2025年12月6日 –, Main Stream 語言: English
Medical AI models often fail in African settings due to hidden dataset and model biases. This talk shows how to detect and measure bias in medical imaging using Python. We will explore real examples with MRI and X-ray data, learn subgroup performance analysis, apply fairness improvements, and discuss practical lessons for building more reliable and equitable healthcare AI.
Bias in AI is not just a fairness buzzword, it is a clinical safety issue. In this talk we will work through practical Python workflows for uncovering hidden biases in medical imaging models. You will learn:
• How to explore dataset composition and label distribution with Pandas and Seaborn
• How to compare subgroup performance using MONAI and TorchMetrics
• Lightweight bias mitigation strategies, including reweighting and LoRA based fine tuning
• What fairer performance actually looks like in practice using African brain MRI and chest X-ray examples
I will share successes, failures, and unexpected findings from evaluating segmentation models across datasets. You do not need a medical background, only curiosity and some Python experience. You will leave with practical tools to ensure your models work more equitably for diverse patient populations, especially within the African context.
I’m a Mechatronics Engineering student passionate about using Python to solve real-world problems. I’m actively involved in several Python-focused communities, including ML Collective Nigeria, where I currently lead the Medical Imaging Focus Group. I’m also a Microsoft Learn Student Ambassador and a member of AfricaAI, the largest community of Africans improving medical imaging with AI. On campus, I serve as the co-lead of machine learning at the Google Developer Student Club, where I help others explore the power of AI. My Python journey began with web development using Flask and Django, but my current focus is on research in deep learning for medical imaging—applying computer vision to health challenges in low-resource settings.