Explainable AI for Biomedical Image Processing
summary
Conference claim: over thousand FDA approved AI algorithms exist but clinical uptake lags. Built a validation platform called Glenwell AI to run external model checks and expose generalizability failures. Shows that common saliency maps sometimes match untrained or random models and behave like edge detectors, undermining trust. Demonstrates workable alternatives: counterfactual explanations generated by GANs/diffusion models that produce realistic image edits clinicians can read to link visual changes to prediction flips.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:01AI in biomedical imaging
- 3:08Clinical adoption challenges
- 5:57Black box dilemma
- 8:24Why explainable AI matters
- 10:09High-dimensional image challenges
- 10:52Post-hoc explainability methods
- 16:13Saliency map limitations
- 18:19Testing with parameter randomization
- 20:28Evaluating saliency reproducibility
- 22:00Assessing saliency repeatability
- 24:28Counterfactual explanations
- 27:21Generative models for counterfactuals
- 30:15Clinical utility of segmentation
- 32:46Q&A on saliency maps
- 34:10Accuracy vs visualization flaws
- 36:20Constraining generative AI
- 38:30Expert validation limits
- 39:15Explainability vs knowledge discovery
- 41:06Saliency vs SHAP values
- 43:17Closed-loop model feedback
