limited population representation preventing generalizability to the community as a whole.