Enhancing Segmentation Fairness Through Curriculum Learning and Progressive Loss

Medical image segmentation is crucial for precise anatomical delineation in diagnostic and therapeutic procedures. Despite significant advancements in medical image segmentation models, achieving both high accuracy and fairness remains a challenging and underexplored area, as improvements in one metric often lead to reductions in the other. This study addresses the challenge of enhancing both accuracy and fairness in segmentation by mitigating demographic biases through supervised curriculum learning and progressive loss. We employ a manually annotated dataset from the Osteoarthritis Initiative (OAI), including both hip and knee radiographs. By applying various curriculum learning strategies and distinct progressive loss functions that shift focus from easier to more challenging examples, we aim to improve the models’ accuracy and fairness. By considering demographic factors such as race and gender, we evaluate and mitigate biases in segmentation outcomes, leading to enhanced segmentation accuracy. Our findings contribute to the advancement of medical image analysis and the promotion of fair AI models for healthcare applications. Co-sponsored by: XDI Lab (www.xdilab.com) Virtual: https://events.vtools.ieee.org/m/504152