New study: Grad-CAM localizes chest X-ray findings at 0.575 vs radiologists' 0.888
zakkohane · x · 2026-11-05
Saliency maps — the go-to 'explainable AI for images' method — take another hit. A new study comparing DenseNet-121 and ViT-B16 attribution methods against radiologist annotations finds Grad-CAM localizes findings far less often than radiologists (hit rate at most 0.575 vs 0.888). Validated on an NIH internal test set (21,834 radiographs, 5,161 patients) plus external CheXpert (Stanford) and VinDr-CXR (Hanoi) data, the study covers localization, explanation shift, and failure detection.
More from Research
- UW study: leading AI models rate Chinese and Nigerian accents lower in speech scoring — lazowska · 2026-11-05
- Paper asks whether LLMs are rational: coherence norms published in AI journal — peterbhase · 2026-11-05
- The Data Efficiency Gap Between Humans and AI: No System Has Both Targets — dyamins · 2026-11-05
- 0.8B open-source HaloGuard 2.0 claims streaming-guard SOTA, beating models 5-10x its size — const_reborn · 2026-11-05
- Anima Anandkumar to keynote SC26 supercomputing conference in Chicago — AnimaAnandkumar · 2026-11-05
- MIT's Weight-Space Programming: LLMs edit weights natively, fixing Qwen math with 0.016% of weights — dbreunig · 2026-11-05