MEDebiaser: A Human-AI Feedback System for Mitigating Bias in Multi-label Medical Image Classific...
summary
An interactive system called MEDebiaser lets physicians upload datasets, view label distributions and co‑occurrence, and directly correct model attention by drawing polygons on images; the system fine‑tunes a one‑vs‑rest classifier using Grad-CAM heat maps and a joint loss that combines prediction with attention alignment loss and a customized ranking that ranks problematic cases. User studies with six clinicians show reduced labeling effort, improved rare‑label performance, and clearer shifts in attention after iterative feedback.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:12Research background
- 4:25System design Medvisor
- 7:37Evaluation results
- 9:44Discussion and insights
- 11:08Q&A
related talks
18mMedical AI Beyond Prediction: From Normative Learning to Agentic Healthcare AI
47mFight AI Bias With… Bias
47mAI for making decisions in clinical medicine
04mBlack Box: How AI will Amplify the Best and Worst of Humanity
38mSocial Implications of Bias in Machine Learning
10mMIND MELD: Personalized Meta-Learning for Robot-Centric Imitation Learning
20mDiagnostic Medical application of ML
20mA Machine Learning Approach to Protect Patients from Blood Tube Mix-Ups