Study finds VLMs dangerously overconfident in brain MRI analysis
Vision-Language Models fail on 46% of high-confidence brain MRI errors
Six instruction-tuned vision-language models, including one medical specialist, were audited on 4,102 images: 4,032 brain MRI slices from 250 subjects plus 70 non-brain/noise controls. The models answered almost everything, but their verbalized confidence was dangerously miscalibrated: expected calibration error ranged from 0.27 to 0.40, mean confidence on wrong answers hit 0.82–0.97, and 33–46% of all answered items were high-confidence errors. The most accurate model was also the most confident on its own mistakes, and while medical specialization improved tumor-presence detection, it didn't improve confidence reliability. Open-ended responses showed hallucination and abstention moving independently from multiple-choice accuracy, arguing that medical-image VLM evaluation must track confidence reliability, confident errors, hallucination, and abstention alongside accuracy.
- Tested 6 VLMs (5 general + 1 medical specialist) across 4,102 brain MRI slices from 250 subjects
- 33-46% of errors were high-confidence mistakes with incorrect confidence scores between 0.82-0.97
- Medical specialist models showed improved accuracy but not better confidence calibration
Why It Matters
Medical AI deployments need confidence calibration checks before clinical use to prevent dangerous overconfidence in diagnostic errors