Preliminary Assessment of Racial Disparities in AI-based Prostate Cancer Detection on bpMRI

Motivation: AI models are being developed for prostate cancer (PCa) detection on MRI but potential racial bias in these models is not well understood. Goal(s): This study evaluates potential racial bias in a deep learning (DL) classifier for clinically significant PCa detection on bi-parametric MRI (bpMRI) Approach: An AI model using 3D ResNet50 encoders was trained on bpMRI data. Model performance was assessed, focusing on White and Black or African American patients. Results: The model achieved an AUC of 0.83 for White patients and 0.77 for Black patients, indicating higher predictive accuracy for White patients. Impact: This study reveals potential racial bias in an AI model for Prostate cancer detection on MRI, with lower predictive accuracy for Black patients. These findings emphasize the need for further work to ensure equitable AI algorithms for PCa detection.