# Preprint reports improved evidence checks for AI-generated chest CT findings

_Published Wednesday, October 28, 2026 at 2:05 AM EDT · Science, AI · Latest · Tier 2 — Notable_

Researchers report that EviGround-CT, a system for checking AI-generated chest CT reports against scans, raised evidence-supported precision from 0.263 to 0.365 compared with conditioning on finding categories. The preprint evaluated verification on a held-out 100-case ReXGroundingCT subset containing 137 extracted in-scope positive findings.

The system extracts sentences describing positive findings, then predicts a three-dimensional evidence mask for each sentence. Using complete sentences preserves anatomical location, laterality and descriptive modifiers, linking individual statements to scan regions that can be inspected. The unsupported positive finding rate fell from 0.737 to 0.635 under the available annotations, leaving substantial unsupported content.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.33038)

---
Canonical: https://techandbusiness.org/newswire/KNpoI7k19Ib56GpQp87TS_
Published: 2026-10-28T06:05:36.736Z
Story chronology: 2026-10-28T04:00:00.000Z
Retrieved: 2026-10-28T10:24:17.131Z
Publisher: Tech & Business (techandbusiness.org)