Title: Investigating Demographic Bias in Skin Lesion Classification

Abstract

This research assesses how demographic biases within training data influence the efficacy of ResNet-based convolutional neural networks for skin lesion classification, with a specific focus on variations in patient age and sex. To facilitate a systematic examination of these bias effects, we employed linear programming to construct datasets with precisely controlled demographic profiles. The study compares three distinct learning methodologies: a standard single-task model, a reinforcing multi-task framework, and an adversarial learning approach.

Our analysis regarding sex reveals that training models on sex-specific datasets yields the best performance. Interestingly, incorporating male subjects into the training set enhanced outcomes for the male subgroup, even when the dataset was predominantly female. Furthermore, both the reinforcing and adversarial learning methods successfully reduced or entirely removed bias gaps in datasets that were either balanced or skewed toward females. However, these mitigation strategies were less successful in male-dominated environments, where models continued to favor male patients over female ones, showing only marginal improvements over the baseline in such populations.

Regarding age, the baseline performance across the three model types remained similar, but accuracy declined as patient age increased. Younger cohorts consistently achieved the highest performance metrics, irrespective of the training data’s distribution. While balanced training produced the best results for the youngest age group, performance dropped for older categories. The findings suggest that sex-based disparities stem primarily from data imbalances, whereas age-based biases inherently favor younger individuals regardless of dataset composition. These differing underlying mechanisms necessitate tailored mitigation approaches. Finally, cross-dataset validation using two external datasets highlighted that domain shifts significantly impact both overall performance and the specific patterns of demographic bias.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC