Zero-shot multimodal large language models underperform a domain-trained CNN baseline in pediatric wrist fracture detection
Matteo Haupt, David Weiß, Tim Bellersen, Martin H Maurer · 2026
Read Zero-shot multimodal large language models underperform a domain-trained CNN baseline in pediatric wrist fracture detection by Matteo Haupt, David Weiß, Tim Bellersen, Martin H Maurer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13276184.txt.
Description
Matteo Haupt and co-authors report a 2026 journal article testing whether big artificial intelligence chatbots can read children's wrist X-rays. Models that handle text and images are often proposed for medical use. The team gave three commercial models no training examples, a strict zero-shot test, for spotting wrist fractures.
Their test set held one thousand children and nearly twenty-three hundred radiographs, evenly split between fracture and no fracture. A specialized neural network trained on this task scored high accuracy. All three chatbots, by contrast, performed close to chance, getting barely half the cases right. The boxes they drew around suspected fractures often missed expert markings. The authors conclude these general models are research toys here, not dependable tools for reading children's X-rays.
Catalog record
| CREATOR | Matteo Haupt, David Weiß, Tim Bellersen, Martin H Maurer |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38781 |
| LICENSE | https://creativecommons.org/licenses/by/4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |