# Medical AI training reduces false lesion reports but misses critical external cases

_Published Wednesday, November 11, 2026 at 12:13 AM EST · AI, Science · Latest · Tier 2 — Notable_

Researchers report in a preprint that training MedScope-4B on images with lesions erased reduced its tendency to report findings absent from an image. On 507 internal cases across five imaging types, it claimed an erased lesion in 9.6% of cases, compared with 41-62% for the frontier models tested, without significant loss of accuracy on real lesions.

The training combined erased lesions, control edits elsewhere and unedited lesion-free images to teach the model when to withhold a finding. External testing exposed a critical limitation: MedScope-4B detected 0 of 48 pneumothorax films in a fresh test. The authors conclude that the training changes reporting behavior without improving perception.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2611.13006)

---
Canonical: https://techandbusiness.org/newswire/CmHDqJdpBzjRdQ5W9T_j5P
Published: 2026-11-11T05:13:26.899Z
Story chronology: 2026-11-11T05:00:00.000Z
Retrieved: 2026-11-11T07:00:32.470Z
Publisher: Tech & Business (techandbusiness.org)