This study audits and mitigates fairness issues in cardiac MRI segmentation across SIEMENS, Philips, and GE scanners. A baseline 2D U-Net showed spurious vendor bias, particularly for the minority GE domain. Implementing a Domain Adversarial Neural Network reduced F1-Score disparity, stabilizing recall and improving clinical safety.
auditing pytorch mri segmentation deeplearning equity bias fairness mitigation u-net medicalai dann medicalimaging domain-adversarial-neural-networks healthcareai cardiacmri dataimbalance vendorbias clinicalai
-
Updated
Dec 20, 2025 - Jupyter Notebook