Towards Fair DeepNet: Evaluating Gender Bias in Deep Learning Networks

Deep learning models have become indispensable in healthcare, particularly critical applications such as disease diagnosis.However, these models are vulnerable to biases associated with sensitive attributes, including gender, race, and age.Such biases can lead to inequitable outcomes, disproportionately affecting certain demographic groups and compromising the reliability of AI-driven healthcare systems.This study evaluates the effectiveness of our UnBias method in comparison with four widely recognized fairness metrics, including Statistical Parity Difference (SPD), Equal Opportunity Difference (EOD), Average Odds Difference (AOD), and Disparate Impact (DI) using the NIH Chest X-ray dataset for gender-biased data in lung infiltration disease detection.We evaluated four deep learning architectures, ResNet50, DenseNet121, InceptionV3, and Xception, and found that UnBias outperformed traditional fairness metrics in detecting bias, especially with just 20 sample images.UnBias showed greater sensitivity to biases, particularly in smaller datasets, with CIs ranging from 0.20 to 0.60.Additionally, three bias mitigation techniques were assessed: Disparate Impact Remover (DIR), Adversarial Debiasing (Adv), and Calibrated Equalized Odds (CEOdds).Results indicated that adversarial debiasing was the most effective method of reducing bias and improving fairness across biased datasets.These findings highlight the importance of incorporating UnBias and effective mitigation techniques to ensure fairness in healthcare applications, where biased models can lead to critical clinical consequences.