Reproduce “MEDebiaser: A Human-AI Feedback System for Mitigating Bias in Multi-label Medical Image Classification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shaohan Shi, Yuheng Shao, Haoran Jiang, Yunjie Yao, Zhijun Zhang, Xu Ding, Quan Li
MEDebiaser is a human-AI feedback system designed to mitigate biases in multi-label medical image classification by enabling physicians to directly refine AI models. The system leverages local explanations and pixel-level annotations to improve model accuracy for underrepresented conditions, enhancing collaboration between medical professionals and AI developers.
Key insights
- Direct, pixel-level modification of AI model attention by medical experts, guided by visual explanations, effectively mitigates biases related to imbalanced label distribution and co-occurrence in medical images.
- An intuitive, attention-guided human-AI feedback system can significantly streamline collaboration between physicians and engineers, reducing communication gaps and lowering the learning curve for domain experts.
- The iterative integration of expert feedback into the model's loss function (combining prediction and attention loss) allows for targeted refinement, leading to more accurate, reliable, and trustworthy medical AI models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.