A new study published on arXiv has found that current medical Vision-Language Models (VLMs) fail to provide faithful visual explanations for their predictions on chest X-rays. Researchers evaluated several VLMs, including MedGemma-4B variants, LLaVA-RAD, and Qwen3-VL-8B-Instruct, alongside specialist models like CheXagent-2-3b and traditional classifiers. The findings indicate that while some VLMs use image data, their attention heatmaps do not accurately reflect the image regions crucial for their predictions, and they often miss annotated anatomy. This suggests that visual explanations from these models are more reassuring than informative, highlighting the need for rigorous localization metrics and causal perturbation methods for reliable clinical explanations. AI
IMPACT Highlights critical limitations in VLM explainability for medical applications, necessitating new evaluation standards.
RANK_REASON Academic paper detailing novel evaluation methodology for VLM explainability. [lever_c_demoted from research: ic=1 ai=1.0]
- Binesh Kumar Sadanandan
- CheXagent-2-3b
- CheXlocalize
- DenseNet121
- LLaVA-RAD
- MedGemma 4B
- Qwen3-VL-8B-Instruct
- ResNet50
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
